Last month, I published The Latest Thoughts From American Technology Companies On AI (2026 Q3). In it, I shared commentary in earnings conference calls for the third quarter of 2026, from the leaders of US-listed technology companies that I follow or have a vested interest in, on the topic of AI and how the technology could impact their industry and the business world writ large. 

A few more technology companies I’m watching hosted earnings conference calls for 2026’s third quarter after I prepared the article. The leaders of these companies also had insights on AI that I think would be useful to share. This is an ongoing series. For the older commentary:

• 2023 Q1 – here and here
• 2023 Q2 – here and here
• 2023 Q3 – here and here
• 2023 Q4 – here and here
• 2024 Q1 – here and here
• 2024 Q2 – here and here
• 2024 Q3 – here and here
• 2024 Q4 – here, here, and here
• 2025 Q1 – here and here
• 2025 Q2 – here and here
• 2025 Q3 – here, here, and here
• 2025 Q4 – here, here, here, and here
• 2026 Q1 – here, here, here, and here
• 2026 Q2 – here, here, here, and here

With that, here are the latest commentary, in no particular order:

Alphabet (NASDAQ: GOOG)

Alphabet’s AI investments are driving growth in Search, including commercial queries; the growth rate in queries in Search is the highest seen in the last 5 years; management is building a seamless experience across AI Overviews, AI Mode, and Search; queries in AI mode are 3x longer than traditional searches; 20% of AI Mode searches globally are multimodal, and Alphabet has crossed 50 billion visual searches per month; management is looking to bring agentic experiences to Search; many queries on AI Mode are upper funnel, which allows Alphabet to capture new, high-value commercial intent; management is also seeing that the majority of AI Mode queries are incremental, which creates more monetisation opportunities; management has found that users are comfortable with advertising on AI Mode and AI Overviews, with similar click-through rates between sponsored and organic product list items within AI Mode; management is now scaling high-quality ads within AI Overviews and AI Mode; management is taking the monetisation of AI Overviews and AI Mode global for the 1st time; AI Mode is currently monetising at a lower rate than traditional search, but management feels very good about the trajectory 

Search has strong momentum. Our AI investments are driving increased usage and overall query growth, including commercial queries. In fact, our year-over-year query growth rate is the highest it’s been in the last five years…

…We are building one seamless experience across AI Overviews, AI Mode, and traditional Search. People are using our AI features more, searching in entirely new ways. Their queries are longer, personalized, and multimodal. Queries in AI Mode are on average three times longer than traditional searches, and one in five AI Mode searches globally use voice, images, or documents. We have now crossed over 50 billion visual searches per month…

…People already put their trust in Google to get things done in their daily lives, and our leadership here remains an important differentiator. We focus first on getting these agentic experiences right with the Gemini agent for enterprise, because the bar for security, safety, and governance is extremely high. As a next step, we are bringing these features to Search…

…We think of Search holistically, a unified product spanning traditional results and AI-powered Search. On AI Mode, many of these queries are upper funnel queries, allowing us to capture new, high-value commercial intent. In addition, the majority of AI Mode queries are incremental, creating more monetization opportunities over time. Additionally, our user research shows that people are comfortable with ads on these surfaces and find them highly relevant. Click-through rates are already similar between sponsored and organic product list items within AI Mode. Based on these collective results, we are now scaling high-quality ads within AI Overviews and AI Mode to more users and queries. This includes taking our monetization footprint global for the first time. Today, we are announcing the expansion of ads in AI Mode into international markets alongside multi-turn ads in AI Mode in the U.S…

…It’s worth noting that the strong growth in AI Mode queries had an impact because of the lower monetization we were seeing in AI Mode. To be clear, we feel very good about our trajectory here. Our ad experiments have shown promising results, and we are now scaling ads within both AI Overviews and AI Mode to more users and more queries, including, as I mentioned, international markets and multi-turn ads in the U.S.

Alphabet’s 1st-party model APIs processed 32 billion tokens per minute in 2026 Q3, up 40% sequentially (was 22 billion tokens per minute in 2026 Q2), and up 4x from a year ago; Alphabet’s 2 latest models, Gemini 4 Argon and Gemini 4 Carbon, are leading the frontier; Argon’s usage in its 1st week ramped up 3.5x faster than Gemini 3, despite capacity constraints; management sees Argon performing well in finance, legal, cybersecurity, and science; customers are using Argon to untangle legacy code bases, accelerate scientific research, improve products, and more; Alphabet has used Argon internally over the last few weeks to ship 35% more code with 30% fewer errors; Gemini 4 Carbon is currently Alphabet’s strongest frontier model and it ranks No.1 on the recently updated Artificial Analysis Intelligence Index at 59; Carbon outperforms Argon across nearly all domains; management is rolling out Carbon to a trusted cyber defenders and the US government first, before expanding access; management recently introduced Genie 4, Alphabet’s most advanced world model; management sees Alphabet being at the point where AI is accelerating the company’s AI training and launch speed, and management expects continued acceleration because of RSI (recursive self-improvement); Alphabet will soon be releasing more efficient models, including the Carbon family of models; Alphabet’s Gemini Flash cyber models are frontier standard for cybersecurity and can patch vulnerabilities 2.6x more reliably than other models; the Gemini Flash cyber models will become even better when Argon and Carbon are rolled out

Our first-party model APIs are now processing nearly 32 billion tokens per minute, 40% more than last quarter and more than four times what we processed a year ago. With Gemini 4 Argon and now Carbon, which we announced today, we are excited to be leading at the frontier…

…We are seeing strong demand and usage for Argon, which we launched two weeks ago. In fact, on our developer platforms, usage in the first week ramped up 3 and a half times faster than Gemini 3, even with capacity constraints. Argon is proving to do really well in areas like finance, legal, cybersecurity, and science. Developers and Cloud customers are already using it for things like untangling legacy code bases, accelerating scientific research, robotics, discovering product improvements, and domain-specific tasks requiring in-depth reasoning. At Google, just in the last few weeks, teams across Chrome, Ads, and Search have used Argon to write and optimize code much faster, shipping 35% more code with 30% fewer errors. 

Argon is just the start of the Gemini 4 series. Today, we announced Carbon. It is our strongest frontier model with state-of-the-art cybersecurity and long-horizon capabilities. It ranks number one on the recently updated Artificial Analysis Intelligence Index at 59. Carbon outperforms Argon across almost every domain, like full-stack web development, multimodal understanding, complex knowledge work, and scientific reasoning. We are rolling Carbon out first to a subset of trusted cyber defenders before we expand access, and it is being tested by the U.S. government.

Beyond our Gemini 4 series, yesterday we announced Genie 4, our most advanced world model yet that simulates safe, realistic, interactive worlds with native audio in real-time. This is an important building block for training robots and agents…

…We are at the point where AI is helping us increase our training and launch velocity. I would expect the next 12 months to see continued acceleration because of RSI. We have more models coming soon, including more efficient versions to help developers optimize for speed and cost and to deploy across our billions of users…

…We are working on distilled versions of our Carbon models so that we can deploy it across a billion-user scale efficiently, both from a compute, from a latency standpoint, and from a quality standpoint…

…Our Gemini Flash cyber models are frontier models that automate vulnerability discovery across 20 programming languages and patch 2.6 x more reliably than other models. There is even better performance to come as we roll out Argon and Carbon.

The Gemini app is now at 1 billion users, and is Google’s fastest-growing product ever; management is attempting to bring agentic experiences to the Gemini app; the integration of the Gemini 4 models into the Gemini app has been a big step forward for Alphabet in introducing agentic experiences into the app

The Gemini app is now at 1 billion users and is the fastest-growing product in Google’s history…

…People already put their trust in Google to get things done in their daily lives, and our leadership here remains an important differentiator. We focus first on getting these agentic experiences right with the Gemini agent for enterprise, because the bar for security, safety, and governance is extremely high. As a next step, we are bringing these features to Search and also, of course, to our billion users on the Gemini app. People are already using the Gemini app to do more than ask questions. They are using it to do things, including complex multistep agentic queries. Our Gemini 4 series has been a really big step forward here and has already been handling these complex tasks really well.

Alphabet’s management is focused on providing the right agentic experience with the Gemini agent for enterprise; management thinks agents will be a unifying layer for Alphabet’s products; the Universal Commerce Protocol (UCP) for agentic commerce currently enables direct checkout for hundreds of thousands of brands; management is combining UCP with the Agent Payments Protocol for agentic commerce; management has launched Universal Cart for retailers to maintain their relationships with customers while helping to capture sales and avoid abandoned carts in the agentic commerce experience; management wants to enhance the product discovery experience for users through the agentic experience; management is confident that Alphabet can provide compelling advertising models for agentic use cases; management sees subscriptions and payments as other ways of monetising agentic use cases; management sees agentic use cases as an expansion opportunity for Alphabet; management thinks safety and trust are very important considerations for consumers when engaging with agentic experiences

We focus first on getting these agentic experiences right with the Gemini agent for enterprise, because the bar for security, safety, and governance is extremely high…

…Agents will operate as a unifying layer across our products. You will see us extend these agentic experiences beyond Search in the Gemini app to Chrome, Android, and our Workspace products as well…

…The Universal Commerce Protocol, UCP, is being rapidly adopted by retailers. Today, UCP enables direct checkout for hundreds of thousands of brands across Google, including Chewy, Nike, Tapestry, and many more. We are combining this with the Agent Payments Protocol as we build for the future of commerce. Our focus spans two fronts, unlocking agentic capabilities that act on behalf of the consumer, and designing practical tools that gives shoppers confidence and control. We launched Universal Cart this quarter to support that second track, mirroring how people actually shop today across Google. This allows retailers to maintain their relationships with customers while helping to capture sales and avoid abandoned carts…

…We also want to enhance, not just automate away the discovery experience that people frankly love in so many cases. Users like choice, and they want to consider things way beyond price, like brand loyalty, customer service and so on. We’re really architecting these agentic workflows in a way that we can intentionally design the time and the space for users to explore those choices. So simple example, an agent could respond, “Here’s the insurance product we found, and here’s another strong option, one of which is a higher premium, maybe a lower deductible, and they are currently offering a very special discount.” Given that there’s almost never only one perfect option for a decision, we’re actually confident that we can actually provide really interesting, compelling advertising models for agentic use cases. And obviously you can imagine going beyond us with subscriptions and other payment methods…

…I would view it as an expansion opportunity, right? Any time, just like a few years ago, being able to incorporate AI chat use cases in Search allowed people to ask more and do more in Search, which is what is playing out in our usage. I think agents would again be an opportunity to expand the types of problems we solve for our users… I’m excited about agents and I would view it as a new frontier, which brings in a new set of use cases to Search, and so we are excited to work through that…

…When people are approaching these agentic experiences, we clearly see people are concerned about safety and trust. It is a very high factor in how they want to use these things. Over decades, the trust we have built by delivering these experiences in a safe way across many consumer products with scale in billions of users, as well as this putting in end-to-end as we are building our models and agents, deploying them both with extensive testing, the kind of real-time monitoring we do as we deploy these at scale, I think all of that will end up being a huge differentiator.

Waymo now has more than 300 million fully autonomous miles; Waymo is now operating in 15 major cities in the USA and will be rolling out internationally soon

Waymo is on a great trajectory. We have now driven more than 300 million fully autonomous miles. Waymo now operates in 15 major U.S. cities and just recently opened up service to the public in San Antonio, San Diego, and Tampa, with planned expansions into Canada, Germany, Singapore, and Spain.

Among Alphabet’s other long-term initiatives, Isomorph is using Gemini 4 Agron for drug research

Isomorphic is using Argon to support drug design research.

Alphabet’s management continues to deploy Gemini models across Alphabet’s entire advertising infrastructure; the Gemini models are improving the performance of Alphabet’s advertising business in advertising quality, advertiser tools, and user experiences; in advertising quality, the integration of Gemini models into ads infrastructure is improving how Alphabet decodes user intent and drive ads relevance; in advertising quality, management upgraded product listing ads in 2026 Q3, which drove a 19% increase in satisfaction and a 50% drop in dissatisfaction; in advertising quality, management has delivered 18% incremental advertiser value and an 8% increase in user sentiment on Discover Ads in 2026; in advertiser tools, management is introducing a way to test different budgets and ROI targets across multiple Search campaigns in a single AI Max A/B test; in advertiser tools, management has integrated Gemini-powered AI creative tools into AI Max, and advertisers have generated over 1.5 billion campaign assets in 2026 Q3, up 7x year-on-year; in advertiser tools, Telstra used AI Max to achieve a 52% increase in conversions and 36% reduction in cost per order

We continue to accelerate the deployment of Gemini models across our entire ads infrastructure to boost performance in the three areas I have highlighted before: ads quality, advertiser tools, and AI user experiences.

First, ads quality. The integration of Gemini models into our ads infrastructure is improving how we decode user intent and drive ads relevance. For example, in Q3, we made the biggest update to product listing ads in a decade, adding new AI capabilities to help users make more confident purchase decisions. Already, we have seen satisfaction increase by 19% and dissatisfaction drop more than half. We are also seeing AI models enable better prediction of clicks and conversions, stronger overall performance, and deeper personalization of ads. For instance, through several innovations in personalization modeling, we delivered 18% incremental advertiser value and an 8% increase in user sentiment on Discover Ads in 2026. Overall, AI is accelerating our ads development and drives the majority of our product launches.

Second, advertiser tools. We are ensuring every advertiser, from global brands to small businesses, can seamlessly participate in our AI services. AI Max and Performance Max continue to see strong adoption. To give marketers even more control, we are now introducing a way to test different budgets and ROI targets across multiple Search campaigns in a single AI Max A/B test. To drive better performance, we integrated our AI creative tools directly into AI Max. In Q3 alone, Gemini models helped advertisers generate over 1.5 billion campaign assets, up over 7x year-over-year. For example, by using AI Max to capture long-form and multilingual queries, Telstra saw a 52% increase in conversions and a 36% reduction in cost per order, showing up in key moments however their customers search.

Google Cloud had 103% revenue growth in 2026 Q3 (was 82% in 2026 Q2) driven by growth in AI infrastructure and solutions; Google Cloud is seeing huge demand and management is working hard to bring more capacity online; Google Cloud is seeing strong demand for TPUs across a wide range of customers; Google Cloud now has 18 products with more than $1 billion in annual revenue; more than 2,600 customers have used at least 100 billion tokens in the last 12 months (was 2,000 in 2026 Q2); 75 customers have consumed at least 10 trillion tokens in the last 12 months; nearly 90% of the Fortune 100 are using Gemini Enterprise; customers who used Gemini Enterprise in 2025 Q3 have 4x-ed their spend in 2026 Q3; Google Cloud has sold more than 100 million seats for knowledge work across all its AI services; Gemini Enterprise now includes the Gemini agent that can research and answer questions, write and run code, and automate process flows; BNP Paribas is deploying Gemini Enterprise to 65,000 employees to streamline workflows; Tata Steel deployed 300 specialised AI Agents with Google Cloud in 9 months and cut service times for customers by 50%; Ulta Beauty used Google Cloud to build the Ulta AI shopping assistant, which has a 3x higher conversion rate than traditional browsing; Gemini is driving acceleration for Google Cloud’s database, analytics, and cybersecurity products; Etsy is using Google Cloud’s BigQuery and Icceberg products to perform critical data joins up to 60% faster; 

Google Cloud revenues grew 103%. Our AI infrastructure and solutions are the major drivers for Cloud growth, and our customers see us as the leading provider here, so there’s huge demand. We are working hard to bring more capacity online to meet this long-term demand. At the same time, we are seeing strong demand for our TPU systems across a wide range of customers, and we expect that to continue as we scale…

…Google Cloud now has 18 product lines with more than $1 billion each in annual revenue…

…Over the past 12 months, more than 2,600 customers have used at least 100 billion tokens, and nearly 75 at least 10 trillion tokens on Gemini Enterprise. Nearly 90% of the Fortune 100 use Gemini Enterprise, and customers who use Gemini Enterprise in Q3 2025 more than quadruple their spend with us year over year…

…Across all our AI services, we have sold more than 100 million seats for knowledge work. Gemini Enterprise now includes the Gemini agent that is deeply integrated across our entire Cloud portfolio. It can research and answer questions, write and run code, and automate process flows with one agent, one API, and a single prompt window. It gives companies strong security and cost controls and the widest choice of frontier models. Customers include BNP Paribas, who is deploying Gemini Enterprise to more than 65,000 employees to streamline workflows while maintaining strict enterprise-grade security and data governance. Tata Steel, who deployed more than 300 specialized AI agents in just nine months to enhance manufacturing safety, optimize supply chains, and cut service times for customers by 50%. Ulta Beauty, who built the Ulta AI shopping assistant, resulting in a 3x higher conversion rate than traditional browsing on their site.

Gemini is also driving acceleration for our database, analytics, and cybersecurity products. Gemini can now help users find, analyze, and visualize information from structured and unstructured data. Gemini uses our Knowledge Catalog to find the right sources and definitions of information, improving accuracy, lowering token cost, and making responses more reliable. Gemini uses our unified lakehouse to access information from multiple systems with zero copy. To respond to agentic analysis, BigQuery improved query speed over 40% year-over-year, while reducing query processing cost by over 50% year-over-year. Customers include Deutsche Telekom, which uses BigQuery Cloud Storage and Knowledge Catalog to build a unified data lakehouse across 40-plus legacy on-premise systems. Etsy, which migrated three petabytes of data to unified lakehouse. By using BigQuery and Iceberg, they now perform critical data joins up to 60% faster.

The growing risk of cybersecurity threats is also driving strong demand for our AI-powered security platform. Gemini and Google Threat Intelligence autonomously analyzes millions of external events every day with 98% accuracy. More than 95% of Wiz customers are adopting AI-powered features, and growth continues. These security tools protect critical infrastructure, including financial institutions like VanOorte and Jack Henry, and telecommunication providers such as AT&T. All of this is powered by our leading AI infrastructure, spanning our TPU systems, Axion CPUs, the latest Nvidia GPUs, and our Agent Compute Engine, which is specifically designed for AI agents.

Alphabet’s TPU 8t and 8i have up to 2.7x better training performance and 80% better inference price performance than previous TPUs; Alphabet recently partnered Meta Platforms to deliver TorchTPU, which allows users to train, serve, and deploy PyTorch models on TPUs; nearly 100 customers use TPUs today; Anthropic and Apple are both users of TPUs, with Apple performing pre-training of its foundation models with the latest generation TPUs; new frontier model companies Mirendil and Physical Intelligence are users of TPUs; Alphabet will be delivering TPUs to Oracle

TPU 8t and 8i offer up to 2.7 x better training performance and 80% better inference price performance, respectively, versus our previous generation of TPUs. To make it easy for developers to port workloads to TPUs, we recently worked closely with Meta to deliver TorchTPU, which allows users to train, serve, and deploy PyTorch models on TPUs using the tools they are already familiar with and rely on. Nearly 100 customers use TPUs today, and we are seeing strong and growing demand from many others. In addition to our work with Anthropic, both directly and via Broadcom, and public reference customers like Apple, who are scaling pre-training of their foundation models on the latest generation of TPUs, we are also working with new frontier model companies, including Mirendil and Physical Intelligence, capital markets firms including Citadel Securities and Jump Trading, universities like Caltech, and public sector service providers, including Cassava and Cirrascale. Given continued strong and accelerating demand, we have recently entered into partnerships to deliver TPU systems with both Crux AI and Oracle, with more in the works.

Alphabet had $924 billion in pre-commencement leases, and other obligations of $152 billion, at the end of 2026 Q3; these reflect maximum obligations, and the actual amounts can vary

At the end of the third quarter, we had future purchase commitments of $924 billion in pre-commencement leases and other obligations of $152 billion. These dollar amounts reflect the maximum contractual obligations. However, a number of these agreements have variable pricing and volume flexibility, which could result in lower spend.

Google Cloud had 103% revenue growth in 2026 Q3 (was 82% in 2026 Q2), driven by growth in GCP (including AI infrastructure, AI solutions, and core GCP) and TPU system sales; Google Cloud’s revenue growth was still very strong even if TPU system sales were excluded; Google Cloud operation margin was 36.2% (was 35.6% in 2026 Q2 and 23.7% in 2025 Q3); Google Cloud backlog grew 9% sequentially to $561 billion in 2026 Q3 (was $514 billion in 2026 Q2); most of Google Cloud’s backlog are GCP contracts and just over 55% of the backlog is expected to be recognised as revenue in the next 2 years; a gigawatt of TPU systems appear to cost $35 billion; when Alphabet sells TPUs to 3rd parties, the inventory is built only after a contract is signed; Alphabet receives cash for the TPU system sales after the systems are delivered

The Google Cloud segment delivered record growth in the third quarter. Cloud revenues were up 103% to $30.7 billion, driven by strong growth in GCP, with strength in AI, including both AI infrastructure and AI solutions, and core GCP. TPU system sales were also a meaningful contributor. GCP accounted for the vast majority of the revenue generated in Q3. The Cloud revenue growth rate, excluding the impact of TPU systems, was again incredibly strong. Cloud operating income was $11.1 billion, more than tripling year over year. Operating margin increased from 23.7% in the third quarter last year to 36.2%. The third-party bridging capacity, which we discussed on our prior call, came online only at the very end of Q3 and therefore had limited negative impact to our margin in the quarter. Google Cloud’s backlog was $561 billion in the third quarter, with the majority of the backlog related to typical GCP contracts. We expect to recognize just over 55% of the total backlog as revenues over the next 24 months…

…[Question] If a gigawatt of TPU full system costs around $35 billion for the compute and all the components around it, what percent of that are we actually recognizing as revenue?…

…When we sign an agreement with a customer to sell TPU systems, once we sign that agreement you’ll first see that showing up in our backlog. Then we start building inventory. Think about this from a CapEx versus cash perspective. We build the inventory ahead of getting the revenue stream in. I think I’ve alluded to this on the previous call. You see inventory balances go up, where previously inventory balances primarily were some of our hardware products. Now those TPU systems that are not yet sold will be showing up, so you see cash going out, and we’re building up the inventory base. As we start delivering and billing customers for these TPU systems, then you see cash coming in…

…We procure components to build inventory, we only do that after we have signed contracts. All of the inventory for TPU systems are associated with committed contracts.

Alphabet’s management expects TPU system sales to continue scaling in 2026 Q4, followed by a meaningful ramp in 2027; management expects Google Cloud margins to be pressured in the near term because of the use of bridging compute capacity that will mostly come online in 2026 Q4; management continues to expect 2026 capex to be $195 billion to $205 billion (2025’s capex was $91.4 billion, which was itself up 65% from $55.4 billion in 2024, and 2024’s capex was up 69% from 2023); management continues to expect 2027’s capex to be much higher than 2026’s; management expects Alphabet’s free cash flow to remain under pressure because of AI-related capex, but the AI-related capex is expected to generate an attractive ROIC that meaningfully exceeds Alphabet’s cost of capital

In Google Cloud, we expect TPU system sales to continue to scale in Q4, followed by a meaningful revenue ramp in 2027. With respect to Cloud margins, given the supply-constrained environment, we started utilizing third-party compute capacity at the end of Q3 and will ramp usage in Q4. We expect margin headwinds in the near term as we utilize this capacity…

…Our full year 2026 CapEx guidance range of $195 billion to $205 billion remain unchanged. We continue to expect our CapEx to increase significantly in 2027 and will provide further details on our Q4 call. 

In terms of expenses, the significant increase in our investment in technical infrastructure will continue to weigh on the P&L in the form of higher depreciation expenses and related data centers operation costs such as energy…

…Let me say a few words on how we think about return on invested capital and free cash flow. We are in a unique position to be able to monetize our AI investments across many products that reach billions of users and enterprises. Our AI investments in Search are driving our highest query growth in half a decade. Google Cloud has generated tens of billions of dollars in AI revenues year to date, and Google One is our fastest-growing subscription product, driven by demand for AI plans. Throughout this investment cycle, we expect to generate an attractive ROIC that remains meaningfully above our cost of capital and anticipate generating strong cash flow from operations. Our sizable investments in technical infrastructure and inventory build of TPU systems as we scale that business will put pressure on free cash flows.

Alphabet’s management saw a step-change in the company’s productivity with the introduction of the 3.8 Flash internally; the introduction of Argon internally has led to a huge inflection point in productivity, and Carbon will be a meaningful step up from Argon

[Question] Sundar, on a recent podcast, you said that 2027 could be the golden age of productivity for Google because of the new class of models and how they will integrate with your other products?

[Answer] We definitely started seeing a step change, I would say, with 3.8 Flash internally as our models got better at long-horizon tasks, agentic coding, et cetera. But definitely, it’s seeing a huge point of inflection once we’ve started deploying Argon internally. It is one of the most meaningful internal productivity levers we have seen on software engineering, directly tracking high-quality code that’s getting checked in, the rate at which it’s getting checked in, reducing the number of rollbacks. But more importantly, people are tackling massive problems as well as we are beginning to see innovation on the product and the consumer experience side from deploying those models. But all this happened only about four weeks ago. We are now deploying Carbon internally, and Carbon is a meaningful step up from Argon.

Amazon (NASDAQ: AMZN)

AWS grew 39.3% year-on-year in 2026 Q3 (was 36.7% in 2026 Q2), the 6th straight quarter of acceleration, and is now growing at its fastest pace in 19 quarters; AWS’s backlog is $578 billion in 2026 Q3 (was $496 billion in 2026 Q2), up triple-digits year-on-year; AWS customers are taking up compute capacity as soon as it’s installed; AWS’s AI business has rapidly improving operating margins; AWS’s non-AI (core compute and storage) revenue grew at the fastest rate in several years in 2026 Q3; the operating margin of AWS’s AI business is currently well ahead of where the non-AI business was at a similar age, despite heightened costs and massive depreciation, but the AI business’s current margin is still lower than the overall AWS business; management has continued to accelerate the capex payback for AWS’s AI business and they have even more conviction now compared to a few years ago that the AI investments will yield strong revenue, free cash flow, and ROIC (return on invested capital); management sees AWS having a strong flywheel where its (1) functionality and performance drives customers to run apps and store data on AWS, which is core-AWS (2) leading to customers running AI workloads on AWS where their data and apps are, (3) leading to more usage of AWS because AI workloads require core-AWS; management sees the flywheel as driving attractive long-term economics for AWS regardless of the level of compute supply; management still sees 85% of global IT spend being on-premises and that the spend will flip to the cloud over the next 10-20 years, as AI-driven migration capabilities make the transition easier, and being on the cloud benefits enterprises; overall demand for AWS is exceptionally strong; management sees enterprises as still being early in using inference at scale, but yet inference is already half of AWS’s AI revenue; AWS is now at a $184 billion revenue run rate (was $169 billion in 2026 Q2); management thinks AWS can reach $1 trillion in revenue in the next 5-10 years, and management has a plan to fund this growth; AWS’s AI revenue grew triple-digits year-on-year in 2026 Q3; big chunks of AWS’s capacity continue to be reserved well in advance, and management has line of sight to some of 2029; management has very good visibility into AWS for the next few years

Despite lapping a comparable year-over-year period where we added a lot of capacity, AWS grew 39.3% year-over-year, accelerating for the sixth straight quarter, our fastest growth in nearly five years. Backlog is $578 billion, growing triple digits year-over-year and well-diversified. We’re installing an AI accelerator about every 14 seconds, and almost as fast as we install capacity, customers are gobbling it up. Our core compute and storage revenue grew at their fastest rates in several years.

Our operating margins have long been strong on our core offering, but rapidly improving on the AI side too, well ahead of where they were on core at a similar time of evolution, even with heightened component costs like memory and significant depreciation from the CapEx we’re investing. We continue to accelerate our CapEx payback and have even more conviction than we already had that the investment we’re making in AWS will yield a much larger revenue and free cash flow-generating business than we would have imagined a couple of years ago, with strong return on invested capital…

…We are happy with the progression in both core and AI margins, and as Andy said earlier, they are ahead of where the core was at a similar point in time, but they are still lower, so there is going to be that mix issue….

…In AWS, having the leading functionality, security, and operational performance can’t help but drive customers to run their apps and store their data in AWS, which can’t help but drive AWS AI demand because customers prefer their production inference to be near the rest of their data and apps, which can’t help but drive core AWS demand because AI uses core for work like reinforcement learning, post-training, and tool usage, which can’t help but drive reinvestment back into our broader AWS functionality…

…It also yields very attractive long-term economics for us, regardless of supply scarcity…

…On the core side, it’s important to remember that the vast majority of global IT spend, about 85%, is still on premises, and we’re in a 10- to 20-year cycle where that equation will flip to the cloud. Conventional wisdom has been that linchpin workloads like mainframes or storage area networks would never move to the cloud, but these workloads are moving as AWS has built unique cloud substitutes and AI-driven migration capabilities to make this transition much easier. Enterprises get the benefits of less CapEx, lower operating costs, faster innovation, better developer productivity, and data accessibility to more easily leverage AI…

…Overall demand for AWS capacity remains exceptionally strong. Enterprises are still very early in using inference at scale, and yet inference already accounts for well over half of AWS’s AI revenue and growing, and we have a lot of training revenue as well…

…AWS is a $184 billion annualized revenue run rate business, and yet the opportunity remains extraordinary…

…We believe that AWS can grow north of 40% in each of 2027 and in 2028 and be a trillion-dollar revenue business in the next five to 10 years…

…Would it be more convenient if this type of acceleration and opportunity happened over 25 years and you could spread the CapEx over 25 years? Yeah, it might be more convenient, but that’s not the world that we live in, and that’s not the opportunity. We can fund this plan. I think it will forever change the size of Amazon…

…Our AI offerings are gaining traction with customers, with AI revenue growing triple digits year-over-year…

…Despite how much capacity we are providing, the big chunks of capacity that people take, the hundreds of megawatts or sometimes up to a gigawatt of capacity, those big chunks are reserved well in advance because you cannot leave a gigawatt of capacity hanging around. We have very good visibility into 2027 and 2028 and some of 2029 too, with respect to what customers want in those really big chunks of capacity…

…We have very good visibility into the next few years on what AWS looks like. 

Amazon has acquired DuckLabs, and DuckDB will become the SQL dialect of AWS Storage and Analytics; DuckDB will enable customers, for the 1st time in analytics, to enjoy zero migration or portability costs between query engines; management thinks the zero migration and portability cost feature will help AI agents to efficiently query data

With our acquisition of DuckLabs, DuckDB will become the universal SQL dialect of AWS Storage and Analytics, simplifying data processing with a modern analytics engine that is fast, simple to use and ubiquitous, and where customers will be able to write SQL queries once and then execute using any AWS analytics service. For the first time in analytics, customers will have zero migration or portability costs between query engines. This is a big deal, and you can imagine how it will also help agents do efficient querying as well.

Bedrock is AWS’s fully-managed service for companies to build upon frontier models; Fortune 500 companies are spending 7x more on Bedrock year-on-year in 2026 Q3 with token volume growing 17x; Bedrock is the fastest growing multi-billion dollar service in AWS’s history; Bedrock recently added managed versions of the latest Kimi, GLM, and SpaceX models; the data of customers that use Bedrock is secure and private; management thinks that the future of AI will (1) consist of 6-8 models that are good enough for nearly all workloads, and each of them will have their strengths, and (2) have customers that use multiple models; management sees Bedrock has being relevant for the future of AI; management thinks the Western world are very hungry for Western open sourced models; most of Bedrock’s workloads today are being powered by Trainium

We continue to expand the capabilities of our inference service, Amazon Bedrock, and it is leading to remarkable growth, with Fortune 500 companies spending nearly 7x more on Bedrock year-over-year and total Bedrock token volume growing nearly 17x year-over-year. Bedrock is the fastest growing multi-billion dollar service in AWS’s history. We also just added Bedrock-managed versions of Kimi, GLM, and SpaceX models…

…We have designed Bedrock so customers’ data remains secure and private. Inputs and outputs are never shared with model providers or used to train base models. When customers tune a foundation model, we create a private copy specifically for their use, and data is fully encrypted both in transit and at rest…

…One of the lessons that we’ve learned, I’ll say for 20+ years in offering technology to customers, is that there is not one tool to rule the world. People always assume that whatever is in the lead early on is going to be the only thing or things that people use, and it’s never the case. It’s not true in database, it hasn’t been true in analytics. In the beginning of AI, everyone said the only framework people use is TensorFlow. It turns out PyTorch is much bigger now. It’s not going to be true with models either.

I’ll say in the next two to three years, I think there are going to be 6- 8 models that are about as good as each other, or at least good enough for 95%+ of the workloads. Some will be better at reasoning, some will be better at coding, some will be better at visual pieces, some will be lower latency, and some will be lower cost. Most customers are going to use multiple models, and many will use multiple models in the same generative AI application. What customers really want to do, and we see this validated over and over again when we talk to customers, is they want a service that has really broad selection of those frontier models, and that has very good price performance, and that has very strong security and access control. They want that service so that they can decide what combination of models they want to use and to swap them back and forth as they leapfrog each other, as different constraints of their applications require them. That’s what we see with Bedrock…

…I think the Western world is very hungry for additional open models that are Western open models…

…The majority and most of Bedrock today is being run on top of Trainium. We do that in meaningful part because we really like the price performance that we are getting, and it allows us to both keep prices as low as possible for customers, as well as our ability, we like the margins associated with it.

Amazon’s management is seeing frontier models surface previously unknown software vulnerabilities, and at a much higher frequency; management has built AWS Continuum for customers to address software security; AWS Continuum already has established enterprises as customers; AWS Continuum is being integrated into popular vibe coding tools such as Claude Code, Codex, and Kiro; AWS Continuum is enjoying unprecedented customer interest for a new security service; some people think 20% of all AI spending in the next 1-2 years could be in security

As frontier models get more powerful, they are surfacing software vulnerabilities previously not found and at unprecedented rates. We built AWS Continuum to discover, prioritize, validate, and remediate code vulnerabilities at machine speed. It is already being used by customers like Capital One, IBM, Infor, Luminate, and Robinhood. In Q3, we announced that we will integrate it into the tools where code is being written, Anthropic’s Claude Code, OpenAI’s Codex, and AWS’s Kiro. The interest from customers exceeds anything we have ever seen for a new security service…

…There are some people who think that the total amount of tokens used over the next year or two, 20% of it is some significant percent of it may end up being just security.

AWS customers are using AgentCore to deploy agents at an accelerating rate; AgentCore has tens of thousands of customers, and they deployed 2x more agents in 2026 Q3 than the prior 3 quarters combined; management is seeing companies build their own agents and use turnkey agentic services; AWS has agentic services that customers love; management thinks every modern enterprise production application will be rewritten to be agentically capable

Agents continue to be deployed at accelerating rates, many of which are being built on AgentCore’s building blocks. Tens of thousands of customers currently use AgentCore, deploying 2x more agents with it in Q3 than the prior three quarters combined. In addition to building their own purpose-built agents, most companies will also use turnkey agentic services, and we offer several that customers love. Kiro for coding, Amazon Connect for contact center agents, and AWS Transform for migrations. They are all iterating quickly and picking up customer momentum. As is Amazon Quick, our intelligent AI work companion, which is growing very rapidly as large enterprises like Mondelez, EchoStar, 3M, Southwest Airlines, and DXC Technology are choosing Quick because it is purpose-built for work with no compromises on customer data…

…I believe virtually every modern enterprise production application will be rewritten to be agentically capable, because that’s the way you’re going to have to deploy if you want to move quickly or you want to actually guard in the right way against some of the security risks.

AWS’s chips business grew triple-digits year-on-year in 2026 Q3; revenue commitments for Graviton and Trainium in 2026 Q3 both increased more than 20x year-on-year; management sees AI as being a tailwind for AWS’s Graviton CPU chip, because AI workloads, including agent tool use, requires CPUs; management sees Graviton as the strongest CPU chip for AI, offering 30%-40% better price performance than alternatives; customers using AWS’s 1st-party chips saved more than $2 billion in 2026 Q3; management recently open sourced TorchNeuron for developers to run existing PyTorch code on Trainium without rewriting code; the backlog for prior generation of Graviton chips still grew more than 100% sequentially in 2026 Q3 even after Graviton 5 was launched; short-term capacity for prior generation Trainum chips are selling at a premium compared to 2026 H1; management intends to sell Trainium to 3rd parties outside of AWS in late-2027; Amazon’s chips business is helping AWS to optimise capacity and improve its cost efficiency; management is seeing that enterprises are eager to save on price-performance as they deploy inference into production, and this is where Trainium has an advantage; management believes that lowering the cost of inference is important for enterprises to start deploying inference at scale, and Trainium is an important part of this view; selling Trainium externally will lead to more revenue in the short term, but selling Trainium through AWS will lead to more absolute revenue and free cash flow over time; managemen thinks selling chips is a very different business than running a cloud

Our chips business continues to grow triple-digit percentages year-over-year, and revenue commitments for Graviton and Trainium have each increased more than 20x year-over-year. AI continues to be a tailwind for Graviton because post-training reinforcement learning and agent tool use is mostly done on CPUs versus AI accelerators, and Graviton is the strongest CPU chip, offering up to 30%-40% better price performance than alternatives. Customers using our custom silicon saved more than $2 billion in Q3 alone, and we just open sourced TorchNeuron so developers can run existing PyTorch code, which most already use on Trainium without rewriting a thing. As we launch new generations of chips, demand for prior generations remains strong. After our launch of Graviton5, backlog for prior generation Graviton instances still more than doubled quarter-over-quarter. Short-term capacity for prior generation Trainium instances sell at a premium compared to the first half of the year. With Trainium demand so strong, customers continue encouraging us to make it available outside of AWS. We intend to do so late in 2027…

…Our investments in custom chips and custom networking, along with software and process improvements, are helping us optimize capacity and build a more cost-efficient infrastructure…

…Increasingly, what you’re finding right now is most customers, every customer that uses AI at scale, but particularly enterprises now as they’re starting to deploy more inference into production, they are very eager to save on price performance. Trainium, with the meaningful advantage it has in price performance versus other alternatives, is very compelling to people…

…I think if we all want enterprises to be able to use inference the way we all believe they will, we have to keep working on taking the cost of inference down. That is a huge mission for AWS. It’s one of the significant reasons why we’re building our own custom silicon in Trainium, is that we just really believe it’s incredibly important for customers to be able to use inference in a much more cost-effective way…

…If you sell externally, you are able to realize more revenue in the short term because you get to recognize the revenue as soon as it leaves your inventory. On the flip side, if you decide to use it inside your cloud in the form of EC2 instances or Bedrock, you actually will end up driving more absolute revenue and free cash flow from those chips because you get to use them over five to six years and drive revenue for a longer period of time…

…We believe we are going to have a very significant business selling chips externally, and we are going to start in late 2027. We will be thinking through that balance, I think, probably every year. They will both be very significant. Our ability to create price, performance, scalable chips is going to be a very significant strategic advantage for us at scale, both in our own cost structure as well as in building a separate business and selling external chips. Selling external chips really is a very different business than your cloud infrastructure business.

Amazon’s management continues to expect capex for 2026 to be $220 billion (capex was $128 billion in 2025, and $83 billion in 2024); the majority of Amazon’s capex is for AWS; management expects capex to increase meaningfully in 2027; management has clear line of sight to strong financial returns on the company’s capex

We still expect 2026 cash CapEx of approximately $220 billion, the majority for AWS, and expect to spend meaningfully more in 2027 with a clear line of sight to attractive financial returns,

Robotics and AI are delivering measurable gains in Amazon’s fulfilment network; Amazon doubled its fleet of robotic arms in 2026; Amazon is developing foundation models for robotics to deliver useful robotics at scale; Amazon has over 1 million robots in its fulfillment network; Amazon’s safety record has improved in the past 4-5 years, driven by robotics; management sees the opportunity to deploy robots for delivery for the last 100 yards; Amazon’s new designs for its fulfilment network are robotics-heavy

Robotics and AI are delivering measurable gains through productivity and employee safety, which we’ll expect to compound in the future. In 2026, we doubled our fleet of robotic arms like Cardinal and Sparrow. We’re also developing robotics foundation models that enable faster programming and teach robots increasingly complex physical tasks, with the goal of delivering useful robotics at scale. Looking ahead, we see meaningful opportunities to improve productivity across our global fulfillment network and lower our cost to serve while continuing to get products to customers faster…

…We have over 1 million robots now in our fulfillment network, and while they’re having very significant impact on our cost structure and on our productivity and on safety. If you look at our safety numbers in our fulfillment network, they’re at record low numbers. We’ve made huge progress over the last four to five years, and robotics has been a big piece of that…

….I think we have the opportunity to also assist on some of the delivering of the last 100 yards to customer doors…

…If I look at the new designs that we have coming in our fulfillment network, including our expansion of same-day facilities, they are very heavy on robotics.

Over 400 million Amazon customers (350 million in 2026 Q2) have used Alexa for Shopping (previously known as Rufus) in the last 12 months; customer interactions with Alexa for Shopping was up 570% year-on-year in 2026 Q3 (was up 5x in 2026 Q2); US customers using Alexa for Shopping spend 50% more per order than those who don’t; management anticipates working with other agents (Meta Platforms’ Muse?), but they think Alexa for Shopping will be the primary shopping agent on Amazon

Over 400 million customers have used Alexa for Shopping, our agentic shopping assistant, in the last 12 months, with customer interactions up 570% year-over-year. U.S. customers using Alexa for Shopping are spending 50% more per order than those who don’t…

…Though we anticipate working with horizontal agents, we believe Alexa for Shopping is going to be the primary shopping agent customers use to shop Amazon.

Amazon’s management thinks agentic shopping from horizontal agents has no meaningful impact on Amazon’s advertising business; management will soon launch Amazon Ads Agent, which brings all of Amazon’s advertising offerings into a single AI-first platform; management has launched full-funnel campaigns in Amazon Ads Agent, where advertisers provide initial inputs, and Amazon’s AI does the rest

The natural question is whether agentic shopping from external horizontal agents is meaningfully impacting Amazon Ads. The answer is no. Few customers are doing much shopping this way yet, as the experience is still not great nor more convenient than coming to a retailer directly to research or one-click order there…

…Headlined by Amazon Ads Agent, which brings together our full advertising offering under a single AI-first platform. Backed by trillions of shopping, browsing, and streaming signals, it supports advertisers at every step with insights, answers, and the ability to optimize and act on their behalf. We also launched full-funnel campaigns in Amazon Ads Agent, bringing sponsored ads, display, video, and streaming TV into a single campaign. Advertisers provide initial inputs, including budget and duration, and Amazon’s AI plans, executes, and continually optimizes the rest.

Intel (NASDAQ: INTC)

Intel’s Granite Rapids CPU is ramping really fast while the Clearwater Forest CPU is shipping for revenue; Intel’s CPU products is seeing significant demand for 2027, driven by the compute needs of agentic infrastructure; Intel’s Diamond Rapids CPU is on track to ramp in 2027; Diamond Rapids is seeing strong interest from customers for inference and agentic workloads; Intel’s CPUs will be deployed by Oracle in its AI buildout; Intel is facing some headwinds in meeting the extremely high AI-driven demand for its CPU products; the headwinds include an industry-wide shortage of advanced organic substrates for EMIB (embedded multi-die interconnect bridge) packaging, and a factory incident in Israel; management sees CPUs as far more efficient at performing tasks compared to GPUs; management sees the TAM (total addressable market) for CPUs growing from $40 billion today to $200 billion in 5 years, because of agentic workloads; management sees competition in the CPU market for Intel, but thinks Intel has unique strengths, namely, (1) strong pipeline of CPU innovation, (2) willingness to customise CPUs for custom workloads, (3) the only advanced semiconductor fab in the USA; management sees demand signals from customers continue to strengthen, but supply is expected to be constrained up to 2028; Intel’s CPUs play a central role in enabling agentic workloads, and management expects growth of more than 30% in its server CPU units in both 2026 and 2027, with higher growth to be achieved if supply were available; management expects the DCAI (data center and AI) segment to achieve revenue in 2026 that’s $10 billion higher than initial expectations, driven by broad-based growth and purpose-built silicon

Within our server business, we are making steady progress on our roadmap. Granite Rapids on Intel 3 has become one of the fastest-ramping server products. Clearwater Forest, our first Intel 18A server products, is shipping for revenue. Demand for 2027 is significant, especially in CPU dense agentic rack scale infrastructure. Diamond Rapids, our first Intel 18A-P server product, is sampling and on track to ramp in 2027 with strong interest for inference and agentic workloads. We were pleased to recently announce an agreement with Oracle Cloud Infrastructure to deploy our latest generation CPUs to support their AI build-out.

I also want to acknowledge headwinds limiting our progress to meet this extremely high AI-driven demand. First, Intel 3 yield and cycle time continue to improve, but I am demanding significant execution from the team. Second, even as our Intel 3 wafer outputs ramps quickly, an industry-wide shortage of advanced organic substrates for our EMIB packaging is also impacting our ability to scale output. Finally, early this month, an unfortunate and tragic accident led to an extended shutdown in our Israel factory. While the teams have worked quickly to address the situation, it will impact our Q4 10-nanometer server shipments…

…Increasingly, energy efficiency is becoming the most important metric for the AI infrastructure build-out. While GPUs are great at thinking tasks, deterministic CPU are far more efficient at doing tasks. While GPU reasoning query run as high as multiple watt-hours of energy, a single CPU query could cost only joules of energy. Applying even conservative estimates to the total number of digital agents and agents served per CPU, it is easy to anticipate a massive dollar and unit TAM for CPUs over the coming five years. By our estimates, server CPU TAM, largely driven by agentic workloads, could increase as much as five times from where it is today, from $40 billion to more than $200 billion over five years…

…We are, of course, mindful of fierce competition underway as nearly every fabless company race to meet this unprecedented demand with their own offering. But Intel is uniquely positioned to benefit from this demand in three ways. First, as a custodian of x86, the most widely deployed ISA in the world, we are doubling down on x86 innovation, have an exciting Tick Tock roadmap of CPU innovations in the pipeline. In addition, for the first time, we are making fully modular, verified instances of our x86 CPUs optimized for the custom workloads available to our customers. Second, as one of the just three advanced node semiconductor foundries in the world, and the only one domiciled in the U.S., we are well positioned to support not only our own demand but also the demands of wider fabless industry. We are also open to building any CPUs at scale for our customers fully agnostic to underlying ISA. Third, as modern computing platforms increasingly become complex system in the package, our leadership in advanced packaging offers highly differentiated solution, especially to meet the requirements of the massive AI infrastructure build-out underway…

…Demand signals from our data center customers continue to strengthen. We expect the industry to remain supply constrained throughout 2027 and into 2028…

…The adoption of agentic AI is broadening the role of CPU compute. AI workloads are moving beyond prompt-based responses toward multi-step reasoning, orchestration, and autonomous persistent action. Our CPUs play a central role in enabling these workloads. As a result, we continue to see strong demand for our DCAI products. We now expect industry server CPU units to grow more than 30% in both 2026 and 2027. Growth would be higher in both years if additional industry supply were available…

…Based on our current outlook, DCAI revenue is on track to finish 2026 nearly $10 billion above expectations entering the year. Q3 growth was broad-based across hyperscale, enterprise, and Neo Cloud customers. We also continued to see strong momentum in purpose-built silicon, with revenue increasing more than 50% sequentially and more than doubling year over year.

Intel’s management sees a tripling of Intel’s purpose-built silicon business over the next few years after tripling over the last 3 years 

Purpose-built silicon remains our fastest-growing business segment, tripling over the last three years and poised to triple again over the next few years. We see further growth opportunity with both existing and new potential customers, including AI labs, and expanding our IP portfolio to meet the needs of a wider array of customers…

Intel’s management recently announced the company’s D-Series inference-optimized GPU; the new GPU has encouraging early feedback and performance is tracking well

We recently unveiled our Intel GPU D-Series inference-optimized GPU. Early feedback is encouraging. Performance is tracking well. This is new market for us, and it will take time to scale, but customer engagements are progressing towards firm commitments.

Intel’s PC (personal computer) business is seeing a slowdown because of memory prices and a shortage of memory chips because of AI-related demand for memory chips; management expects the slowdown in the PC business to last for a few quarters

In recent weeks, our customers began signaling a PC slowdown based on the memory prices and availability. Year-to-date system endpoint prices have increased significantly, placing additional pressure on already limited IT budgets. The pace of this PC correction has accelerated over the last two months as our OEM partners are aggressively managing inventory to match the softer sell-through. We expect the current environment to persist until memory availability and prices begin to normalize. While the next few quarters will be challenging for the PC market, our roadmap continues to strengthen, and our long-term outlook remains positive…

…The industry continues to navigate one of the most significant supply shortages in the history, driven by the unprecedented demand for super intelligence.

Intel Foundry’s Intel 3 and Intel 18A yields both increased in 2026 Q3, with Intel 18A’s yield improving at a double-digit rate; wafer output of both Intel 3 and Intel 18A exceeded management’s expectations; Intel 18A-P is on track for launch by end-2026 for Intel’s internal products; management is engaged with a few highly strategic customers on Intel 18A-P; management is confident Intel 18A-P will be an attractive node for external Intel Foundry customers; Intel 14A is on track; management recently released Intel 14A’s 0.9 PDK (process design kit); management is pleased with Intel 14A’s progress with multiple potential customers; management is pleased with Intel Foundry’s progress with EMIB and EMIB-T for advanced packaging for internal use; both EMIB and EMIB-T have significant external customer traction; management is seeing robust output for HBI (hybrid bonding interconnect); Intel Foundry is bringing additional capacity online, ithe US and Asia; Intel Foundry is still overwhelmingly serving Intel and not external customers; Intel 3 costs improved 15% from 2026 Q1, while volume increased mid-teens sequentially; Intel 18A costs have fallen by 40% since 2026 Q1, and its yield has improved ahead of internal projections; management expects Intel 3 to be a profitable node for Intel Foundry in 2027; management thinks Intel Foundry will have some external customers to announce in the next 2 quarters for advanced packaging

Intel Foundry execution continued to improve consistently. During the quarter, Intel 3 and Intel 18A yields both increased again, with Intel 18A improving at a double-digit rate. Wafer output on both nodes exceeded expectations. Intel 18A-P remains on track for launch by year-end across our internal product portfolio, and external customer engagement continues to build. We are engaged with a couple of highly strategic customers, and I am increasingly confident that our Intel 18A-P will be an attractive node for external foundry customers. Intel 14A also remains on track. In October, as committed, we released our industry-standard 0.9 PDK to select customers. It is our highest quality 0.9 PDK, meeting industry standard yield and performance criteria at release. As is standard practice in the foundry business, I have consistently maintained that we will not disclose customer details. That being said, I remain highly encouraged by the progress we are making with multiple potential customers on 14A.

On advanced packaging, we are pleased by the progress we are making with EMIB and EMIB-T for our internal products. Both continue to see significant external customer traction as well. We continue to improve on execution, working with our supply chain to secure the capacity to meet the needs of our advanced packaging customers. We are also seeing robust progress in the readiness of HBI, which is an enabler of 3D chip stacking for our upcoming server products. Across our manufacturing network, additional capacity continues to come online. Intel 3 tool installation is progressing well in Ireland, as is in the ramp of the 18A in both Oregon and Arizona, including Fab 62, our second high-volume 18A fab. In New Mexico and Malaysia, we are adding capacity to support strong customer ramps for advanced packaging next year. We are also seeing robust progress in the readiness of HBI, which is an enabler of 3D chip stacking for our upcoming server products.

Across our manufacturing network, additional capacity continues to come online. Intel 3 tool installation is progressing well in Ireland, as is in the ramp of the 18A in both Oregon and Arizona, including Fab 62, our second high-volume 18A fab. In New Mexico and Malaysia, we are adding capacity to support strong customer ramps for advanced packaging next year. In New Albany, Ohio, the team recently achieved important construction milestone with installation of the first major structural truss. And we continue to evaluate unique and highly differentiated opportunity to accelerate our Ohio build-out…

…Intel Foundry revenue was $6.7 billion, up 16% sequentially, driven by better-than-expected growth on Intel 3 and Intel 18A. External foundry revenue was $411 million, an increase of $118 million sequentially, primarily due to a non-recurring customer program…

…Intel 3 costs improved approximately 15% from Q1 levels, while volumes increased in the mid-teens sequentially. Intel 18A costs have declined approximately 40% since Q1, reflecting continued yield and cycle time improvements. Intel 18A yields continue their upward trajectory to industry-standard levels ahead of our internal projections…

…If you look at Intel 7 today, kind of 10 nanometer 7, that’s already a profitable node to us. Intel 3 is going to be profitable next year for us…

…We are going to have couple of customer ready to announce in next two quarter. Clearly, we always want to wait for the customer, if they want to announce, we will respond to that.

Intel’s management continues to expect capital expenditures to be flat in 2026 (around $20 billion); management expects capital expenditures to be up strongly in 2027; management has increased firm purchase orders to equipment suppliers by 40% sequentially in 2026 Q3; despite the expectation of much higher capital expenditure in 2027, management is still committed to deleveraging in 2027; management will only spend on capital expenditure if there’s clear evidence of customer demand; most of Intel’s capital expenditure will be on equipment

Our 2026 gross capital expenditures remains on track to be $20 billion, and we expect 2027 gross capital spending to be up strongly. We have made good progress securing the equipment required to support our growth plans with firm purchase orders to equipment suppliers increasing nearly 40% sequentially. We’re also accelerating clean room build-outs, including Fab 62 in Arizona, and exploring opportunities to accelerate in Ohio. Despite this increase in capital spending and working capital, we remain committed to deleveraging in 2027. We will remain disciplined as we invest, and we will not spend a dollar of capital unless we have confidence in the health of our products and process technologies and clear evidence of customer demand…

…We have, I think, a significant ramp in equipment spend this quarter, and it is expected to be even more significant next quarter. A lot of what we are spending on now is really equipment. We have made the investments. I would call it, in a lot of cases, putting the finishing touches on some of the space and most of the spend is going towards equipment, with the only exception, we are still in the heavy construction phase of Ohio. But outside of that, it is really about equipment.

Mastercard (NYSE: MA)

Mastercard’s new Digital Risk Protection product helps identify vulnerabilities in an organisation’s digital environment; Digital Risk Protection has nearly 800 enterprise customers after being launched for a few months; Mastercard’s enhanced risk offering combines RiskRecon and Recorded Future’s capabilities (Recorded Future was acquired by Mastercard in 2024 Q4 and it provides AI-powered solutions for real-time visibility into potential threats related to fraud); management is seeing AI change the demand for Mastercard’s security solutions

Within cybersecurity, we continue to scale one-to-one platform sales, but even more so new network products like our newly launched Digital Risk Protection. This capability helps organizations identify vulnerabilities in their own digital environments. After only a few months in market, it now serves nearly 800 enterprise customers. Additionally, our enhanced third-party risk offering brings together RiskRecon and Recorded Future capabilities in one solution. It helps identify, prioritize, and respond to risks before they become business disruptions. AI is not only changing the demand for our security solutions, it’s also changing how commerce happens.

Mastercard’s management is seeing how AI change the way commerce happens; management sees 2 forms of agentic transactions, machine-to-machine, and traditional e-commerce; management thinks both forms of agentic transactions are significant opportunities for Mastercard; management thinks agentic transactions could eventually lead to an entirely new addressable market of micro-transactions; management sees consumers starting to use agents to discover products and make purchases; it’s still early days in consumers’ using agents for purchases, but management expects the trend to ramp up in the holiday season; management is seeing agentic transactions ramping up at impressively weekly growth rates; management believes 4 things are necessary for agentic commerce to function at scale, namely, trust in the payment, the agent, the ecosystem, and ubiquitous reach; management sees Mastercard as having solutions to solve for all 4 elements; management is helping merchants be ready to accept agentic commerce with the new Agent Connect solution; Shopify’s Shop Pay is already using Agent Connect’s features; Mastercard signed a partnership with OpenAI in 2026 Q3 for Agent Connect; management thinks that agentic commerce solutions that take VCNs (virtual card numbers) are not a good long-term solution

AI is not only changing the demand for our security solutions, it’s also changing how commerce happens. The foundation I described earlier, broad participation, global acceptance, and trusted capabilities, positions us to help shape the next evolution of commerce. That includes agentic commerce, where a buyer uses AI agents to discover, select, and purchase a product or service. There are two forms, machine-to-machine and traditional e-commerce. Combined, they create significant opportunity for Mastercard, from more transactions to new B2B carded payments as businesses utilize agents. Eventually, an entirely new addressable market of micro-transactions, all the while creating more opportunity for our services from fraud to longer-term potential in insight and loyalty. Consumers are already beginning to use agents to discover and purchase products. We expect that trend to build as holiday shopping ramps up and to continue well beyond the season. It is still early days, and the numbers are not yet significant in the grand scheme. As the AI agentic assistants have emerged, agentic transactions are ramping up at impressive weekly growth rates. We’re now seeing the early stages of consumer adoption.

For this type of agentic commerce to function at scale, we need four things. Trust in the payment, the agent, the ecosystem, and ubiquitous reach. Mastercard has built capabilities for each of these four. For trust in the payment, tokenization is embedded in Agent Pay and backed by the franchise rules everyone expects today. For trust in the agent, Agent Pay ensures agents are verified and authorized by the buyer and that the merchant and issuer are aware. For trust in the ecosystem, we bring transparency and continuity through the payment flow. Verifiable Intent captures the buyer intent, and our recently launched Agentic Trust Intelligence solution helps banks and merchants evaluate transactions initiated by any AI agent. That allows these transactions to occur wherever Mastercard is accepted. That is the unparalleled and growing reach I talked about earlier.

At the same time, we are helping merchants ready themselves and be more discoverable to agents with our new Agent Connect solution. Many partners are putting all of this into practice. For example, Shopify, through Shop Pay, is using Agent Pay and our network tokens to make agentic payments more secure and seamless by preserving the integrity of the consumer’s credential end-to-end. Shop Pay is already available on multiple agentic platforms with more to follow. With OpenAI this quarter, we signed a broad strategic partnership that includes the use of Agent Pay and goes well beyond. Our agreement establishes collaboration across payments, joint services growth, and AI transformation…

…It is in everybody’s best interest to use Agent Pay and equivalent. We are not supporting solutions that take a one-time VCNs out there as you’re fronting a transaction. That fundamentally breaks a transaction that is important from a consumer perspective. They made a choice. It’s important from a rewards perspective, from a control perspective, and so forth. I think we’re at this turning point where everybody is pushing in the same direction. We are very clear that fully tokenized all the way through agentic pay or an equivalent is the way to go.

Microsoft (NASDAQ: MSFT)

Microsoft’s management announced the biggest update to Copilot in 2026 Q3 (FY2027 Q1); management thinks of Copilot as the OS for work; automated tasks on Copilot has nearly doubled since July; daily Copilot users are up 50% sequentially in 2026 Q3 (FY2027 Q1) and up 5x year-on-year; user satisfaction of Copilot is up 20% sequentially in 2026 Q3 (FY2027 Q1); 90% of Copilot queries are on auto mode, which dynamically routes tasks across models; customers with more than 50,000 Copilot seats increased 8x year-on-year in 2026 Q3 (FY2027 Q1); Copilot is helping to drive momentum with E7; management is bringing Dynamics 365’s business logic and capabilities into Copilot; management introduced 13 new Copilot skills across Dynamics 365 in 2026 Q3 (FY2027 Q1)

We announced our biggest update to Copilot to date this quarter, bringing together chat, co-work, autopilot, and code into one unified experience. It’s one product experience across commercial and consumer and comes with all of Office built in. We think of Copilot as the new OS for work. Customers are increasingly turning to Copilot to get real work done. Automated tasks has nearly doubled since July. Daily users are up nearly 50% quarter-over-quarter and over 5x year-over-year. User satisfaction improved nearly 20% quarter-over-quarter to another all-time high, and 90% of Copilot queries use auto mode, which dynamically routes tasks across models optimizing for quality, speed, and cost. These improvements are translating into larger purchases. The number of customers with more than 50,000 Copilot seats increased over 8x year-over-year, including Capgemini, HCLTech, Hitachi Group, Truist Financial, and U.K. Revenue and Customs Department. Just last week, Volkswagen announced plans to deploy Copilot to more than 330,000 employees, one of the largest deployments to date in Europe.

We’re also seeing strong momentum with E7, which brings together Copilot, E5, Entra, and Agent 365. Thousands of customers in every industry have purchased the suite to date, including Amadeus, BlackRock, Fiserv, New South Wales Health Department, and ZEISS…

…We are extending this vision of Copilot as the new OS for work, bringing the business logic and capabilities of Dynamics 365 directly into Copilot and helping every organization stand up their own infinite SaaS factory. This quarter, we introduced 13 new Copilot skills across Dynamics 365, which enables customers to access CRM and ERP capabilities directly in the flow of work.

Microsoft’s business model now includes consumption, which management thinks expands the company’s TAM (total addressable market) and average revenue per user (ARPU); management thinks the consumption model opens up business-funded demand, compared to the subscription model, which addressed IT-funded demand; Copilot customers on usage-based billing grew 75% sequentially in 2026 Q3 (FY2027 Q1); credit consumption of agentic features on Github has more than 2x-ed since usage-based billing was introduced

Our business model is expanding from per-seat subscriptions to include consumption, expanding our TAM and ARPU. The subscription brings increasingly capable everyday intelligence to every user. Typically, IT funds this, consolidating frontier advances for everyone across the organization. As agents take on higher-value work, consumption opens up business-funded demand, with spending tied directly to the value these agents deliver. The number of Copilot customers actively consuming credits through usage-based billing grew over 75% over the quarter…

…Credit consumption from agentic features on GitHub has more than doubled since usage-based billing went into effect.

Agent-generated PRs (pull requests) on Github have grown 6x in the 6 months ended 30 Sep 2026; agents now open 1/3 of all PRs on Github; management is rebuilding the core Github architecture to address agentic use; management recently announced Copilot Fusion on Github; Copilot Fusion brings multi-model orchestration capabilities, delivering frontier quality up to 40% faster and at 77% lower costs; Github Advanced Security revenue grew 50% year-on-year; Copilot Fusion is ranked highly by Artificial Analysis’s latest benchmark for coding agents

When it comes to software development, all major coding agents have shown up to work on GitHub. Agent-generated PRs have increased 6x over the past six months, and agents now open more than a third of all PRs on the platform. We are very hard at work on reliability as agentic activity on GitHub soars. At GitHub Universe this morning, we shared how we are expanding capacity and rebuilding the core Git architecture for the agentic wave. Also at Universe, we announced broad availability of Copilot Fusion, which is all about bringing multi-model orchestration to software development. It delivers frontier-level quality up to 40% faster and at up to 77% lower cost…

…GitHub Advanced Security revenue increased over 50% year-over-year as customers use its AI capabilities to scan entire repos and fix vulnerabilities in real time…

…In fact, this morning, Artificial Analysis had what I think is the real benchmark, which is the coding agents using our Fusion, and I think we were number 2 or number 3 there, which is awesome to see. Which is you can now get frontier performance 45% faster, 75% cheaper. Who would want to not take that?

Foundry is becoming more important to Microsoft’s growth; Foundry customers reached 120,000 in 2026 Q3 (FY2027 Q1); Foundry revenue more than tripled year-on-year in 2026 Q3 (FY2027 Q1), driven by efficiency gains across Microsoft’s GPU fleet; Foundry helps customers optimise across quality, cost, and performance across models; the number of Foundry customers in 2026 Q3 (FY2027 Q1) using multiple models to build agents grew 5x year-on-year; Foundry’s revenue from open source models increased 10x year-on-year in 2026 Q3 (FY2027 Q1); Foundry customers can use Microsoft’s harness behind Copilot; Rolls-Royce used Foundry to build an agentic system to manage its supply chain; management thinks enterprise AI workloads will be very dynamic in terms of switching between models

It starts with Foundry, which is becoming an increasingly important driver of our growth. Foundry helps organizations build learning loops grounded in their business context and deploy agents safely at scale. We now have over 120,000 Foundry customers, with revenue more than tripling year-over-year. Foundry’s model choice and feature set, coupled with its deep integration across our stack, is driving differentiation. Azure OpenAI directly benefited in the last quarter with big advances in OpenAI models. When you couple this with our model choice built into Foundry, customers are able to build solutions optimizing across quality, cost, and performance. In fact, the number of customers building agentic solutions with multiple models increased 5x year-over-year, and revenue from open source models on Foundry increased 10x…

…Foundry revenue more than tripled year-over-year, with upside driven by efficiency gains across our GPU fleet that allowed us to allocate additional capacity to customers, as well as higher revenue per token that we benefited from in the quarter…

…Beyond models, we are making the harness behind our own Copilot experience available to customers in Foundry. It gives customers a Microsoft-managed way to build agents that reason, use tools, complete complex, long-running work while preserving their choice of model…

…Rolls-Royce deploys a system to really manage all of their maintenance of whatever their jet engines with 30,000 parts are pretty complex supply chain, both planning, execution, inventory management system, all built on Foundry agentic systems…

…A lot of customers are saying, hey, look, take the Unilever one. It is their system. They will ultimately take even the traces after having done even a bunch with frontier models, and they just want to train their own model. So it can be open weights or, in this case, it was MAI model. So I think all of that needs to be as options that are available. Then depending on pricing, depending on the need of the customer, they will exercise it. There is a temporality to it. It is never like, oh, I start one place and that is where I stop. It is like, I may start at the frontier. I may go to an open weight. I may go to an MAI model. I may go back to the frontier for the next drop for a few months, come out. That is how I think how dynamic these workloads are.

Microsoft’s management thinks the company is making good progress with its 1st party models; management’s goal is to deliver capabilities comparable to frontier models within 3-6 months of their release, at a fraction of the cost; Microsoft’s MAI-Cyber model now powers over 40% of the scans in Microsoft’s MDASH code scanner at 18% lower cost per scan; MAI-Transcribe is the cheapest and most accurate transcription model in the world; MAI-Code-Flash has 3 million users in Github Copilot in September 2026; more than 4,000 customers have used Microsoft-Decision-1, a model for structured decision-making; Microsoft recently completed pre-training of MAI-Base-2, which has 4x-12x better training compute efficiency compared to other open weight models; Unilever’s agentic system was built with MAI models

We also continue to make rapid progress with our first-party models. Our goal is to advance the cost to outcome frontier, which means delivering capabilities comparable to leading models within three to six months of their release at a fraction of the training cost, thereby allowing us to transfer those cost benefits to customers. You see this in our products today. MAI- Cyber now powers over 40% of the scans in our agentic MDASH code scanner at 18% lower cost per scan. MAI- Transcribe is the most accurate and cheapest transcription model in the world, supporting Dragon Copilot’s multilingual workflow. MAI- Code- Flash has three million users in GitHub Copilot last month alone, helping developers get more out of their AI credits. Microsoft-Decision-1 is our new model for structured decision-making, outperforming both LLMs and other decision models in latency and quality. Over 4,000 customers, including Atos Group, Equinor, and RBI, have already used it in the first month since its release. We recently completed pre-training MAI- Base- 2, which outperforms other open weight models we tested with 4x- 12x greater training compute efficiency. This base model will be used in the upcoming MAI models across thinking, code, and cyber…

…Take the Unilever one. It is their system. They will ultimately take even the traces after having done even a bunch with frontier models, and they just want to train their own model. So it can be open weights or, in this case, it was MAI model.

Fabric customers grew 50% year-on-year in 2026 Q3 (FY2027 Q1) to 44,000 (was 40,000 in 2026 Q2); 19,000 customers now use both Fabric and Foundry (was 17,000 in 2026 Q2); management is finding that customers are rapidly deploying Microsoft’s AI-optimised databases; PostgreSQL’s growth accelerated for 4 consecutive quarters in 2026 Q3 (FY2027 Q1); the number of PostgreSQL customers who also use Foundry was up 90% in 2026 Q3 (FY2027 Q1); Microsoft has WorkIQ, which makes an entire organisation’s work knowledge available to an agentic workload; Microsoft IQ combines Work IQ, Fabric IQ, and Foundry IQ, and it has more than 60,000 customers in 2026 Q3 (FY2027 Q1), up 25% sequentially 

We now have 44,000 paid Fabric customers, up more than 50% year-over-year…

…Over 19,000 customers now use Foundry and Fabric as enterprises connect and operate agents over real-time operational, analytical, and unstructured data.

Customers are also rapidly deploying our AI-optimized databases, including PostgreSQL, SQL DB, Hyperscale, Cosmos DB, to give agents fast, secure access to the data and context they need. PostgreSQL growth accelerated for the fourth consecutive quarter, and the number of PostgreSQL customers also using Foundry accelerated to 90% year-over-year.

We also have one of the most important databases for any enterprise, which is the knowledge underneath Microsoft 365, spanning people, their relationships, documents, emails, meetings, chats, projects, and other artifacts. Work IQ makes this context available to every agentic workload. With Microsoft IQ, we are bringing together Work IQ, Fabric IQ, and Foundry IQ as a unified context for agentic systems. Nearly 60,000 customers are accessing context across Microsoft IQ, up 25% quarter-over-quarter.

Microsoft has Agent 365 for AI observability and it has over 50 million agents in the registry in 2026 Q3 (FY2027 Q1), up 30% sequentially; Microsoft’s management recently expanded Perception for customers to secure agentic insiders; management recently introduced guardian agents that monitor agentic activity for unsafe actions; management is seeing encouraging feedback on Perception; customers have used Microsoft Purview to audit more than 7.6 billion agent activities in 2026 Q3 (FY2027 Q1), up 5x sequentially; management recently introduced Microsoft Execution Containers to govern and contain agents; agentic products from OpenAI, Anthropic, and Meta Platforms are starting to adopt Microsoft Execution Containers

 With Agent 365, we provide a platform for AI observability, enabling customers to manage, secure, and optimize the performance and cost of their agents. Over 50 million agents are now in the registry, up over 30% quarter-over-quarter…

…Last week, we expanded Perception with new capabilities to help customers secure and contain agentic insiders, including expanded agent discovery across local and cloud environments, dynamic permissions that enforce least privileged access for each task, and continuous risk assessment. We also introduced teams of guardian agents that continuously monitor agent activity to help prevent unsafe actions before they occur. We are seeing encouraging feedback on Perception overall from preview customers, including KPMG, Nationwide Building Society, Northern Trust, Celentis. More broadly, we are helping customers secure and govern AI use across the enterprise. Customers have used Microsoft Purview to audit over 7.6 billion agent activities to date, up 5x quarter-over-quarter…

…We introduced Microsoft Execution Containers, a policy-driven execution layer that governs and contains agents across Windows and other platforms integrated with Agent 365 for governance and observability. Agents like OpenAI Codex, Anthropic Claude Code, Meta Muse are adopting MXC, and customers like Meta are starting to deploy it across their enterprise environments.

Microsoft’s management is building the company’s compute infrastructure to be fungible, such that it can be dynamically allocated to the highest value opportunities; Microsoft added 24 new data centers across 5 continents in 2026 Q3 (FY2027 Q1); Microsoft has reduced dock-to-live times for new GPUs by 20% sequentially in 2026 Q3 (FY2027 Q1); management has increased the throughput for Copilot workloads by over 50% sequentially in 2026 Q3 (FY2027 Q1); Microsoft’s AI infrastructure continues to utilise chips from NVIDIA, AMD, and itself (Maia); Microsoft has docked the first Vera Rubin GPUs from NVIDIA; Microsoft’s Maia 200 is now delivering 1.7x more tokens per watt across Teams and Copilot, with Maia 300 expected to drive a further 40% more tokens per watt compared to leading GPUs

We are building against a growing book of business with a fungible fleet that allows us to dynamically allocate capacity to the highest value opportunities across compute, data, Foundry, and Copilot. Accelerating demand led to us adding 24 new data centers across five continents this quarter. We are bringing capacity online faster, reducing dock-to-live times by about 20% quarter-over-quarter for new GB300s in our biggest regions, which make up the majority of the GPU deployments. At the same time, we continue to get more from the infrastructure we already have by optimizing across the data center design, silicon systems, software, and models. For example, we increased the throughput for Copilot workloads by over 50% quarter-over-quarter.

We continue to scale our fleet with the latest innovations from NVIDIA, AMD, alongside our own Maia systems. We have docked our first NVIDIA Vera Rubin GPUs and are very excited about it. Maia 200 is already delivering up to 1.7x more tokens per watt across the AI experiences in Teams and Copilot, and Maia 300 will drive further gains, delivering 40% more tokens per watt compared to the leading GPU-based systems.

Microsoft is working to create new PCs that can run powerful AI models locally

As more capable models run locally, we see an opportunity to make it more intelligence effectively unmetered. We are working with the broader hardware ecosystem to bring these capabilities to a new generation of PCs across AMD’s Ryzen AI 400 Series, Intel Core Ultra Series 3, Qualcomm Snapdragon X2 Series. Earlier this month, we also announced new PCs powered by NVIDIA’s RTX Spark and DGX Station for Windows. With these new devices, you can run last year’s Frontier models on your laptop and last quarter’s model at your desk.

1/4 of Microsoft’s cloud and AI-related capex in 2026 Q3 (FY2027 Q1) are for long-lived assets, while the other 3/4 are for CPUs and GPUs; management sees the potential for each dollar of Microsoft’s capital expenditure to support more revenue and higher returns over time

Capital expenditures were $45.2 billion, as some third party capacity deliveries we expected in Q1 were delayed. Roughly three quarters of our CapEx was for short-lived assets, primarily CPUs and GPUs, including higher memory and storage costs, as we continue to invest in the infrastructure required for our customers to build and scale AI applications. The remaining spend was for long-lived assets. This quarter, finance leases were $7.4 billion. As a reminder, we recognize the full principal value of finance leases in CapEx at the time of lease commencement. Cash paid for PP&E was $39.4 billion…

…As we continue to improve the efficiencies of our infrastructure, each dollar of CapEx has the potential to support more revenue and higher returns over time.

Azure’s growth was broad-based across business-size and workloads; Azure grew revenue by 48% in 2026 Q3 (FY2027 Q1) (was 43% in 2026 Q2); Azure’s revenue growth was better than expected because of efficiency gains in the GPU fleet, which drove outperformance in Foundry, and in the CPU fleet; Azure’s growth in 2026 Q3 (FY2027 Q1) benefitted from higher token pricing; management is starting to see the breadth of enterprise AI workloads starting to emerge; Azure’s demand continues to exceed supply

All-up Azure growth was broad-based across both enterprises and small and medium business segments and spanning both core cloud services and AI infrastructure….

…Azure revenue was ahead of expectations and increased 48%, even with the 3P capacity delays noted earlier, driven by Foundry and improved CPU monetization. Foundry revenue more than tripled year-over-year, with upside driven by efficiency gains across our GPU fleet that allowed us to allocate additional capacity to customers, as well as higher revenue per token that we benefited from in the quarter. Across our CPU fleet, engineering work enabled us to identify and monetize underutilized capacity this quarter, helping us serve more of the growing demand with our existing capacity. The timing of engineering efficiencies like these, as well as 3P GPU lease deliveries, our own capacity build-outs, and ARPU changes from model mix can create quarterly variability in the year-over-year growth rate…

…We had higher token pricing than we expected in the quarter. With some high-quality models that we are able to take advantage of and see good margins on do benefit us in the quarter. That did happen in Q1…

…I was even in Europe just last week, and it was great to see the quality of the agentic systems that are just broadly getting deployed. When Unilever deploys what is their automated order taking and trade coupon dispute management stuff, all as a fully agentic touchless system. Rolls-Royce deploys a system to really manage all of their maintenance of whatever their jet engines with 30,000 parts are pretty complex supply chain, both planning, execution, inventory management system, all built on Foundry agentic systems. RBI, which is an Austrian bank, showed me 20+ business processes end to end. They were talking about whether it is in for credit or what have you, how they are able to now do that so much more optimally. So the breadth, the quality, their ability. Oh, by the way, the multi-model nature of it. They start with a frontier model. Many of them are, in fact, the Unilever one, they just dropped an MAI model trained using our traces. So even the degree of sophistication of what I think of as model optimization, workload optimization, token optimization, was just great to see. So the breadth is starting. Finally, it does feel like the enterprise workloads.

They feel like to me, like the database applications enterprises built in the past that I grew up with are happening now when it comes to AI. In fact, if anything, some of the shiny object stuff is waning, which is it is not about the next big thing. They are really more about, wow, can I build a system? Can I own the weights? Can I manage my costs? Can I have reliability? What about security? So in that sense, the enterprise breadth motion is starting…

…In Azure, we expect revenue growth of 49%- 50% in constant currency, with continued efficiency gains across our CPU and GPU fleet, helping to offset some 3P GPU as-a-service contracts that have been delayed to H2. Customer demand continues to exceed available supply. 

OpenAI accounted for 11% of Microsoft’s revenue in 2026 Q3 (FY2027 Q1)

[Question] At the same time, you have OpenAI at 11% of revenue in the quarter.

Visa (NASDAQ: V)

Visa’s management is seeing AI be one of the top 3 focus areas of its clients; empowering agentic commerce is one of management’s top priorities; management thinks agentic commerce requires trust, and that the trust barrier can be solved by using a Visa credential each time an agentic commerce user pays or gets paid; management is currently seeing a limited number of agentic transactions, even with the introduction of Muse and Instinct; Visa has nearly 50 AI partners going live with Visa Credentials and management expects the number to double in the coming months; Visa is collaborating with Anthropic to bring its commerce blueprint to Visa’s merchant network; management thinks B2B is an area for fast adoption of agentic commerce, and Visa has already launched with several partners; Ramp and Airwallex are using Visa Intelligent Commerce to provide B2B agentic commerce solutions; Visa Intelligent Commerce provides trust for agentic commerce users, trust for agents, and trust for the transaction; cross-border e-commerce volume was up 17% in 2026 Q3 (FY2026 Q4), up from 16% in 2026 Q2, driven by retail, including spending on AI platforms; management thinks that agentic commerce will bring incremental opportunities for Visa, but it will take some time to show up; Visa Intelligent Commerce’s role is to ensure that issuers, acquirers, and other ecosystem participants are ready to enable agentic transactions at scale

It continues to be among the top three focus areas for our clients, alongside AI and cybersecurity…

…One of the key areas of innovation for Visa has been empowering agentic commerce. It is clear that in agentic, trust is the biggest barrier to scale. We believe the best way to break through that barrier is with one simple but powerful idea. Every time someone pays or gets paid, it is with a Visa credential. As we look at the albeit limited number of agentic transactions today, Visa credentials are being used by consumers, businesses, and agents. We have nearly 50 AI partners live and in production, including more recently, Shopify’s Shop Pay, with that number expected to double in the coming months…

…I think both Muse and Instinct experiences are good examples of that. I think it is still very early. If you just look at just even the user numbers that are public for Muse, it is still very early in terms of the number of consumers that are doing this, and certainly in the big scheme of things in our overall business, it is a very small amount…

…We are also collaborating with Anthropic to bring their commerce blueprints to our merchant network, combining the intelligence of Claude with the trust, security, and global reach of the Visa network. One of the areas we have identified for fast adoption of agentic has been B2B, and we have had several partners launch and expand offerings. Ramp’s agentic bill pay use case is now generally available, and as Ramp businesses use Visa Intelligent Commerce to automate bill payments volume is growing. Airwallex is testing Visa Intelligent Commerce-enabled credentials and integrating them into their native wallet, Airwallex. BILL’s payment agent is piloting Visa Intelligent Commerce to pay invoices, and Billtrust’s supplier portal will identify the agent and accept payment. Our Visa Intelligent Commerce solution is enabling all of this. It provides trust for the user with credentials, agentic tokens, and transaction identifiers, trust for the agent with the Trusted Agent Protocol and Agent Directory, and trust for the transaction. Once we have verified the user and agent identity, we need to make sure that the purchase instructions are clear by capturing consumer-authenticated intent. Of course, like on any Visa transaction, our rules clarify the roles and responsibilities and processes should something go wrong. It is clear that agentic commerce is here, and Visa’s role is clear: delivering trust…

…Cross-border e-commerce volume was up 17%, more than one point above Q3, primarily driven by retail, including spending on AI platforms…

…[Question] I wanted to ask kind of a follow-on question related to how you are investing in agentic commerce and all the new money movement capabilities. As these flows continue to develop, how much of an opportunity do you expect it genuinely be incremental to Visa versus potentially a migration of existing card volumes?

[Answer] We do have conviction that there is real incremental opportunities. But as we think about the next four quarters in 2027, it all remains very small in the big scheme of the enormous size of our business. A couple other notes in terms of where we see incremental opportunities. One, you mentioned some of the work we have been doing on Visa Direct and things like that. Those are big incremental opportunities. Two, when you think about machine-to-machine payments and the CLI as a commerce interface, there is a lot of incremental opportunities there. Then just more broadly, in the big scheme of things with agentic commerce, are agents finding more buying opportunities for us? Are agents making more purchases at more seller locations because they are optimizing based on value, speed, delivery, whatever it is important to the user? Those are opportunities. I think we remain very excited and have conviction about the opportunities, but I think we all just need to realize that the development of all of this is still quite early…

…[Question] I want to hit on this a little bit more on the agentic network token. Our understanding is that this token is sort of a bolstered version of the network token. It’s more of a scoped network token, if you will. It will have all the programmatic aspects, meaning a time window, merchant category code, dollar amount. It will have the upfront intent, as you mentioned in your prepared remarks. All of that gets kind of bundled up into the agentic network token. The question is, how ready are the issuing banks, and what do they need to do in order to say, “Yes, I’m willing to have an agent issue an agentic network token for one of my customers.” Related, what about the merchant acquirers that are out there, the PSPs? Are they ready to accept the tokenization handshake, if you will?

[Answer] That was a great summary of a lot of the objectives of Visa Intelligent Commerce and our agentic tokens. In short, we are doing the work across the ecosystem to ensure that issuers, acquirers, and other ecosystem participants are not only ready but are ready to enable these transactions at scale. We know how to do this work. If you look at our history, even just with tokenization pre-agentic commerce, this is the work that we did broadly across the world. I think I mentioned in my prepared remarks now that we have tokenized e-commerce transactions in 200 countries and territories around the world. We know how to do this work to enable the ecosystem. I also just want to come back and build on your points around these agentic tokens and just put it in context of what we are delivering through Visa Intelligent Comme

Visa’s new agentic-AI-driven way of working has made big impacts on its initiatives for stablecoins; the Visa Stablecoin Platform enablement for OUSD was designed, built, and shipped in seven weeks with 12 people with the new agentic-AI toolchain and would have taken much longer otherwise; more than 60% of Visa’s global product development and engineering teams have adopted the new agentic-AI-driven way of working; management intends for 100% of the teams to adopt this new way of working; Visa strengthened its defenses in FY2026 with the Visa Vulnerability Agentic Harness, which has been open sourced; Visa has used AI to deliver over 350 product releases in FY2026, up 25% year-on-year

I spoke last quarter about the fundamental changes to the way that we build and ship products enabled by utilizing our agentic toolchain. For example, the Visa Stablecoin Platform enablement for OUSD was designed, built, and shipped in seven weeks with 12 people. This would have taken much, much longer and required many more people without utilizing our agentic toolchain. Today, more than 60% of our global product development and engineering teams have adopted this new way of working with the intent of reaching 100% this year…

…This year, we tested and strengthened our defenses through our Visa Vulnerability Agentic Harness, which we then open-sourced to the world…

…We demonstrated our ability to significantly accelerate product development with the true birth of tokenized agentic commerce transactions powered by Visa, with blockchain infrastructure and platforms designed for payments, with the increased utilization of AI across our business to deliver over 350 product releases, up 25% year-over-year.

---

Disclaimer: The Good Investors is the personal investing blog of two simple guys who are passionate about educating Singaporeans about stock market investing. By using this Site, you specifically agree that none of the information provided constitutes financial, investment, or other professional advice. It is only intended to provide education. Speak with a professional before making important decisions about your money, your professional life, or even your personal life. I have a vested interest in Alphabet, Amazon, Mastercard, Microsoft, and Visa. Holdings are subject to change at any time.

Mastercard (NYSE: MA) and Visa (NYSE: V) are two of the largest payments companies in the world. As a result, they have a great view on consumer spending that’s taking place. With both companies reporting their earnings results for the third quarter of 2026 earlier this week, the bottom line is that consumer spending remains strong in the USA and other parts of the world. Here’s what they are seeing.

*What’s shown in italics between the two horizontal lines below are quotes from Mastercard and Visa’s management teams that I picked up from their earnings conference calls.

---

From Mastercard

1. Mastercard’s management sees consumers and businesses continuing to spend despite economic and geopolitical uncertainty; management sees the economy as having a resilient backdrop; Mastercard’s business fundamentals remain strong; management is seeing a supportive macroeconomic environment with balanced labor markets, strong household balance sheets, and continued wage growth; management sees European consumers as being very resilient  

While there continues to be economic and geopolitical uncertainty, consumers continue to spend, businesses continue to invest, and commerce continues to expand. The economy, our overall business, and our network have a resilient backdrop…

…As mentioned, consumer and commercial spend was healthy in Q3 and so far into Q4. Our business fundamentals are strong, and we continue to deliver solid growth across both our payment network and value-added services and solutions…

…This is all underpinned by a supportive macroeconomic environment with balanced labor markets, strong household balance sheets, and continued wage growth. To date, consumers and businesses have demonstrated resilience. There is heightened geopolitical uncertainty, which has persisted for some time, driving energy prices higher and putting more focus on monetary policy around the world…

…If I take a look at the macro environment right now, the European consumers are very resilient.

2. Worldwide GDV (gross dollar volume) was up 7% year-on-year in 2026 Q3 in constant-currency basis; cross border volume was up 11% globally in constant currency, driven by both travel and non-travel cross-border spending (cross border volume growth was 12% in 2026 Q2); switched transactions was up 9% year-on-year in 2026 Q3; card growth was 5% in 2026 Q3, with Mastercard ending the quarter with 3.8 billion cards in circulation (there were 3.7 billion cards in 2026 Q2, and year-on-year growth was 5% then); domestic assessments were up 13%, cross-border assessments were up 16% and transaction processing assessments were up 14%

Worldwide gross dollar volume, or GDV, increased by 7% year-over-year. In the U.S., GDV increased 6%, with credit growth of 10% and debit growth of 1%, which includes the impacts of the Capital One debit migration. Outside of the U.S., GDV increased 8%, with credit growth of 6% and debit growth of 10%. Cross-border volume increased 11% globally for the quarter, reflecting continued growth in both travel and non-travel related cross-border spending…

…Switched transactions grew 9% year-over-year in Q3…

…Card growth was 5%. Globally, there are now 3.8 billion Mastercard and Maestro-branded cards issued…

…All growth rates are described on a currency-neutral basis unless otherwise noted. Looking quickly at each key metric. Domestic assessments were up 13%, while worldwide GDV grew 7%. The six PPT difference is primarily driven by pricing and mix. Cross-border assessments increased 16%, while cross-border volumes increased 11%. The five PPT difference is driven primarily by pricing in international markets and mix. Transaction processing assessments were up 14%, while switched transactions grew 9%. The five PPT difference is primarily related to pricing and a favorable mix, partially offset by lower revenue from FX volatility.

3. In 2026 Q3, Mastercard’s operating metrics were in line with Q2 and remained strong (switched volume was up 9%, switched transactions was up 9%, and cross-border volume was up 11%); in October so far, Mastercard’s operating metrics continue to be strong with worldwide switched volume growth of 10% (9% in the USA, and 11% outside of the USA), switched transactions growth of 9%, and cross-border volume growth of 10%

Starting with Q3, switched volume and transaction metrics are generally in line with Q2 and remained strong. Of note, the sequential increase in the U.S. switched volume was primarily due to recent commercial wins. Looking at cross-border. First, lower yielding intra-Europe volumes were down sequentially from Q2 to Q3, driven by the impacts related to a new customs fee in the E.U., timing of large retail promotional events, and lower crypto purchases in the quarter compared to last year. These impacts on intra-Europe cross-border were the primary reasons for the decline in card not present ex-travel sequentially. Growth in higher yielding other cross-border remained stable and slightly higher quarter-over-quarter.

As we look to the first three weeks of October, spend remained strong with metrics generally in line. A few points to note. In the U.S., switched volume improved sequentially as we begin to see the lapping of the Capital One debit migration, the lift from recent commercial wins, and higher spend on gas. As a reminder, the Capital One debit migration began towards the end of Q3 last year and continued through Q1 2026. Looking at cross-border, the sequential impacts remain predominantly intra-Europe. Growth in cross-border card not present ex-travel continued to see the impacts from two of the areas I noted earlier, notably the customs fee and a tougher comp due to higher spend on crypto purchases last year. The slight deceleration in cross-border travel was due to timing of holidays. Overall, underlying spend remains strong and consumers and businesses are spending.

From Visa

1. US payments volume growth was good at 10% in 2026 Q3 (FY2026 Q4); both discretionary and non-discretionary spend growth remained strong; there was good growth in both US credit and debit volumes; growth across consumer spend bands was consistent with prior quarters, with the highest spend band continuing to grow the fastest; there was no sign of weakening in spend among the lower-spend consumer band

U.S. payments volume grew 10% year-over-year, generally consistent with Q3, as we saw several factors from Q3 continuing to play a role, including the cost of fuel, retail, strong Visa Direct growth, and FIFA-related spend. From the FIFA World Cup round of 32 matches on July 1st through the final on July 19th, we continued to see both U.S. host cities as well as destination cities benefit from the influx of FIFA fans, with total U.S. card-present spend payments volume growth elevated above pre- and post-tournament levels. For the quarter, both U.S. discretionary and non-discretionary spend growth remained strong. Fourth quarter U.S. credit grew 11% year-over-year, and debit grew 9%. Growth across consumer spend bands remained relatively consistent with prior quarters, with the highest spend band continuing to grow the fastest. We did not see signs of the lower-spend consumer weakening in our volumes.

2. Visa’s cross-border volume growth remained strong in 2026 Q3 (FY2026 Q4) at 14%, up from 12% in 2026 Q2; cross-border e-commerce volume was up 17% in 2026 Q3 (FY2026 Q4), up from 16% in 2026 Q2, driven by retail, including spending on AI platforms

Q4 total cross-border volume was up 14% year-over-year, more than two points above Q3, primarily due to travel. Travel-related cross-border volume was up 13%, three points of acceleration from Q3, primarily driven by improvements in both CEMEA inbound and outbound. Overall, U.S. inbound cross-border growth for July remained elevated following strong growth in June, reflecting in part the demand around the FIFA World Cup. Focusing on inbound cross-border card-present spend, U.S. host cities increased by over 20% year-over-year from July 1 to July 19, with the most significant increases driven by fans from Argentina and Colombia. Cross-border e-commerce volume was up 17%, more than one point above Q3, primarily driven by retail, including spending on AI platforms.

3. Payments volume on Visa’s network continues to grow in October 2026, with US payments volume up 10%, cross-border volume up 15%, e-commerce volume up 17%, and processed transactions up 10%; management will soon stop discussing intra-quarter payment volume trends

Let’s look at the drivers through October 21st with volume growth in constant dollars. U.S. payments volume was up 10%, with credit up 11% and debit up 9% year-over-year. For constant dollar cross-border volume, excluding transactions within Europe, total volume grew 15% year-over-year, with e-commerce up 17% and travel up 13%. Processed transactions grew 10% year-over-year…

---

Disclaimer: The Good Investors is the personal investing blog of two simple guys who are passionate about educating Singaporeans about stock market investing. By using this Site, you specifically agree that none of the information provided constitutes financial, investment, or other professional advice. It is only intended to provide education. Speak with a professional before making important decisions about your money, your professional life, or even your personal life. I currently have a vested interest in Mastercard and Visa. Holdings are subject to change at any time.

The way I see it, artificial intelligence (or AI), really leapt into the zeitgeist in late-2022 or early-2023 with the public introduction of DALL-E2 and ChatGPT. Since then, developments in AI have progressed at a breathtaking pace.

We’re thick in the action of the latest earnings season for the US stock market – for the third quarter of 2026 – and I thought it would be useful to collate some of the interesting commentary I’ve come across in earnings conference calls, from the leaders of technology companies that I follow or have a vested interest in, on the topic of AI and how the technology could impact their industry and the business world writ large. This is an ongoing series. For the older commentary:

• 2023 Q1 – here and here
• 2023 Q2 – here and here
• 2023 Q3 – here and here
• 2023 Q4 – here and here
• 2024 Q1 – here and here
• 2024 Q2 – here and here
• 2024 Q3 – here and here
• 2024 Q4 – here, here, and here
• 2025 Q1 – here and here
• 2025 Q2 – here and here
• 2025 Q3 – here, here, and here
• 2025 Q4 – here, here, here, and here
• 2026 Q1 – here, here, here, and here
• 2026 Q2 – here, here, here, and here

With that, here are the latest commentary, in no particular order:

ASML (NASDAQ: ASML)

ASML’s management has completed value-based pricing conversations with key customers but the outcomes will not be visible in 2026; most of ASML’s expected growth in 2026 is coming from memory net system sales with, specifically DRAM, with 75% growth; management expects the ratio between logic and memory to shift towards memory, but logic demand remains strong; management expects ASML’s advanced logic business to increase by over 25% in 2026; management expects the logic business to be larger than the memory business in 2027

[Question] We also talked last quarter about higher value-based pricing and the contribution that might play on the gross margin. How is that showing up as part of the outlook? 

[Answer] We completed those conversations. So that’s been done with the key customers. The effect of that is not really visible in 2026. It will start to become visible partly in the course of 2027. The full impact of that you will start to see in 2028 and beyond…

…Turning to end use, this year, we expect logic net system sales to grow around 10% and memory net system sales to grow around 75%. The growth in memory is driven by the combination of capacity expansion of our customers and the increasing adoption of EUV on advanced DRAM nodes…

…We are still on trend where memory become more and more important, 75% increase, which is of course significant. You start to see the results of this huge memory investment. That, by the way, is mostly DRAM, to be clear. We continue to see that moving forward. I think we have said in the past that the ratio between logic and memory will most probably increase towards memory. Still, I would say that logic is also very strong. To be able to predict exactly where this is going to land is difficult because I like to stress that we see also a very strong demand on the logic side, because of course, as you know, the two are very much coupled when it comes to delivering AI capacity to the market…

…Logic went up around 10%, but the advanced logic piece, included in there is far better. So that grew over 25%, is expected to grow over 25% in this year. You are right, memory is getting closer to logic, but logic remains very important. When we look at 2027, memory comes close. But I still believe that logic will outweigh memory in the total book of business for next year.

ASML’s customers are seeing their demand rising both in the short and long-term; management thinks ASML’s customers have very good visibility into their demand; ASML’s customers think they will be in a supply-shortage until at least 2028, and so they are building up capacity faster; the faster capacity build up by ASML’s customers are leading to more demand for ASML in the short-term across its product portfolio, especially for EUV machines; ASML has already taken in a lot of orders for 2028 and there are even orders for 2029 and 2030; the higher visibility that ASML’s customers have on their demand has given ASML better visibility too; management thinks the demand they are seeing is real and not inflated by buying-behaviour from customers; customers that put an order with ASML have to make 2 serious commitments; management has not accepted the 2029 and 2030 orders because they are still understanding the overall demand picture for that time frame; management does not see ASML’s customers’ clean room constraints being a major factor on the company’s demand; the interest management is seeing from customers for 2029 and 2030 comes equally from DRAM and logic; the DRAM customers talking to ASML for 2029 and 2030 are from the US and Korea (so Micron, SK Hynix, Samsung)

We did indeed spend a lot of time with our customers and all of them are commenting that they see their demand increasing on the short term but also extending in time. I think that their visibility is very good at this point of time. Today they expect that their supply will not meet the demand at least until 2028. So this is a very, very strong demand on their side. And of course this means that they want to build up more capacity faster. So they are accelerating their plans. They are announcing new fabs. For us this translates on the short term into more demand, as Roger mentioned. Overall, basically for our holistic product portfolio and on the long-term increased visibility for our demand, especially with EUV. So a lot of our discussions with our customers now are shifting in fact beyond 2027. A lot of order intake for 2028. But also some strong interest to already place orders for 2029 and sometimes even 2030. ..

…Because they have very high visibility on what is needed in the future. And as a result we also have this very high visibility…

…[Question] Do you worry that customers are pulling in orders or giving you bigger forecasts just to secure a slot?

[Answer] When it comes to the demand we are referring to, the increased demand we are referring to, I think we are very much capable to validate that demand with some more detailed plan our customers are sharing with us in term of fab construction, release, et cetera. I think that when we talk about demand, when we refer to demand in those discussion, we really refer to demand we believe is real and not inflated by any potential buying behavior from our customer…

…If you put an order with ASML, that’s not a frivolous piece of paper that you put your name under. It comes with two serious commitments. One serious commitment is a commitment to make a down payment. As I qualified it before, it’s a down payment that is enough to hurt, right? So that’s one element. Secondly, a PO is a real commitment on both sides of the equation, and it also means that if you want to get rid of that commitment, that’s not going to be easy. That comes at a price, not just the down payment, but more than that. Therefore, before customers sign a PO with ASML, they’ve made up their mind, and they’re not doing this frivolously…

…Customers are requesting us, and therefore we have not yet accepted those, is that we want to make sure that we have a good picture of what the overall demand in 2029 and 2030 looks like that customers are currently signaling, and mirror that against the capacity expansion as we see it at that point in time. We want to do that because we don’t want now to accept orders from customer A, B, and C, and then to tell customer D, who hasn’t given orders yet for 2029 and 2030, “Sorry, we’re sold out.” That would be totally inappropriate for customers that have been business partners for us for a very long period of time. We’re now really using this period of time to mirror, to get a good understanding, what is the total demand that our customers are signaling for the 2029, 2030 timeframe? What does it look like?…

…If we look at the longer term, as I said before, beyond those little move or those tactical move, we have confirmed that the number our customer are giving us are really lined up with their plan.  In fact, in many cases, when the supply is tight, we make sure before we fully allocate a tool to a customer that the space will be there, the facility also will be there, so that we don’t ship a tool that will be waiting for weeks for a customer to use it, because, of course, other customer will be very disappointed with that. We do a lot of work to make sure that in all the plans we have at ASML, we don’t have delay of more than a few days or weeks, I would say, in shipping our product. Short answer is, I think, no. We don’t see our customer clean room constraint being a major factor on our demand…

…[Question] Just to understand the discussions that you have on 2029 and 2030, if it is more because you seem to be very positive at the moment on memory. If it is very balanced between memory and logic, or if it is still memory manufacturer who are giving you some visibility on 2028– 2029, sorry, and 2030.

[Answer] You need both memory and logic to get some AI chips. I think we have discussion with both customers because they get the same message from their own customer. I think today, DRAM is a bit more aggressive on the buildup of those capacity. But yeah, we have discussion with both. I think we see equal interest, equal, I would say, demand trajectory for both DRAM and logic, because we need both basically to get the product out…

…[Question] DRAM, is it at the same time the U.S., Korea, and even China, or is it more concentrated from a geographic standpoint?

[Answer] We are mostly talking about EUV in this case. So this excludes China.

ASML’s management has been planning for a 30% increase in capacity in each of 2027 and 2028, but management is also clear that the planned capacity increase is not enough to meet demand; ASML’s management is looking to increase capacity in 2028 beyond the 30% that had been planned; ASML’s Installed Base business is able to help its customers increase manufacturing capacity; management will be introducing the NXE:3800F EUV machine in 2027 that will also help customers increase manufacturing capacity; ASML’s customers are asking the company to increase capacity; management will soon be introducing the next generation DUV system that will also help customers increase manufacturing capacity; management does not see Carl Zeiss as a bottleneck for ASML’s capacity expansion; the capacity-increases that management has in mind are already validated with ASML’s supply chain

Last quarter, we talked about 30% more capacity next year, and also the aspiration to have another 30% increase in 2028. A lot of focus there. But to be honest, it’s becoming very clear that 30% in 2027, 30% in 2028 will not allow us to meet our customer demand. So we have to do more and we are looking basically at ways, especially in 2028 and beyond, to further extend our capacity…

…In addition to increasing unit shipments that Christophe talked about, one important driver obviously is the Installed Base business that we’ve talked about quite a bit. So of course that is a major driver. It’s something that customers really like. Because what it does, it gives you within your existing footprint, you get pretty fast incremental capacity there. So that’s been very successful so far and we expect that to continue to be quite successful on a go forward basis. We’re also very aggressively developing upgrades. So we will get more and more upgrades on our installed base available to our customers. So in that way, we try to help customers in getting instantaneous access to more capacity…

….We talked in the past about the NXE:3800F, which gets you from the 230 wafers per hour that we have today to 260 wafers per hour. That will be introduced in 2027. That will be in small quantities. But obviously for 2028 and beyond you will see that really helping to increase the capacity of our customers…

… [Question] What then would you say are the key things that you want to focus on for the rest of the year but also in preparation for next year?

[Answer] I think it’s very clear, it’s more capacity. As we discussed today, our customers are asking us every single day to get more tools…

…We also plan to introduce the next generation critical Deep UV immersion, running at 330 wafers per hour. These new products will provide incrementally more wafer capacity to the industry as they ramp in the coming years…

…We do not see Zeiss today as a bottleneck for the capacity expansion we are discussing, even beyond the 30% we mentioned for 2028. We have done, in the recent years, many work with all the supply chain to make sure that the supply chain will not be an issue for us…

…Every number we share with you is the result of us having validated that the supply chain can, of course, follow those numbers. 

ASML’s customers are starting to use the company’s high-NA EUV machine for high-volume manufacturing; TSMC has asked ASML to co-lead a consortium to introduce a 12-inch reticle for high-NA EUV; the 12-inch reticle is slated for introduction in 2031 and to enter production in 2033; management thinks the existence of the 12-inch reticle project is a sign that the entire ecosystem wants to more and more adoption of high-NA EUV; TSMC will introduce high-NA EUV in 2030 for high-volume manufacturing; all of ASML’s other major customers are also in the consortium; Samsung will be using high-NA EUV in high-volume manufacturing in 2028; Intel continues to use high-NA EUV in high-volume manufacturing, and the performance of the high-NA EUV system meets or exceeds comparable layers pattern using low-NA EUV; management now expects to see adoption of high-NA EUV increase node-on-node moving forward

The High NA program started more than ten years ago. We are all very happy to see that our customers are now about to start using the tool in high-volume manufacturing…

…So TSMC has asked us to lead with them a consortium basically to bring a 12-inch reticle to the industry. This is what is happening sometime in the first half of the next decade. But it’s also the sign of how important High NA has become in their plan. That’s also why they have basically confirmed that they will also introduce High NA around 2030 in high-volume manufacturing. So this is a very important announcement because it shows basically that over time, customers are looking at using High NA more. This is why they want to use the 12-inch reticle. Now you have noticed that all other major customers also joined that consortium…

…To further ensure chip makers can fully leverage the advantage of High-NA EUV, we also, in collaboration with TSMC, began an industry-wide initiative to develop larger formats EUV photomask. This collaborative effort aims to establish a 12-inch mask pilot line by 2031 and enter production in 2033… 

…I think the announcement, and the drive of TSMC first on the 12-inch critical mask, is very interesting there. Because the first signal it gives is that, in the mind of the customer, there will be more and more use of High-NA. Because if it was not the case, no one will really bother to go to 12-inch mask on High-NA. The 12-inch mask drive and very wide adoption by customer in the consortium, is just a sign that everyone has in mind over time to develop basically more and more adoption of High-NA…

…Samsung has also announced that they will be using High NA in high-volume manufacturing starting in 2028. And Intel, I think we talked about that before, they continue to run the system in high-volume manufacturing, increasing the number of layers. Of course, we look now at a significant amount of wafers being run there…

…Intel announced that it has exposed more than 1 million wafers using its High-NA EUV systems, with performance that meets or exceeds comparable layers pattern using low-NA EUV…

…With availability across the existing High-NA EUV fleet approaching 90%, we now expect to see adoption increase node and node moving forward. 

ASML’s management sees very strong demand in China for domestic consumption for 2026 and beyond; management sees ASML’s Chinese customers all wanting to work with the company; management thinks the likelihood of strong competition for ASML in China is still low; management thinks China is still a fair distance away from being able to compete with ASML, even though there have been rumours for 25 years that China is on the cusp of being able to compete with ASML

If you look at what China needs in terms of mainstream logic, in terms of memory, there is very strong demand in China for their domestic consumption. If you look at the fabs that are being built in China, there is a lot coming online there. Therefore, we would expect the demand in China to be strong. This year, 20%. I would argue that also for 2027 and beyond, based on the intrinsic drivers, it seems that the demand situation in China remains very strong…

…Where we are today, I think, with our Chinese customer, is that they all want to work with us. I think this is very clear. I won’t give any market share number, but you can imagine it’s extremely high, especially on immersion. Therefore, as long as we have a good market access, as long as our customer can trust they can work with us, I think that the likelihood of a major competition in China is still pretty low…

…If they were to not have access to some of our tools, then you get a lot more motivation. But it’s still a very, very long way to go. I think that we have been having this discussion about China having a tool to compete with us for the last 25 years, and every other years, we get the news that this is coming. If you stick to fact, I think the fact this year are not different from the fact a year ago or even five years ago. So I think that the progress is not very strong.

Intuitive Surgical (NASDAQ: ISRG)

Intuitive Surgical’s management is developing AI capabilities on da Vinci 5, including a voice user interface, and an automated fourth-arm control; management thinks the AI capabilities will improve efficiency and outcomes, but the technologies must earn the trust of the ecosystem before they can be scaled; the da Vinci 5 has multiple programs with advanced visualisation amd AI features that can add revenue per procedure

Our teams continue to develop additional digital and AI capabilities on da Vinci 5, including a voice-activated user interface and automated fourth-arm control to further enhance efficiency and the user experience. AI-controlled motion in surgery has the potential to improve efficiency and outcomes, but it is early and must earn the confidence of regulators, surgeons, care teams, and patients before we deploy it at scale…

…Another area that we are very excited about when it comes to da Vinci. We have multiple programs going there with advanced visualization technology around hyperspectral imaging, fluorescing molecules, and some AI features that we do think will bring that incremental value, and we are going to be able to add some revenue per procedure as a result.

Meta Platforms (NASDAQ: META)

The vision of Meta’s management for personal superintelligence is for every person to have an extremely capable personal agent that helps them achieve their unique goals in life; Meta launched its personal agent product, Muse, recently and in 1.5 months, more than 10 million people have set up Muse personal agents; Muse has been #1 in the App Store nearly every day; management thinks Muse’s engagement and retention are off to a strong start, with 25% of people who have set up Muse using it every day; the longer people use Muse, the better their retention; Muse’s retention appears to be at least 2x better than where other AI apps started at; the most common use case for Muse so far is shopping, with an average purchase size of more than $200; payment technology company Plaid said Muse has shopping activity never seen before on other services; Muse’s shopping volumes in the first week were much higher than management expected; other common Muse use cases are personal productivity and connecting to other personal services; Muse has state-of-the-art privacy and security; management intends to implement Muse Confidential VM (virtual machine) soon, so that even Meta cannot see users’ content; the introduction of Muse Wonder, Meta’s latest frontier model, has improved Muse’s quality, engagement, and retention; management believes Muse will perform better over time as Meta releases better models; management is seeing that businesses are very excited about enabling agentic commerce; management has been positively surprised by how many businesses want to connect with Muse

We have been discussing our vision for personal superintelligence for more than a year now. We want to help every person have an extremely capable personal agent that understands your goals and can help you achieve them and improve your life in whatever ways you want — whether that’s improving your relationships, helping you save time, save money, build a business, create new things, or anything else that you can dream of. Muse is the product form of this vision…

… In just a month and a half, more than 10 million people in the US and Canada have set up Muse personal agents. Growth continues to be strong and we have been #1 in the App Store almost every day. Engagement and retention are also off to a strong start. About 25% of people who have set up Muse use it on any given day. For our earliest adopters, the longer that they use Muse, the better that retention seems to get. To put this in perspective, this is more than twice the level that Threads started at or where we believe that other AI apps started at as well. 

The most common use case for Muse is shopping, with more than a quarter of people who are using Muse engaging in this with an average purchase size over $200. Plaid told us they hadn’t seen activity like they’re seeing with their connector on any other service before. And in the first week we had to work with Stripe to raise credit limits since the volumes were above what any of us expected. Beyond shopping, the most common connectors are for personal productivity and connecting Muse to your other services like email and calendar. This shows the importance of the Muse Secure VM that we’ve built to provide state-of-the-art privacy and security, and we’re looking forward to following up soon with the Muse Confidential VM, which will make it so that even Meta cannot see your content.

We have also seen that Muse Wonder improves quality, engagement, and retention further, which suggests that as we continue to improve intelligence, Muse will perform better and better over time…

…[Question] When you think about the friction points around agentic commerce, how do you guys as a company think about solving for friction that’s rising where there are aggregators or suppliers that don’t want agents to interact with their platforms or their services and how you think about lowering that friction or creating the right balance between the existing services that consumers access today relative to how this might all be changed or altered by agentic behavior over the medium to long term? 

[Answer] Right now what we’re seeing is that, overwhelmingly, businesses are very excited about enabling this because they recognize that it’s going to be a new opportunity for them not only to increase the sales and the quality of the experience for existing customers but also potentially to bring in new customers… We’ve been very positively surprised by how quickly, I think a huge amount of businesses have wanted to not only figure out how to make it so that Muse can use their services more effectively but have reached out to us to integrate around it.   

Meta’s management sees Muse entering an operational phase that involves scaling the product, growing revenue, and optimising compute costs; management expects Muse to have billions of users; management wants to create a large Muse network as quickly as possible because they think a scaled network will get more done for users; management expects the average revenue per user for Muse to eventually exceed the few hundred dollars per year seen for American users of Meta’s family of apps; management expects Muse to be free for most users and monetisation will come from (1) transaction-based fees from free users, (2) subscriptions from users who want better models and/or more tokens, (3) subscriptions from businesses who want more features, and (4) connecting Muse to improve personalisation and quality of content and advertising across Meta; Muse is very compute-intensive currently but management expects to improve its cost efficiency by 10x in the next year; management thinks Muse will become highly profitable if compute efficiency improves over the next few years; there are at least 500 million small businesses using Meta’s family of apps, and management thinks Muse can help these businesses; Muse is currently powered by Muse Spark, and Muse Wonder is available only on the paid tier; management wants to introduce a more efficient version of Muse Wonder for Muse; management recently previewed the Personal Agent Protocol, an open standard for Muse and Meta business agents; management sees a few ways to improve Muse’s compute efficiency, including (1) leverage off-peak and spare compute capacity, and (2) using smaller modles; management expects to introduce monetisation to Muse faster than Meta’s normal playbook

Now that Muse has clear product-market fit, we’re entering a new phase and a more operational phase that is focused on scaling it. The goals of this next phase play to many of Meta’s strengths. We need to do three things: scale the product, grow revenue, and optimize compute costs. We expect to grow Muse to serve billions of people…

…We think that billions of Muses in a network at scale will be able to get more done for people and businesses by working together. So we’d like to create a large scale Muse network as quickly as possible…

…The average revenue per American user for our existing Family of Apps users is several hundreds of dollars annually, and we expect that Muse will reach an even higher level over time due to the amount of commerce that we already see taking place. This will take some time, but we’re starting to work on this now…

…For most people, the product should be free and we will build commerce tools so we get paid a fee based on the transactions that we help facilitate. We expect this will be paid for by businesses. For people who want the most advanced models or to use extra tokens, they subscribe directly and pay. For business owners who want additional features to use Muse to grow their businesses or for companies that want people to use Muse for work, they will be able to subscribe directly and pay as well. And for those who want Muse to help improve the personalization and quality of their content and ads across Meta, they will be able to connect to those experiences as well…

…On the compute side, we are building out the capacity to serve more than a billion people over the next few years. In the near term, Muse is very compute-intensive to run. But we expect to improve its cost efficiency by more than 10x in the next year, which will enable it to be profitable once the revenue lines are at scale. If compute efficiency continues improving quickly for another few years, which we think is highly likely, then Muse will be highly profitable over multiple years…

…We also see a significant opportunity to extend Muse to small businesses. Based on a recent user survey, we estimate that at least 500 million self-identified small businesses use Meta’s family of apps to reach customers and generate revenue. Muse could help these businesses acquire customers and operate more efficiently, while helping people complete transactions. Over time, consumer and business agents could connect and transact directly, creating opportunities around paid usage, commerce, and business services…

…We launched Muse first on Muse Spark, and that’s what is currently the default across the free tier. And we rolled out Muse Wonder, and people can access that in the paid tier today. And we know that it performs a lot better, so we’re working on getting a more efficient version of Muse Wonder that we’ll be able to roll out more broadly…

…Earlier this month, we previewed the Personal Agent Protocol, an open standard across Muse and Meta business agents that we are developing with industry partners like Sierra, Walmart, Stripe, Shopify, et cetera. The protocol is intended to define how a consumer’s personal agent can interact securely and reliably with the business’ agent and preserve appropriate controls for both the business and customers…

…I think there are a couple of dimensions of achieving that efficiency. I think some of the more nearer-term pieces include things like leveraging offpeak or spare computing capacity to run some sets of jobs, using smaller models. There’s a lot of work to be done across different inference efficiency initiatives, like all in service effectively of higher GPU utilization and other product efficiency work we can do also. And I think over the longer term, I think this is a place where we expect that we’re going to be able to, over time as LLMs scale, get more tasks done with fewer tokens and also rely on AI models themselves to help us improve the efficiency of the software that we use to run the model…

…We are going to work on monetization somewhat sooner on this than our normal playbook for a couple of reasons. One is just the size of the numbers that we’re talking about in terms of both the costs and the opportunity, I think make it kind of important to make progress on that in order to be able to scale it. The other is just that such a large part of what people are doing is commerce. And to your question, that does include travel, too, I think we might consider travel a different case, but travel is actually one of the other top use cases for Muse, is basically people planning that out and planning different kinds of things that they would get than they might get through our Family of Apps. 

Meta’s management is building the Meta Enterprise Platform to serve as much high-margin inference demand as possible to leverage on the large fixed costs that come with training frontier models; management thinks that traditional enterprise companies do not have AI usage locked down yet

Since training is a large fixed cost, now that we have leading models, we want to serve as much high-margin inference demand as possible. That is why we are building the Meta Enterprise Platform, led by CJ Desai. Muse Wonder and even more advanced models that we are training can serve additional high-margin services beyond the Muse network, including our API, Muse Code, and other enterprise services…

… AI is creating huge amounts of disruption in enterprise today. And we don’t — from what we can see, we really don’t believe that traditional enterprise companies have this locked down. 

Meta’s management sees the business having many good uses for its compute capacity; management believes that building multiple ways to deploy advanced models will increase the chances that one or some of them will significantly outperform management’s expectations; Meta has yet to reach the limit on effectively applying compute to accelerate its core business

A fundamental feature of our business is that we have many good uses for this compute. It is difficult to forecast the exact growth rates for the Muse network, for our enterprise platform, and for how Muse Wonder and beyond will accelerate our core business. But through building out multiple ways to deploy these advanced models, we increase the chances that one or more of them significantly outperforms our expectations. Beyond serving other businesses, we have not yet found the limit of how effectively we can apply compute to accelerating our own core business. 

Meta’s management is using Muse Wonder, the company’s latest frontier model, to improve its ability to autonomously improve the company’s recommendations and advertising systems; management expects an increasing rate of improvement in Meta’s systems as the company builds better models; Muse Wonder is state-of-the-art across many dimensions important for personal super intelligence; management thinks Muse Wonder is proof of the importance of Meta training its own models; Muse Wonder has led Meta AI’s engagement per person to more than double since April; the introduction of Muse Wonder, Meta’s latest frontier model, has improved Muse’s quality, engagement, and retention; management believes Muse will perform better over time as Meta releases better models; management plans to focus on pushing the frontier with Meta’s models as well as distilling the frontier models into smaller, efficient product models for personal agents; Meta is already working on models larger than Muse Wonder that are impressive and capable; Meta is already working on smaller models that will deliver similar intelligence to Muse Wonder more efficiently

We have implemented an eval with Muse Wonder to improve how effectively it can autonomously improve our recommendations and ads systems. And as our upcoming models get even more advanced, I expect that we’re going to see an increasing rate of improvement in these systems…

… Muse Wonder is state-of-the-art across many dimensions that matter for personal superintelligence. This validates the importance of training models ourselves to deliver the best experience. In addition to Muse, we’ve also seen that engagement per person in Meta AI has more than doubled since April when we started rolling out our new models…

…We have also seen that Muse Wonder improves quality, engagement, and retention further, which suggests that as we continue to improve intelligence, Muse will perform better and better over time…

…We plan to focus on both pushing the frontier of intelligence and distilling those models into much smaller and more efficient product models that will be optimized for personal agents. My view is that we will be able to build superintelligence through scaling, so we are methodically working our way up the ladder there. We are already working on larger models after Muse Wonder, and they’re going to be very impressive and capable. We’re also working on new smaller models that are coming out soon that will deliver Muse Wonder-level intelligence more efficiently. 

Meta has been shipping many models in the last few months

I’ve been very pleased with our pace of execution here. In the last few months we’ve shipped Muse Spark 1.2 and 1.3, Muse Code, two open weights models with Muse Glimmer and Muse Spark Open, and also Muse Voice. 

Meta’s management has refocused Reality Labs on delivering personal superintelligence; the Muse agent will soon appear in Meta Glasses; management expects Muse to be the most important experience within Meta Glasses; management sees Muse as being a central pillar of the Meta Glasses business; management does not want to make a large profit on the sale of the Meta Glasses hardware; management wants to make profit on Meta Glasses through subscriptions to Muse; Reality Labs will now be shipping new devices only if it can achieve major breakthroughs; Reality Labs recently introduced the Meta VR Glasses, which is the first ever VR experience in a pair of glasses; management expects Reality Labs’ operating loss to be $15 billion in 2027 (was $19 billion in 2025) and to trend towards profitability from there

Our work on personal superintelligence is now also the focus of Reality Labs. Muse is coming to Meta Glasses starting next month, and that will be an important step towards this. People already love many features of the glasses already — music, calling, capturing moments, hearing assistance, and more — but over time we expect Muse and having your personal superintelligence with you throughout the day will be the most important part of the experience. Muse is also central to the business model for the glasses. Again, our mission and strategy are to serve everyone here. We believe in putting the power of this technology in everyone’s hands. So we have priced our devices not to make a large profit on the hardware, but to build a large software and superintelligence business on top of them over time. Glasses sales continue to grow rapidly, and now with Muse, we expect that paid subscriptions will increase and that this business will begin improving meaningfully from here.

I also shared at Connect that we are changing our approach to virtual reality to only ship new devices when we have major breakthroughs like we had with the Meta VR Glasses this year — which of course are the first VR experience ever in a pair of glasses…

…We expect that the P&L of Reality Labs to improve from here, with operating loss coming down to around $15B next year and trending towards profitability from there. 

Meta continues to see gains from its content recommendation initiatives; global time spent per user on Instagram is at an all-time high and was up double-digits year-on-year in 2026 Q3; time spent on Facebook Reels reached all-time highs globally and in the US, and both were up double-digits year-on-year in 2026 Q3; Meta’s upgraded model architecture that ranks sequences of Reels at a time helped grow time spent on Instagram Reels by over 4% in 2026 Q3; 45% of all recommended content on Instagram Reels is now less than one day old, up from 34% in 2026 Q2; Facebook Reels watch time in the US increased by 1% from a doubling of the amount of history the ranking model can consider; Meta has used LLM-based content-understanding systems to improve engagement with Instagram search results by 1% in 2026 Q3; management has started moving Meta’s content-understanding models into the Muse family; the Muse family of models has reduced per-unit content processing costs by more than 80% across Meta’s apps

On the first, we continue to see significant gains from our content recommendation initiatives. On Instagram, global time spent per user reached an all-time high and grew double-digits year-over-year in Q3, with strength across both Reels and Feed. On Facebook, Reels time spent reached all-time highs both globally and in the US, with both growing double-digits year-over-year…

…Engagement gains across our services continue to be driven by improvements we’re making to the relevance, personalization and timeliness of the content we recommend. Investments into our recommendations engine across retrieval and ranking have continued to yield positive results on Instagram and Facebook.

For instance, on Instagram Reels, our relevance launches in Q3 increased Reels time spent by over 4%, driven in part by our upgraded model architecture that ranks sequences of Reels at a time instead of evaluating each Reel independently, as well as improvements we made to add richer watch time signals. On Facebook Feed and Reels, we completed the global launch of a new generative retrieval architecture which, inspired by LLMs, labels pieces of content with semantic representations and generates potential recommendations, helping our systems surface relevant new posts before they’ve accumulated significant engagement. This new architecture is now rolled out to roughly one-quarter of unconnected Feed views and half of Reels views and increased Reels watch time by 1%…

…Nearly 45% of recommended Instagram Reels are now less than one day old, up more than 11 percentage points from Q2. On Facebook Reels, we doubled the amount of history our ranking model can consider, driving a 1% increase in Reels watch time in the US…

…We continue to process public Instagram posts each day through LLM-based content-understanding systems. These signals have improved the engagement with Instagram search results by 1% through the end of this quarter. This quarter, we also began moving our content-understanding models onto our first-party Muse model family. In offline evaluations, these models are outperforming our prior production models and are reducing per-unit content processing costs by more than 80% across Instagram, Threads, Facebook, and ads. Beyond improving quality, owning more of this stack gives us greater flexibility to develop models around the signals that matter the most for personalization.   

Meta’s management introduced Meta Generative Recommender, which uses LLMs to reason about ad content and user preferences simultaneously, in 2026 Q2; management launched the same approach to generative retrieval on Facebook Feed in 2026 Q3, driving a 2% increase in advertising conversions across targeted segments; management has expanded Meta’s GEM (Generative Ads Model) model to additional surfaces, driving a 1.7% increase in conversions for offsite conversion campaigns and a 9.6% increase in app ad conversions; improvements to Andromeda in 2026 Q3 increased conversions by 1.4% on Facebook

Last quarter, we introduced Meta Generative Recommender, which uses LLMs to reason across ad content and user preferences to predict the best ad for each person. In Q3, we launched this approach to generative retrieval on Facebook Feed, driving a nearly 2% increase in conversions across targeted segments. In addition, we expanded our GEM model for ads ranking to additional surfaces, driving a 1.7% increase in conversions for offsite conversion campaigns and a 9.6% increase in app ad conversions…

…In Q3, improvements to Andromeda such as scaling model complexity and throughput increased conversions by 1.4% on Facebook. 

Advertisers who adopt Meta AI business assistant’s recommendations had 5% more conversions at the same cost per conversion and increased their advertising investment by 9%; management will be rolling out more features to Meta AI business assistant; maple syrup producer Simply Organic Maple used Meta AI business assistant for campaign messaging and saw a 57% increase in purchases and a 36% reduction in cost per purchase

Meta AI business assistant is becoming a more proactive and capable partner for advertisers. We estimate that advertisers who adopted its recommendations drove almost 5% more conversions at the same cost per conversion and increased their advertising investment by 9% compared with similar non-adopters. We are rolling out additional capabilities, including creative insights, stronger business context and more agentic actions. These features will allow marketers to plan, create and optimize campaigns, as well as schedule tasks and alerts, through a conversational experience with approvals and previews that keep them in control…

…Simply Organic Maple, a family-run maple syrup producer in Vermont, used Meta AI business assistant to improve its campaign messaging. After the business assistant recommended testing longer, more descriptive advertising copy, the business saw a 57% increase in purchases and a 36% reduction in cost per purchase. Its strongest revised ad delivered a 3x return on ad spend, a 52% improvement from the original.  

10 million small businesses are now using at least one of Meta’s Gen AI advertising creative tools (was 9 million in 2026 Q2)

In Q3, nearly 10 million advertisers used at least one of these tools. Advertiser adoption of our video generation features grew 25% quarter-over-quarter, driven in part by ready-made videos that reduce generation time to just a few seconds from a single image upload. 

Guidance for capex in 2026 has been raised to $135 billion to $145 billion (previous guidance was $130 billion to $145 billion) or growth of 87%-100% (2025 capex was $72.22 billion); management expects to continue aggressively investing in AI infrastructure in 2027; management thinks Meta’s AI investments are already delivering substantial incremental advertising revenue; management thinks Meta’s AI investments for MSL (Meta Superintelligence Labs) can unlock significant new revenue opportunities while strengthening the core business, but the monetisation will take time; management expects to sustain strong overall revenue growth in 2027

We anticipate 2026 capital expenditures, including principal payments on finance leases, to be in the range of $135-145 billion, narrowed from our prior outlook of $130-145 billion. 

Looking beyond 2026, we are still developing our budget and capacity plans for next year, but we expect to continue investing aggressively to support both our core business and emerging AI initiatives. Our current capacity and what is scheduled to come online gives us the flexibility to allocate resources across model training, core business growth, and the development and support of new products. Our AI investments are already delivering substantial incremental advertising revenue through stronger engagement, better recommendations, and more efficient monetization. Building on this progress, our MSL investments are creating the technology and infrastructure needed to unlock significant new revenue opportunities and strengthen capabilities across the business. While these investments will take longer to generate returns than those in our core business, the breadth of monetization opportunities and internal efficiencies gives us confidence in their long-term potential. 

Looking ahead, although we are not providing a full-year 2027 revenue outlook, the benefits our AI investments are delivering to the core business, combined with emerging opportunities from model advances and product launches, position us to sustain strong overall revenue growth throughout 2027.

Meta’s management sees 3 broad themes for how its core advertising business can continue improving; the 1st theme is around traditional machine learning, where management sees a robust and healthy pipeline of improvements to be made; the 2nd theme is around incorporating LLMs into Meta’s ranking and recommendation engines; the 3rd theme is around generative AI tools found in the Advantage+ suite that make it easier for advertisers to quickly and seamlessly generate creatives

I think our ads business has been benefiting for now many years from our ongoing investments in both kind of what we would have called traditional AI machine learning models and then, of course from the sort of more recent generation of models. I would say that there are three themes there that I think are worth outlining. 

The first is just the kind of traditional roadmap of performance improvements that we have in the pipeline based — that is what I would describe more as business-as-usual machine learning work. That continues to be a robust and healthy pipeline. And as we are looking into 2027, again we’re not providing an outlook, but I would say that our investments and product improvements across both engagement and ads initiatives give us confidence that we have meaningful runway to grow our core business throughout 2027…

… The second theme I would say is that we are excited about some of the new research bets that we have as we move into sort of a more LLM-oriented AI infrastructure to help us more deeply reason about people’s preferences and what they’re interested in both across organic content and recommendations, and we expect that we will be able to unlock benefits to our ranking and recommendations engine from that research, too…

…The third thing is that we expect that generative AI tools which have already very valuable for advertisers will continue to grow in terms of the features that we make available to advertisers primarily through the Advantage+ suite but features that make it easier for advertisers to quickly and seamlessly generate, whether it is creative or take an image and generate video or generate sound, all, I think in service of making the ads increasingly personalized for the person that they’re being served to.  

Taiwan Semiconductor Manufacturing Company (NYSE: TSM)

TSMC’s capital expenditure is always in anticipation of growth in future years; management has maintained capex guidance for 2026 at US$60 billion to US$64 billion; TSMC’s capex in the last 3 years was ~US$100 billion, and management continues to expect the next 3 years to have much higher capex (previous guidance was for it to be “much higher”); the majority of TSMC’s capex is for long-term demand and management is also looking for revenue growth to outpace capex growth; management expects TSMC’s business to continue to be supported by leading-edge nodes, including 2nm; management now thinks the AI accelerators business is likely to have a CAGR for 2024-2029 of even higher than the high-50% (previous expectation was for even stronger CAGR for for 2024-2029 than the high-end of the mid-to-high-50% CAGR communicated in the 2026 Q1 earnings call)

At TSMC, a higher level of capital expenditures is always correlated with higher growth opportunities in the following years. Given the continued robust multi-year demand from our customers, including the emerging agentic AI market, we reiterate our full year 2026 capital budget to be between $60 billion and $64 billion. We continue to expect our CapEx investment from 2026 to 2028 to be even more significantly higher than the past three years as we continue to invest heavily to support our customers’ growth…

…The majority of the CapEx is always related to the long-term demand. That will be the composition going forward. Capital intensity-wise, as we always said, if do our job right, the increase in revenue will outpace the increase in CapEx, and that is still the case from what we are looking at…

…Moving into fourth quarter 2026, we expect our business to be supported by continued strong demand for our leading-edge process technologies, including the steep ramp-up of our 2 nm technology…

…[Question] Any update on your AI CAGR growth? I recall previously we talking about like a high 50%.

[Answer] All I can say right now is the demand is extremely robust, okay? We did not see any sign of slowdown. Whether it is a high 50 CAGR, it does not matter. It probably will be higher…

…There is no ceiling for TSMC’s profitability, but a dramatic increase will not happen overnight either.

TSMC’s management thinks recent developments in the AI market are very positive for AI’s long-term demand, in particular, the rise of agentic AI; management thinks there’s great adoption of open source AI, which will lower the barrier to AI usage and lead to more demand for leading-edge silicon; management sees agentic AI and networking driving further silicon demand, in the form of CPUs, and networking chips; management thinks the value of semiconductors is set to increase in the future because of the entire AI trend; management thinks AI-related demand is structural and will last for multiple years; management thinks TSMC is well-positioned to capture the full AI-related demand since all AI players are working with the company; management has not included CPU demand in their assessment of TSMC’s AI-revenue CAGR

The AI market continues to be very strong. Recent developments are very positive to AI’s long-term demand outlook. We observe increasing AI demand fueled by the rising adoption of agentic AI as we transition from chatbot to AI assistance. In addition, we are observing great adoption of open source AI models. This will drive greater efficiency and help lower the barrier to future AI usage, and result in more and more demand for leading-edge silicon. Thus, AI-related demand is being supported by strong underlying demand for AI accelerators. In addition, the emergence of agentic AI CPUs and the continued demand for networking chips are driving further silicon demand. Therefore, semiconductor will remain essential in the AI mega trend. The value of semiconductors in the global supply chain is set to continue to increase in our view. This will create more value opportunities for our customers to capture and more value opportunities for TSMC as well.

No matter which approach is taken, you will require use of the most advanced semiconductor process technology. With our technology leadership, manufacturing excellence, global capacity expansion, and customer trust, we are well-positioned to capture the full multi-year structural AI-related demand. Therefore, all the AI players are working with TSMC…

…[Question] On the agentic AI demand on the CPU. So I wonder, according to our previous framework, demand forecast framework, the mid to high 50s growth including the CPU or not, and if we heard our clients increase the CPU forecast significantly, when is the next time we refresh our AI demand growth numbers?

[Answer] It is not included in our AI demands calculation. The reason is very simple, because we cannot identify which CPU user in the conventional server CPU or AI server CPU. Also, it depends on our customer’s application and their own allocation. So, no. So far today, we stick to our initial very narrow definition on the AI’s demand.

When planning for TSMC’s manufacturing capacity, management collaborates closely with its customers, and customers’ customers, and now they are even checking on the development progress of data centers and power supplies in the USA; management thinks the cloud service providers (CSPs) have been well-prepared; management has a very good idea of the multi-year product roadmap and production plans of TSMC’s customers and continue to hold strong conviction in the AI mega trend; management is moving at full speed to increase TSMC’s capacity, with TSMC expanding at a record pace; TSMC has nearly 20 fabs in progress worldwide, with 13 leading-edge and advanced packaging fabs in Taiwan to be built over the next several years; management intends to capture every piece of business possible; management thinks the demand for TSMC is still higher than supply and they are working very hard to close the gap

To address the structured increase in the long-term semiconductor market demand profile, TSMC collaborate closely with our customers, yet our customer’s customer to plan for our capacity. In addition, we are now going one step further and sending our own teams of people to check on the data center build-out progress and power supply situation in the U.S. The feedback is so far so good, and the CSPs has been well-prepared. Internally, TSMC employed a disciplined capacity planning system to assess the market demand from both a top-down and bottom-up approach. This is a continuous and ongoing process. Given the fundamental complexity of leading-edge technologies and the design in and lead time involved, we already have a very good idea of customers’ multi-year product roadmap and production plans. Thus, we continue to hold strong conviction in the AI mega trend.

We are moving full speed ahead to increase our capacity to meet our assessed demand, and stepping up our CapEx investment to support our customers’ future growth. TSMC is expanding at unprecedented pace with a nearly 20 fabs worldwide in progress. We are building 13 leading-edge and advanced packaging fab in Taiwan over the next several years, and another five to six fabs overseas…

…We have no intention to leave food on the table for others, and we want to capture every piece of business possible…

…We have a very sophisticated capacity planning process, and we also have a very disciplined on our consideration. Let me share with you, we put every possibility we can, talking about the CSPs, CapEx, data center build-out, financial, power supply, and also other component supply, et cetera, into our consideration. We make our own judgment, and then based on our assessment the demand, we still think the demand is higher than the supply, and we are working very hard to narrow the gap.

TSMC’s management will only be utilising ASML’s high-NA EUV lithography systems in 2030 because they think that is when the technology will be sufficiently mature with a good cost curve; management is confident that TSMC’s relatively late adoption of high-NA EUV compared to competitors will not disadvantage TSMC

[Question] You did touch upon it, the High NA agreement with ASML from 2030 onwards. I’m just curious, historically, TSMC was more ahead of the curve with the EUV adoption. It looks like by the time you adopt High NA EUV, it would be 2030, 2031, but some of your competitors would already been using it by then. Do you see that as a disadvantage, or you’re not worried about competition getting ahead with High NA EUV?

[Answer] We always evaluate a new tool, and if we want to put into the mass production, it got to be the best technology with a very reasonable cost, and most important is the best maturity. Okay? We say in 2030 we will start to use it because we also think at that time, High NA will be mature, and we are also working very closely with our ASML to establish a 12 in photomask by 2031. At that time, we think that High NA’s cost will be reasonable and for TSMC to make it into the mass production. This is a very good tool, let me emphasize that. In the smaller geometry, it definitely has some advantage. Whether that our technology, because we are using it later than our competitor, so disadvantage to TSMC? No, because we have confidence that we are still the best company to offer the most advanced technology to our customers.

TSMC’s management thinks the company’s advanced packaging technology is very competitive; management is willing to be the backup plan for some customers that are using Intel’s EMIB-T advanced packaging technology

[Question] If some customer, they initially choose to use Intel’s EMIB-T, and for some reason cannot get to sufficient demand, are you also support customers for the CoWoS version or capacity.

[Answer] I still think that TSMC’s advanced assembly, or they say packaging technology, are very, very competitive. But the most important thing is that we using that one to help us help our customer’s product can be introduced in the market smoothly. That’s our number one priority. We have to offer them the best technology with enough capacity. So far, let me share with you today’s situation. We are in the real-world production of 5x reticle size. I don’t believe that anybody can do that today except at TSMC, and the yield is excellent. So our customer are very satisfied with the result. We are working on the 9x reticle size and moving into the 14x reticle size. Of course, we are also working on the CoWoS, and to make sure that in the future we can offer a reasonable cost reduction. But so far, the CoWoS is still the number one choice for our customer. As time goes by, yes, our competitor offer a very competitive solution. Still, I want to say that today’s situation, we are the only one that makes sure that the advanced packaging is a 5x reticle size, and still, we are very competitive. I want to emphasize that…  Let me say that, yes, we will be. Number one priority I already indicate that TSMC’s customer’s product can be introduced into the market smoothly.

TSMC’s management thinks the learning curve for the company with the new technology nodes is actually better now with the current situation of having both smartphone and AI HPC (high-performance computing) applications drive demand for the new nodes

[Question] In the past a new node is mostly ramped with smaller die size and simpler design. But today we have both the smartphone and also with the AI and HPC applications coming in. Does that increase the complexity of the node process architecture, make the learning curve more challenging or difficult, and therefore the implication to the margin?

[Answer] Today’s situation actually is a little bit different from previous years. Yes, we now have a very strong demand from smartphone and also AI HPC applications. So the smaller die size per se or the big die size in the accelerator, they happen at the same time. This is to TSMC’s advantage, let me say that. Because we are good. We have been prepared for this kind of thing, and happy to working with us. So the learning curve, yes, but we are doing a reasonable job, and so the customer are very happy today.

TSMC’s management thinks the shortage of power supply in the USA for AI data centers does not drive additional demand for the company’s backside power rail solution

[Question] Could this tighter and tighter power grid surprise make your customer actually accelerating adoption of your Super Power Rail product for the better performance power?

[Answer] Whether this kind of situation change our customers’ behavior and move more into the backside power rail? Answer is no. Whether you use the backside power, those kind of metallization scheme is based on your need for your product. Okay? Some of the accelerator might be using it, some of the product like mobile phones, they don’t use it. So it depends on customer’s product.

Tesla (NASDAQ: TSLA)

Tesla’s management sees Full Self-Driving (FSD) software as the  biggest driver of demand for the company’s vehicles; management thinks there’s no competitor to Tesla’s FSD; FSD is approved in many countries in Europe, but Tesla is pushing for an EU-wide approval; management thinks an EU-wide approval could happen in the next few months; more than 75% of Tesla vehicle deliveries in North America in 2026 Q3 included a paid FSD subscription at the time of delivery, up 20% from 2026 Q2; every Tesla car delivery comes with a free month of FSD, but vehicle owners are still signing up for the paid subscription; FSD now has 1.7 million paid customers globally, and more than 50% of them are paid subscriptions; management wants to sell certain cars in Europe with some FSD benefits before a broader approval in Europe is obtained; Tesla is building AI-based mapping and navigation into FSD v.15 (version 15) and it will be very personalised, and will also tap on the collective intelligence of all Tesla cars; management thinks FSD is an incredibly hard problem to solve for other vehicle manufacturers; management is seeing many customers walk into Tesla’s showrooms and ask for an FSD-drive, instead of any specific car model; management thinks Tesla’s FSD is much better than what any other Chinese car company can deliver; FSD is already approaching $1 billion in ARR (annualised recurring revenue)

Obviously, we are going to be ramping, and we are ramping production meaningfully. The biggest driver for demand is Full Self-Driving. Today, there is really no competitor for FSD in any car that you can buy…

…We are pushing very hard for FSD approval in additional markets. We are now approved in many countries in Europe, and we are hoping for an EU-wide approval. It is hard to predict, but maybe within a few months, but I would be surprised if it takes longer than Q1. Obviously, that is not a timeframe we control, but we do anticipate being fully approved throughout Europe and many other countries…

…In Q3, over 75% of deliveries in North America included a sign-up for paid FSD subscription at the time of delivery, which is approximately 20% higher than Q2. Overall, the FSD attach rate continued to improve, reaching over 1.7 million paid customers globally, of which for the first time, 50% is based on subscriptions. Given the strength of FSD in the U.S., we are taking extra steps to sell certain cars in Europe with some benefits to try and get more fleet out there before a broader FSD approval is obtained…

…One thing which I’d also like to make sure it’s clear when I say in North America that 75% attach rate at the time of delivery, this is every car which we’re delivering comes with a month of free FSD subscription. Still, people are signing up even though they have that free available…

…We are building our own maps, but not the same way as OpenStreetMap. We are building an entirely new architecture with essentially AI-based mapping and navigation. It’s going to be super awesome. It’s going to be integrated with voice. You can talk to it, ask for different routes. It’s going to learn your personalized preferences, plus also the fleet preferences. The collective intelligence of all the cars are going to flow into making the maps and navigation be super cool. You’ll see the start and end of trips already use this technology pretty soon as part of the V15 release. Over the next several months, it’s just going to take over all of the navigation functions, parking, destination handling, and just will know you better than you know yourself in some regard for the driving task…

…It’s easy to get to the first 90% and make a cool-looking demo. But when you get in the long tail of this problem, it’s so long that people don’t realize anyone with their sane mind should not attempt to solve this problem. It’s so difficult, and so long tail. Especially to get to a safety point where you can drive safely for millions of miles with no injuries and so on, that’s going to be extremely hard. As we release unsupervised FSD to the broader customer fleet, it’s just going to be I don’t think anyone can match it in several years…

…Especially in our showrooms here in the U.S., customers are now coming in and saying, “Can we get into an FSD drive?” They’re not asking for 3/Y or Cybertruck. They’re saying, “Can we get an FSD drive?” That has become the key driver because what it does is it actually opens up a lot more flexibility for different parts of the population wherein they had issues driving conveniently before. Now they’re getting their freedom back by being able to use FSD…

…The leapfrog which we have in terms of this technology is not there with any of the other Chinese OEMs today. We are working very diligently to try and get broader FSD approval in China. It is hard but for certain cars, we have been able to have FSD operating and whoever has tried it there, they understand how much better it is as compared to any of the competition. In fact, I do not know if you have seen this, there have been some competitor CEOs who have come and done FSD drives here in the U.S., and they themselves have been blown up as to what FSD is capable of doing…

…[Question] Are you at some point going to disclose for us the annual recurring revenues that FSD represents. I did my own math, and I’m thinking it’s probably now approaching $1 billion of ARR. Is that the right order of magnitude?

[Answer] I think the numbers which you shared are correct. As it scales further, we will start sharing more metrics around FSD, not just the revenue, but also other key metrics around it.

Robotaxi’s fully autonomous miles driven are scaling rapidly on a weekly basis; management thinks no other company can match Tesla’s scale with autonomous vehicles; Tesla’s Cybercab fleet in Austin has grown by 10x since launch a month ago; Cybercabs are in use at scale in Austin, Texas; Robotaxi’s unsupervised miles increased by 8x in 2026 Q3; management is preparing for a broader Cybercab launch in other markets; management thinks Cybercab provides the best Robotaxi experience, but they will augment the Robotaxi fleet with other vehicles over time; management’s ambition with Robotaxi is to take over every part of terrestrial transportation, not just transporting people, and so they are preparing a huge investment cycle for the business; management launched Cybercab with the help of data from Tesla’s other vehicle platforms; management is seeing very good scaling in Cybercab’s AI capabilities; Cybercab will be launched in Dallas in November; management thinks consumer vehicles with FSD can become part of the Robotaxi fleet in 2027 Q2; management is trying to get the Cybercab to customers by the end of 2026; management sees most Cybercabs as being for Tesla’s internal use to deliver transport-as-a-service; the Robotaxi service is still not at the $0.30 per mile cost-target management has, but they see a path to get to the target; management has released a safety report on unsupervised Robotaxis in China, and the data is hard to refute by any regulator; management thinks Robotaxi can start moving the needle in Tesla’s financials by 2027 Q2; Tesla’s engineers think scaling the Cybercab is going to be relatively easy; the Cybercab has 50% less parts than any other Tesla vehicle so that’s 50% less things that can go wrong; the listed Cybercab capacity of 125,000 units a year is not near the actual capacity of that vehicle line; there are still software challenges for the Robotaxi service; management thinks Tesla will end up doing most of the passenger miles on in any country it is allowed to operate freely

We are seeing the Robotaxi miles driven or fully autonomous miles driven scale exponentially, meaning scaling very rapidly. It would be hard to overstate at this point, it really is an extremely fast week-over-week growth in autonomous miles. I do not think there is any company in the world that can match our scale with autonomous vehicles as we continue to ramp. The Cybercab fleet in Austin, for example, has grown by a factor of 10 since launch about a month ago…

…Come into Austin and you’ll see Cybercabs in use at scale in Austin. As I think a lot of people saw with the Austin City Limits concert recently…

…Accumulative unsupervised miles 8x in the quarter, as we had weekly double-digit growth throughout the quarter. We’re taking all the learnings from our Austin expansion to help us prepare for broader Cybercab launch in other markets. Our priority is progressing the density and expansion of the network through vehicles that provide the best experience, like the Cybercab. We will also continue to augment the fleet with other variants over time to address the terrestrial transportation needs of the total population. I’d like to stress that what we are trying to do is literally take over every part of terrestrial transportation. Not just getting people. We’re starting with getting people from one place to the other, but we want to basically own this transportation of any kind on Earth. We are in a huge investment cycle to set us up for the future…

…We launched Cybercab this quarter, which is very exciting. We achieved that by mostly learning from data from the other vehicle platforms. For example, Model Y, Model 3, et cetera. All of the data helped make the Cybercab without requiring a ridiculous amount of vehicles to produce the data. We are seeing very good scaling in terms of Cybercab’s AI capabilities. We launched in Austin, but that’s only because we had some cabs, and we want to swarm Austin with Cybercabs everywhere. There’s no fundamental reason that we are limited to Austin. In fact, next month we’ll be launching in Dallas…

…The software for Cybercabs and Model Y is the same. Pretty soon we will be able to release the same software as part of the V15 release to customers. The only additional requirement is to make sure that, for example, the customer Model Ys can be part of the geo-fences and the fleet management that Tesla has for it to be fully autonomous. We suspect somewhere around Q2 next year is when we can turn on some of the first customer vehicles to be part of the Robotaxi fleet…

…We are trying to get Cybercabs to customers by the end of this year, 2026. Obviously, we are producing them in our line in Texas, but all of those are currently going to our own internal Robotaxi fleet. As Tesla grows, we will mainly deliver fully designed autonomous vehicles to our customers. Cybercab will be the first…

…The vast majority of the Cybercab stuff will be essentially Tesla-owned and transport as a service…

…Week by week, as Elon was talking about, we look at the miles and how many paid rides are happening. Are we at our goal of $0.30 today? No. Do we see a path to get to that? Yes, we do. It is all about scaling operations. What you will see over the next few months is that today, Cybercab, which is the primary Robotaxi product, is available only in Austin. We do plan to bring it out in some other cities before the end of this year. As we add more Cybercabs, we are also learning from the operational experience of managing a fleet, pick-up and drop-off, cleaning, all those things. All those costs are being dialed in, and that is why, like I said before, we definitely see a path to get to our $0.30 target…

…We are actively working with the regulators in China because I do not know if you saw this, but we have released our safety report on FSD before, but we have also released our unsupervised, our Robotaxi safety report. We just think this is the right way to get people into a safe environment because you know the amount of accidents which happen because of driver that was not being alert. This is the only way we can ensure that there are less casualties in the world. Any regulator who looks at the data, it would be very hard for them to refute it…

…[Question] I am curious if there is a way you can triangulate that for us into revenue and maybe profitability of the Robotaxi business. If you sustain, let us say, 15% week-over-week growth in these miles, you will probably cross 10 million miles by the end of the year. How many miles would you need to generate gross profit?

[Answer] I would say it’s not going to move the needle this quarter. But if you say maybe, I think it starts to become noticeable probably in Q2 and then probably a big deal by Q3 or Q4 of next year…

…[Question] How do you expect the ramp-up curve to be on the Cybercab line?

[Answer] I would give kudos to the engineering team. They have designed a product because not only do I talk to the engineers, I also talk to their manufacturing guys. They think this is one of the easiest product for them to scale in terms of manufacturing. Scaling would not be that much of a challenge… We have 50% less parts than any other vehicle we’ve ever built, so that’s 50% less things to go wrong. We’ve definitely been bursting the Cybercab line to see how fast we can run it. I think the listed capacity is more of a rate at which we’re capable of running today, but the actual capacity of that vehicle line is significantly higher…

…[Question] A follow-up on Robotaxi. It does seem like you’ve hit an inflection point here in Austin over the last month or so since you launched the Cybercab. Would you generally say the bottlenecks at this point to growth are more regulatory-driven as you think about expanding into new geographies, and that you’ve kind of gotten through most of the hurdles on the software side?

[Answer] For example, we had initially had the Robotaxi service from 6:00 A.M. to 9:00 P.M., then we expanded to 10:00 P.M., 11:00 P.M. Now it’s going to be pretty soon 2:00 A.M. So we are slowly expanding the availability of Robotaxi, and that is driven by the capability of the AI software. We just want to make sure that there’s not a single safety incident. That is our operating principle. Even if there’s the slightest possible risk at all, we don’t take on that risk. So we are very risk-averse in that sense, as in we want zero injuries or anything bad happening. That is the driving principle. We will see that over the rest of this year, we’ll unlock 24/7 operation, go to higher speed roads. We think all of that is necessary to scale and offer a great product to a lot of users. That’s why we wanted to both do the breadth-first of doing multiple cities. We do six cities now. We are also doing the depth of having a lot of vehicles in a single city, which is in Austin. We’ll expand to more weather conditions, day and night conditions, highways and high speeds, and all of that will be done hopefully by the end of this year, at which point you will see a lot more scaling, a lot more cities being covered by Robotaxi. So it’s not purely regulatory or something like that. It is technology-driven, but we do see that pretty quickly become regulatory, perhaps in Q1, Q2 timeframes…

…Unless we are stopped by governments, which is certainly a possibility, if they are trying to protect a domestic industry. I think Tesla will end up doing a majority of all passenger miles on Earth in any country where it is allowed to operate freely.

Tesla has started early production on the high-volume Optimus line in Fremont; management thinks the ramp of Optimus’ manufacturing will depend on the slowest or unluckiest component involved in the manufacturing process; Tesla already has the vast majority of a million units per line production line landed in Fremont, but management thinks it will take a year or more to get to that annual production rate; management is still tweaking Optimus’ design, and AI software, and believe they will be done by the end of 2027, which is when Optimus can be delivered outside of Tesla; management thinks Optimus can get the training data needed only by Tesla manufacturing 10,000-20,0000 of the robots and having them do self-play; management is building the Optimus Academy for the Optimus robots to practice tasks and these data can be used to improve Optimus’ capabilities; management is convinced that Optimus will be the biggest product on Earth; management thinks there’s no other general-purpose humanoid robot that is close to Optimus; management thinks Tesla can release a video of Optimus doing things it was not pre-programmed to do in about 1-2 months’ time; management thinks a lot of humanoid robot demonstrations are pre-programmed, and the robots are not really using any advanced AI or SI (superintelligence); management thinks Tesla will integrate the Tesla AI 5 chip into Optimus in April 2027; management thinks the AI 5 chip will improve Optimus by an order of magnitude because AI 5 has an order of magnitude more compute than AI 4; management thinks Optimus’ initial and most valuable customers will be industrial companies; the hand of the Optimus robot remains one of the toughest problems for Tesla to solve; management is already building Optimus robots for training purposes even when the hands are not perfected; management thinks there will eventually be more humanoid robots on Earth than humans; Tesla is building a new Optimus factory in Fremont that will have a much larger capacity than the Austin factory

Regarding Optimus, we’ve started what I’d call early production on our high-volume line in Fremont. I want to elaborate on this point because I have said this before, that the growth of Optimus, it’s an S curve, but the initial part of the S curve has a shallow slope and is somewhat bumpy because it is such a radically new product. The way to think about it is there’s 10,000 things that have to go right for production of the robot, and it’ll move as fast as the slowest, least lucky of the 10,000. We do have the vast majority of a million unit per line production line landed in Fremont right now. I would expect that it takes us probably at least a year, maybe a little longer, to get to that annualized rate, to get to 20,000 units a week…

…There’s still some things we need to tweak about the design because you need to deliver a robot that is safe and reliable and has long duration. It can’t wear out after a month. There’s still quite a lot of work to do with regard to the AI and the robot for it to be truly useful for open-ended tasks. Like I said, I would expect that the AI, the electromechanical design, and the reliability and usage duration, the usage life of the robot, I do expect these issues to be solved around the end of next year. That’s the time to expect when we’re delivering Optimus to companies outside of Tesla…

…There’s a lot of things we have to get right about the AI architecture that we’re still working on. We’re making good progress, but that needs to be sorted out. Really, in order for Optimus to get the training data we need, we actually need to make a lot of Optimi. We need to make, I think, on the order of 10,000 or 20,000 bots and then have them do self-play. We’re building out an Optimus Academy, where Optimus can practice doing a very wide range of tasks. Then we can use that data in sort of a flywheel to make it better and better at these various tasks. A big difference with Optimus versus the cars is we don’t have this vast fleet of cars from which we can gather training data. With Optimus, we have to bootstrap it by making the robots and having them do self-play, which is really how a human learns to navigate the world…

…I remain convinced that Optimus will be the biggest product on Earth, and humanoid robots in general will be the biggest product on Earth…

….To the best of my knowledge, there is no one that is even close to Tesla with respect to a general-purpose humanoid robot…

…I think we may be ready to release an Optimus video in a month or so. I want to make sure that when we release this video, that it is not like a canned demo that. When you see these humanoid robot demonstrations, some of them look very impressive, but they are not really using any kind of advanced AI or SI. They are pre-programmed deterministic tasks, and if you actually try to use that humanoid robot in the real world, you would have a lot of trouble. It would not work, basically. So I want to release a video where it is Optimus doing things that are not pre-programmed. I think that is maybe a month or 2 for that to have a demo that you can believe what you see type of thing.

There is another milestone that I think important we want to get to, which is to integrate the Tesla AI 5 chip into Optimus, which, if things go according to plan, will be in approximately April of next year. That is a giant improvement in capability with an order of magnitude more operations per second. It is an order of magnitude more compute, AI 5 versus AI 4. That will help Optimus a lot because the degrees of freedom that a humanoid robot has to deal with are far greater than what a car has to deal with. It is a bit of a struggle to squeeze everything needed to operate with all the degrees of freedom of Optimus and the wide range of tasks. If you think of a humanoid robot is going to do 1,000 times more things than a car would do. That is some very big number, basically. I think the additional compute of AI 5 will make a big difference…

…In terms of the initial customers for Optimus, I think it probably would be things that are industrial applications at first. Things like dangerous repetitive tasks, things that people really don’t want to do, and that you want done basically 24/7. So, around the clock stuff. That is going to be the highest value for Optimus in the beginning. As we scale production, of course, it would be something that people can buy for their home. But the highest value use, I think, will be heavy industry, dangerous tasks, or things that may cause repetitive stress injuries, and that you want done 24/7…

…[Question] A quick follow-up on Optimus. I know you mentioned that we could see a video soon, that there’s quite a bit of progress, but you’re still making some refinements. I guess, just generally speaking, where do you see most of the design changes still happening going forward? Is it still the hand that’s an issue? Anything you can call out there?

[Answer] It is mostly the hand, and just the fine-tuning of the hand design, and the fingers especially. It is a hard problem, because humans have self-repair. Our body repairs itself, but robots don’t. Our fingers are kind of squishy and have lots of sensors in our skin. It is hard to make a hand that can handle lots of wear and tear. But it is still squishy enough to pick up delicate objects, and has all the sensors and doesn’t damage the sensors when you do repeated actions. There is a lot of little things that you realize actually what an incredible thing the human hand is at the end of the day. So it is mostly from an electromechanical standpoint, the hand. There are other things that are improving as well throughout the robot, but that is the main thing…

…[Question] Can you talk about the strategy around ramping Optimus production just a little bit to start training, even if the hands aren’t fully baked with the final design? It seems like there’s a lot that you could end up doing in terms of the training cycle and maturing some of the simulation technology, even without perfected hands.

[Answer] That’s what we’re doing, is we’re building robots that do not have perfected hands, and we are making progress with Optimus training and ironing out all aspects of the supply chain and the production system. We’re doing all these things in parallel, like this giant production line at our Fremont factory, massive supply chain, and then fine-tuning the design and engineering simultaneously with the AI software. All these things are happening simultaneously.

…I think at some point there will be more humanoid robots on Earth than there are humans…

…You should know that not only the Fremont factory, we’ve also started constructing a factory here in Austin, which would be able to do a lot more robots as compared to the Fremont one.

The research fab for Tesla’s Terafab project is progressing quickly; the research fab will commission machinery int 2027 H1 and management hope to start manufacturing in May/June 2027; management’s aim is to manufacture useful chip volume around 2028 Q3; management thinks Terafab will be the most valuable building in the world; management thinks Terafab will be necessary for the scaling of Optimus, as it can address bottlenecks across the semiconductor value chain; management thinks TeraFab will need significant capital to meet the goal of reaching useful chip volume around 2028 Q3, in the order of $100 billion over the long term; Tesla has already placed most of the orders for the first set of equipment for TeraFab

Regarding the Terafab, the Austin-based research fab is progressing quickly. We’ll begin commissioning machinery in the first half of next year, and hopefully, if things go well, we’ll make our first test chip just to see if we can make any chip at all around May or June. Our aspiration is to be making some kind of useful chip at volume around Q3 of 2028. As I’ve said publicly on X, the Terafab I think will be the most valuable building in the world and will probably have an economic value that will place it in the top 20 of countries. So it’s an epic project. This will be an incredibly important asset, if not a necessary asset, in order to scale, especially Optimus, as we can address bottlenecks in semiconductors, whether it’s memory, packaging, logic, or whatever the case may be…

…[Question] You mentioned that you had an an aspirational goal of 2Q 2028. Do you have a sense on how much capital needs to be invested to get there? Also what kind of milestones we should be on the lookout for on the way there?

[Answer] It’s going to be some pretty significant capital. It’s certainly in the tens of billions for Terafab. We’ve got to figure out what the relative split is between SpaceX and Tesla in terms of who pays for what at Terafab. But it’s not going to be less than tens of billions in capital equipment up through early 2028. Then it’ll probably be north of $100 billion long term…

…The first set of equipment for the fab which Tesla is building, we’ve actually already placed most of the orders.

Tesla’s energy business is seeing extreme demand from data centers, and this is driving improvements in selling prices and margins

On the energy business, we had our second-best quarter of deployments of 13.7 gigawatt hour of energy storage… the business is witnessing extreme demand from data centers, and we are doing our best to increase production, and accordingly, we’ll see improvements in ASP and further improvements in margins as we move forward.

---

Disclaimer: The Good Investors is the personal investing blog of two simple guys who are passionate about educating Singaporeans about stock market investing. By using this Site, you specifically agree that none of the information provided constitutes financial, investment, or other professional advice. It is only intended to provide education. Speak with a professional before making important decisions about your money, your professional life, or even your personal life. I have a vested interest in ASML, Intuitive Surgical, Meta Platforms, and TSMC. Holdings are subject to change at any time.

JPMorgan Chase (NYSE: JPM) is currently the largest bank in the USA by total assets. Because of this status, it is naturally able to feel the pulse of the country’s economy. The bank’s latest earnings conference call – for the third quarter of 2026 – was held earlier this week and contained useful insights on the state of American consumers and businesses. The bottom-line is this: the US economy continues to experience strong business investment and consumer spending, but there are risks to the global and US economy, including large fiscal deficits, sticky inflation, rising interest rates, and more.

What’s shown between the two horizontal lines below are quotes from JPMorgan’s management team that I picked up from the call.

---

1. The US economy continued to experience strong business investment and consumer spending in 2026 Q3, but the risks (large fiscal deficits, sticky inflation, high asset prices, rising interest rates, and geopolitical tensions) are shifting; JPMorgan’s debit and credit spend were both up 10% in 2026 Q3, driven by consumer demand; management is not trying to predict higher interest rates, but they think the chance is higher than others think; management thinks if interest rates stay very high for a long time, the consumer will be impacted; the US consumer is currently very healthy because of a strong labour market; management thinks the balance sheets of JPMorgan’s middle-market clients are strong; management sees the corporate credit outlook as stable, even though there are pressures

The U.S. economy continued to be fueled by strong business investment and ongoing consumer spending. However, large fiscal deficits, sticky inflation, elevated asset prices, rising interest rates and geopolitical conditions continue to shift under the surface. These forces may be manageable, but we remain prepared for a wide range of outcomes…

…Debit and credit spend grew 10% year-on-year, with continued momentum across spending categories, including gas and travel, supported by consumer demand…

…[Question] As we assess the risk of interest rates maybe moving even higher than what we saw in the third quarter, maybe your latest thoughts around how you assess the resiliency of the economy to handle maybe a 6% rate on 10-year and just the Fed hiking a few times. 

[Answer] Of course things like that could happen. I would never take that off the table. They could happen for good reason, which is a lot of demand for capital, a lot of productivity, and things are fine. They could happen for bad reason, in sticky inflation around the world. They can happen not because of what is in the U.S. It could happen because of what happens overseas. So we are not predicting it. I have always thought there is a higher chance than other people…

…If you keep real rates very, very high for a long time, at some level, that’s designed to slow down economic activity to fight inflation. It’s hard to imagine that that wouldn’t have an impact on the consumer. So far, what we’ve seen in the current data and what’s been true, I would say, for the last few years is that the consumer has been quite resilient in the face of a number of headwinds that under different circumstances you might have expected would present a challenge. But I think the conditional expectation of consumer credit health as a function of higher rates is a fairly nuanced question. But fundamentally, the picture is pretty simple. The consumer is currently very healthy, primarily because of a healthy labor market. But if the central bank needs things to cool off, and that involves an aggressive extended rate cycle, almost by definition, that’s going to involve higher unemployment, which will produce weaker consumer credit results…

…I think there, the theme for some time now has been people worried about other people while doing fine themselves. The theme is always like, we run these surveys and they’re always like, “Oh, I’m worried about this, I’m worried about that.” And it’s like, “How are you doing?” “Oh, I’m doing great.” So it’s kind of an interesting dynamic there. No question that higher rates, I think input costs, inflation, the diesel narrative, these things are tough for middle-market clients. So we see them trying to manage that to the best of their ability, and I think it does create a little bit of caution in that space. But I think balance sheets are strong…

…[Question] Could you talk about the corporate credit health? It looks like you built some reserves on the wholesale side.

[Answer] The wholesale build is actually just from new activity, so it’s not downgrades or change in the outlook or anything like that. I would say that the corporate credit outlook is currently pretty stable. But obviously there are issues out there, and we’ll have to see how those flow through the performance of our borrowers. Energy prices, fuel inputs, higher rates, these are all things that will create challenges in different pockets of the economy and C&I space, as well as the CRE space for that matter. So we’ll see. But so far, it’s very early obviously, but in general, we think corporate balance sheets are in quite good shape, and so there’s been resilience.

2. Net charge-offs for the whole bank (effectively bad loans that JPMorgan can’t recover) was down at US$2.2 billion compared to a year ago (charge-offs was $2.4 billion in 2026 Q2)

Credit costs of $2.7 billion with $2.2 billion of net charge-offs and a $539 million net reserve build.

3. JPMorgan’s investment banking fees were up 14% in 2026 Q3 from a year ago because of strong performance in advisory; management still has a positive outlook, but they’ve seen a slowdown in capital markets activities, which has led to short-term uncertainty; management thinks the capital markets pipeline is fine at the moment, but it can open or close very quickly

IB fees were up 14% year-on-year, reflecting strong performance across products, notably in advisory. While our outlook remains constructive, supported by ongoing client engagement, evolving market conditions have recently slowed activity a bit, so we’ll have to see how the quarter develops…

…The pipeline is fine, but I pointed out many times before, the pipeline is like an accordion. It can open and close pretty quick. You have seen in the U.S., as Jeremy said, the word delayed, probably. We do not really know. We will see. But I was in Europe recently, and the pipeline there is actually quite healthy, and even more so in IPOs than we would have thought. We just do not know…

…We do not obviously know exactly what investment banking fee revenue is going to be in the fourth quarter. I think that we do know that the overall environment is robust and supportive, and there is a lot of business that wants to get done and needs to get done. As you point out, Erika, not surprisingly, given the rate sell-off, you are going to have the market take a little bit of a breath and a pause to make sure buyers and sellers can meet on valuation terms. We did see a little bit of that slowdown in the second half of September, basically. We will see how that plays out into December. I mean, sorry, into the fourth quarter.

4. Management continues to expect credit card net charge-offs for 2026 to be 3.2% (previous expectation was 3.2%; was around 3.3% in 2025)

We continue to expect the card net charge-off rate to be approximately 3.2%.

5. Management feels good about investing long-term in areas that improve the security and resilience of the USA

I want to brag on the Investment Bank a little bit. They have hired a tremendous amount of very talented people in SRI, and we are lucky to have Todd Combs here, investing some of the capital. They have had a lot of ideas. It is not just meetings at this point. They have helped fund some interesting types of things. We have more ideas. They are doing a really great policy around it too, which is what can you do to accelerate security and resilience in the U.S. around, whether it is shipbuilding or —we had a big announcement in Philadelphia. I am proud of them. We have had a bunch of people who retired, who raised their hand, and they are coming back more for patriotism than anything else. It has been just a great effort for the company. As we pointed out, it is global, and the global response has also been exceptional. I have something across my desk which is, alone, over $10 billion or $20 billion of financing in one country. A lot more to come here. I would put in the same category, American Dream Initiative, the innovation economy initiative, the commercial banking initiative around the world. These things are working. We think we are doing the right kind of good business at good returns.

6. Management is going to be stricter with deploying capital for credit because they are seeing credit terms becoming a little loose; the loose credit terms appear to be related to too much AI-related investment activity

[Question] Just following up on the capital allocation. Maybe a comment or two on where you are less comfortable allocating capital to?

[Answer] I am not going to give you very specific stuff, but we are tightening up a little bit on credit. I think some of the credit terms have gotten a little bit loose. You have heard us talk about that before, and other people. A little too much pay, a little too weak covenants, a little too much priming good assets, a little too much pushing on it. We just do not feel you have to chase that kind of stuff anymore…

…To Jamie’s point about tightening up credit, that is a little bit of an overall point about just making sure we are really scrubbing everything. In this moment where AI is such an overwhelming theme, it is producing so much innovation in so many different pockets. There is capital being tapped through all kinds of different mechanisms. We just need to be really, really on top of things to make sure that nothing slips through.

7. Management thinks this latest rate hike cycle is even more unusual than the last one and there isn’t much historical precedent to study

I think it is an important thing for people to be thinking about because this particular rate cycle, this latest rate cycle, is arguably even more unprecedented than the last one. We just don’t have a lot of historical data about how people behave when you see rate hikes from the launch point that we’ve seen this cycle. So it’s been an unusual 10+ years, I would say, and I think it’s quite important for people to realize, especially people who are doing this by training models on historical data, that the historical data set’s just not that rich. So it’s very important to be kind of thinking out of the box.

8. Management thinks no financial institution is immune at this point to a major change in sentiment in the AI space

…I think there, in addition to the obvious single name focus, the other thing to be very careful about is just the AI theme. Because the AI theme is so broadly diffused into the marketplace. It’s not just traditional construction lending. Now, increasingly, it’s working capital revolvers. You also see it in the equity markets. It’s kind of everywhere. So you really need to be running stresses that test your resilience to a meaningful change in the overall theme across the entire enterprise. That’s something that we spend a lot of time thinking about. Currently, our exposures are still quite modest in the context of our overall wholesale exposure, but that’s not the same as saying that we would be immune to a major change in sentiment in the AI space. I think I would be surprised if anyone, frankly, was immune to that at this point.

---

Disclaimer: The Good Investors is the personal investing blog of two simple guys who are passionate about educating Singaporeans about stock market investing. By using this Site, you specifically agree that none of the information provided constitutes financial, investment, or other professional advice. It is only intended to provide education. Speak with a professional before making important decisions about your money, your professional life, or even your personal life. I don’t have a vested interest in any company mentioned. Holdings are subject to change at any time.

Last month, I published Even More Of The Latest Thoughts From American Technology Companies On AI (2026 Q2). In it, I shared commentary in earnings conference calls for the first quarter of 2026, from the leaders of technology companies that I follow or have a vested interest in, on the topic of AI and how the technology could impact their industry and the business world writ large. 

A few more technology companies I’m watching hosted earnings conference calls for 2026’s second quarter after I prepared the article. The leaders of these companies also had insights on AI that I think would be useful to share. This is an ongoing series. For the older commentary:

• 2023 Q1 – here and here
• 2023 Q2 – here and here
• 2023 Q3 – here and here
• 2023 Q4 – here and here
• 2024 Q1 – here and here
• 2024 Q2 – here and here
• 2024 Q3 – here and here
• 2024 Q4 – here, here, and here
• 2025 Q1 – here and here
• 2025 Q2 – here and here
• 2025 Q3 – here, here, and here
• 2025 Q4 – here, here, here, and here
• 2026 Q1 – here, here, here, and here

With that, here are the latest commentary, in no particular order:

Adobe (NASDAQ: ADBE)

In the Business Professionals and Consumers group, Adobe’s management is attempting to transform Acrobat from a leading PDF tool into an AI-powered document productivity platform; more than 400 billion PDFs are opened with Acrobat annually; management believes integrating AI into Adobe’s products will help customers realise the value of AI faster; Acrobat’s capabilities are integrated into major platforms including ChatGPT, Chrome, Claude, and Whatsapp; management recently announced major advances in Acrobat, powered by the Adobe Productivity Agent, that helps users understand information better and create work easily

Our differentiation is grounded in a deep understanding of the PDF format and three decades of document expertise to transform Acrobat from the world’s leading PDF tool into a comprehensive document productivity platform powered by AI. Today, people open more than 400 billion PDFs with Acrobat every year to read, edit and create content. By integrating AI deeply into our products, we are helping customers realize the value of AI faster, more accurately, securely and cost efficiently. We’re meeting users wherever they work and expanding our reach across the digital ecosystem by extending Acrobat’s capabilities to major platforms including ChatGPT, Chrome, Claude, Microsoft Edge and WhatsApp…

… Yesterday, we announced significant advances in Acrobat, powered by the Adobe Productivity Agent, that help people and teams understand the information in their documents faster, create polished work more easily and access trusted knowledge at scale. Acrobat now turns complex documents into visually rich interactive reports and summary slides, as well as audio formats like personal podcasts, enabling users to quickly absorb key insights. New features like Knowledge Base and Analyzer empower teams to ask questions across massive document collections and surface structured insights from thousands of files. This unlocks the value of existing organizational knowledge, driving smarter, faster decision-making at scale, all with enterprise-grade governance, security, trust and compliance. 

In the Creators & Creative Professionals group, Adobe’s differentiation lies in its deep domain expertise across media types; management is delivering personalised value to users with agentic AI with Firefly and Creative Cloud; Adobe offers intelligent routing to the best models for users’ needs; management has expanded the Adobe Creative Agent to additional Creative Cloud flagship apps; Adobe now has a new AI Assisted Editor in beta mode for users of its generative tools; management recently added native generation of music, speech and sound effects to Adobe Firefly; management recently released new creative skills and tools for social creators and solopreneurs; management recently enhanced Firefly AI Assistant’s personalisation and context-awareness; management recently added Adobe Firefly Graph Enterprise Edition and Firefly Creative Production Enterprise Edition to Firefly Enterprise; management recently added a new Simulate feature in Adobe Brand Intelligence to predict campaign performance before the content goes live; Disney Imagineering will be integrating Firefly Foundry for its theme park design toolkit; management is seeing a lot of usage in video generation in Firefly

Our differentiation is rooted in our deep domain expertise across media types, including imaging, design, video, photography, illustration, animation and 3D and complemented by our depth of intelligence across creative workflows. With Firefly and Creative Cloud, Adobe is reinventing creative software with agentic AI to deliver immediate, personalized value— meeting users where they are, with tools and agents that adapt to their needs. Adobe remains committed to delivering intuitive, context-aware experiences that honor creative craft, accelerate creative velocity and ensure that the next generation of Creators can thrive in an ever-changing world. Adobe’s strategy is to offer customers choice as well as intelligent routing to use the best models for their needs, fully integrated in Adobe applications, eliminating the friction of switching between workflows and platforms.

The Adobe Creative Agent empowers customers by handling the orchestration and execution of complex, repetitive creative workflows. We expanded the Adobe Creative Agent to additional Creative Cloud flagship apps, including Photoshop and Premiere, extending conversational, agentic workflows deeper into our professional applications.  

We also introduced several innovations in Photoshop, giving professionals more choice and control at every stage of the creative process. This includes a new AI Assisted Editor, now in beta, offering customers the option of using a dedicated interface and natural-language prompt bar alongside our generative tools powered by leading AI models. 

Adobe Firefly is the all-in-one creative AI studio for the next generation of Creators. In Q3, we added native generation of music, speech and sound effects, bringing commercially safe AI audio to creative workflows. We released new creative skills and tools like Create Storyboard and Create Brand Kit, purpose built for social creators and solopreneurs, along with additional enhancements that help make Firefly AI Assistant more personalized and context-aware over time…

…Firefly Enterprise, spanning Firefly Services, Adobe Firefly Foundry and Adobe Brand Intelligence, helps the world’s largest brands industrialize content production with brand safe custom models. In July, we expanded Firefly Enterprise with Adobe Firefly Graph Enterprise Edition and Firefly Creative Production Enterprise Edition, giving businesses reusable, nodebased creative workflows and mass-scale production capabilities, alongside a new Simulate feature in Adobe Brand Intelligence that predicts campaign performance before content goes live. We announced that Disney Imagineering is integrating Firefly Foundry into its theme park design toolkit, bringing Disney stories, characters and experiences to life in Parks faster than ever…

…If you look at the amount of innovation that is going into Firefly, it is magical. Video in particular, you are seeing a lot of usage of video, so that certainly helps across that.  

Adobe recently acquired Topaz Labs; Topaz Labs adds state-of-the-art AI enhancement models to the Adobe Firefly, Firefly Services, and Creative Cloud apps; Topaz Labs has over a million users; Topaz Labs’ AI technology has won an Emmy Award; Adobe’s management will integrate Topaz Labs’ technology across Adobe’s creative AI portfolio; management expects the deal to close in 2026 Q3 (FY2026 Q4); some of Adobe’s customers are also Topaz customers and they have been praising Topaz

We announced our agreement to acquire Topaz Labs, which would add state-of-the-art AI enhancement models in Adobe Firefly, Firefly Services and Creative Cloud apps, providing Creators, designers, video professionals, photographers and enterprises the tools to achieve exceptional quality across every format and workflow. With over a million users, Topaz Labs and its Emmy Award-winning AI technology will be integrated across Adobe’s creative AI portfolio, enabling customers to enhance footage, restore and remaster archival content, and blend AI-generated and traditionally captured content into seamless final productions. We expect the transaction to close in Q4, subject to regulatory approvals and other customary closing conditions…

…A number of our customers have been using our technology along with Topaz. We had a lot of great feedback on them from our customers.

Under the Creators & Creative Professionals group AI credit consumption is accelerating sequentially in 2026 Q2 (FY2026 Q3) across Creative Cloud and Firefly; Firefly’s ending ARR in 2026 Q2 (FY2026 Q3) was up 40% sequentially (was nearly $300 million in 2026 Q1); under the Marketing Professionals group, ending ARR was up 20% year-on-year in 2026 Q2 (FY2026 Q3) for each of AEM & agentic web apps, Adobe GenStudio and AEP & apps  

Additional Creators and Creative Professionals highlights include:

• Creative freemium MAU now surpassing 100 million and growing over 70% year over year;
• Intensity of AI usage continues with credit consumption accelerating quarter on quarter, across Creative Cloud and the Firefly App;
• Firefly ending ARR across Firefly App and Firefly credit packs grew 40% quarter over quarter;
• Enterprise wins this quarter include Academy Sports, Disney, Jet2, Perficient, Premier League, Publicis, Tennis Australia and T-Mobile…

…Additional Marketing Professionals highlights included:

• Ending ARR growth of over 20% year over year for each of AEM & agentic web apps, Adobe GenStudio and AEP & apps;…
• …Q3 industry analyst recognition, being named the top leader in two Forrester Waves: including Customer Data Platforms for B2C and Experience Optimization Solutions, as well as two IDC Marketscapes for Worldwide AI-Enabled Customer Data Platforms for B2B Users and B2C Users; 
• Global enterprise customer wins in Q3 included Academy Sports, Alpine Racing, BNP Paribas, Humana, IKEA, Jet2, Marriott, MSC Cruises, Publicis, Royal Bank of Canada, Vanguard and Wells Fargo. 

Adobe’s management’s recently delivered the first unified brand visibility solution, Adobe Brand Visibility, by combining Semrush’s AI visibility intelligence with Adobe’s agentic content optimization capabilities; Adobe Brand Visibility can tap on Adobe’s database of nearly 300 million real-world AI search prompts; Adobe Brand Visibility is seeing tremendous interest from brands to understand how they are showing up across leading AI platforms; management recently introduced GenStudio for Commerce Media Networks for retailers to provide self-serve, on-brand creative for advertising partners; management recently expanded Adobe Brand Intelligence so users can predict audience reactions to content before it goes live; CX Enterprise Coworker was recently made generally available; CX Enterprise Coworker coordinates AI agents and workflows across analytics, content creation and journey orchestration; CX Enterprise Coworker already has 1,700 customers and early adopters

In Q3, we delivered the first unified brand visibility solution bringing together Semrush’s AI visibility intelligence with Adobe’s agentic content optimization capabilities, grounded in a deep understanding of business and brand context. Adobe Brand Visibility is a comprehensive solution for Generative Engine Optimization that combines the industry-leading capabilities of Adobe LLM Optimizer with Semrush’s AI Optimization. We’re seeing tremendous customer interest in tapping our database of nearly 300 million real-world AI search prompts to understand how their brands are showing up across leading platforms such as ChatGPT, Google AI Mode, Microsoft Copilot and Perplexity AI. 

At the Cannes Lions International Festival of Creativity in June, we introduced a major expansion of Adobe GenStudio to help brands meet customers wherever they are in the moment of purchase, in a stream or in a scroll. GenStudio for Commerce Media Networks gives retailers a way to attract advertising partners with self-serve, on-brand creative built directly from product listings and category context. We also expanded Adobe Brand Intelligence capabilities so marketing teams can preview how content will resonate with audiences before it ever goes live. This is what an agentic content supply chain demands, at the speed our customers need. 

We are seeing tremendous customer interest in CX Enterprise Coworker, the specialized AI agent in CX Enterprise that autonomously executes complex marketing and customer engagement workflows. CX Enterprise Coworker, which became generally available in June, meets marketers where they are, embedding agentic intelligence directly into the engagement lifecycle, offering choice and control while enabling teams to act faster, scale personalization, and continuously optimize outcomes. CX Enterprise Coworker synthesizes insights from Adobe and third-party applications, while coordinating AI agents and workflows across analytics, content creation and journey orchestration. Early customer interest is strong with over 1,700 customers and early adopters of CX Enterprise Coworker. 

Adobe’s management is seeing customers look for 4 things in agentic software, namely, (1) a user interface of their choice, (2) model flexibility, (3) connection to their enterprise data so that there’s no hallucination, and (4) the ability to realise the value of the software

What I hear is when they look for agentic software, what they mean by that is they are looking for, one, the product architecture, the product architecture evolving so that people can use the user interface of their choice, whether it is conversational or a traditional interface, that could be ChatGPT or a cloud or a Copilot, along with the functionality they get from Adobe. We are making that possible, as you have seen our announcements across creativity, across productivity, as well as the CX announcements like the Adobe Marketing Agent. We are making that possible through other interfaces as well as our own, like we are doing with our apps as well.

Second, they are looking for model flexibility. They’re looking to make sure that they can have access to the best models available and that we, as the provider of the agentic software, are matching the best model available to the best task and making it easy for users. That’s something that we are doing extremely well, with the Adobe Creative Agent we talked about and the Adobe Productivity Agent as well. 

Then they’re looking to make sure that the apps are using their enterprise data and context so that there’s no hallucination, that they get usable results, and that they get predictability out of the use of AI. That’s something that we are doing really well with, for example, the CX Enterprise Coworker we talked about, because that’s built on top of the Adobe Experience Platform, where we serve over 1 trillion experiences every year.

Then they’re looking for value. They’re looking to make sure that the provider of agentic software is helping them realize the full value of this software, bringing it all together across all of the complexity of AI and software. That’s what we’re doing with the growth we’re seeing across our forward -deployed engineering and our value -realization teams.

Adobe’s management sees a wide variety of pricing models for its AI products

[Question] One question we get a lot is you kind of meet customers where they are or how they want to consume in terms of how you monetize AI across either Creative Cloud, Acrobat, or your enterprise products, and that is through bundled credits, credit packs, premium tiers, per-seat pricing. There is also usage-based contracts. Over time, which pricing model do you expect to become the primary driver of AI revenue?

[Answer] It really depends on what exactly they are doing, what are they trying to get done, what is their skill level, what they are doing it in the context of. So that can vary based on whether they are individuals or enterprise customers and so on. So it really then starts to focus on what they are actually doing with it. So we are happy to work with, say, for example, something like Microsoft Copilot in the enterprise to ChatGPT for consumers and so on, and making sure that they are getting value and driving up that intensity of usage and the engagement. Beyond that, depending on the product area, once we have the value, we can map it to the monetization that Shantanu talked about, which helps us with better segmentation, understanding what the segments are, and how they are getting value. So we have a wide variety of ways, as you pointed out, to monetize that, and we are going to be applying that both based on the context of the customer and the context of what they are doing.

MongoDB (NASDAQ: MDB)

Enterprise Advanced (EA) revenue growth was 36% year-on-year in 2026 Q2 (FY2027 Q2); the strength in EA was widespread across MongoDB’s installed base; when management brought search and vector search to EA, demand for EA came in immediately from customers who want to build AI in their own self-managed environments; a US bank was an existing customer of EA and recently started using EA for GenAI and semantic search; AI self-managed use cases represent net new demand for MongoDB, while hybrid development opens the door to Atlas for existing EA customers; the introduction of AI features in EA was driven by customer requests

EA and other had a standout quarter, growing 36% year-over-year due to widespread strength driven by our run anywhere capabilities…

…Turning to Enterprise Advanced, this quarter’s strength was widespread across our install base, particularly within financial services, tech, and the public sector. Two patterns in how customers are using EA stand out, and both point to why this business is strategic for us.

The first is AI in governed, self-managed environments. This quarter, we brought search and vector search to EA, closing a gap between our cloud and self-managed experiences. Demand came in immediately and across industries from customers looking to take a consolidated approach to building AI in their own governed, self-managed environments. A major U.S. bank shows what that looks like in practice. EA already serves as the standardized data platform for more than 100 production applications across payments, fraud detections, document processing, customer and account services. This quarter, that bank extended that same environment to GenAI and semantic search for employee advisors, chatbots, product search, and document intelligence. By bringing operational data, search, and vector retrieval together, self-managed with EA, they keep sensitive customer and conversational data inside their own governed environment without sending up separate systems. That gives them a practical foundation to expand AI across the bank on the same platform already running their most critical operations…

…Bringing AI self-managed opens net new demand for us, and hybrid deployment often means that the strong EA estate opens the door to net new Atlas conversations within the same customers…

…The reason we invested in EA roadmap that we outlined, and it is nice to see it is working out, is that customers said to us, many customers, even in my early days, that you must invest in EA. If EA gets to being AI-ready with search, vector search and so on, they are asking, “Hey, can we also make Voyage AI available in a self-managed type of an environment?” That was very customer-driven, and we are meeting customers where they are.

Voyage’s customer count was up nearly 100% sequentially in 2026 Q2 (FY2027 Q2); Atlas Vector Search adoption  is outpacing growth in the rest of MongoDB; Atlas Vector Search’s performance is a sign of strong early momentum in AI workloads; financial services, healthcare, technology, and AI native companies are increasingly choosing MongoDB to run AI workloads; Voyage’s embedding and reranking models are consistently at the top of leaderboards; management recently brought automated Voyage embeddings to Atlas; management recently launched Voyage Code 4, a coding model; management recently shipped a new reranking API for Voyage; some of Atlas’s largest existing customers are starting to use Voyage for AI use cases, while a large number of new Voyage customers are AI natives that are new to MongoDB; Voyage grew its customers by around 100% sequentially for the 2nd consecutive quarter in 2026 Q2 (FY2027 Q2); Claude Code and Codex are driving developer referrals to Voyage; coding agents love Voyage; there is low awareness that Voyage is part of MongoDB, but once customers realise this, they start to also look at Atlas

Voyage customer count nearly doubled quarter-over-quarter, and Atlas Vector Search adoption continues to outpace the growth of the rest of the company, showing our strong early momentum for AI workloads…

…Enterprises across financial services, healthcare, tech, are running their most demanding mission-critical workloads on MongoDB, and we are winning more workloads each quarter. Increasingly, these same enterprises, as well as AI natives, are choosing our platform for AI workloads, evidenced by the adoption of Atlas Vector Search and Voyage AI embeddings…

…We are also seeing strong traction with Voyage, our embedding and reranking models, which consistently rank at the top of independent leaderboards. In August, we brought automated Voyage embeddings to Atlas for one-click vector search setup, launched Voyage Code 4, a model purpose-built for code, and shipped an upgraded reranking API, all keeping Atlas retrieval accuracy for AI ahead of the market. 

Voyage traction is showing up on both ends of the market. Some of our largest existing Atlas customers are beginning to adopt Voyage for AI use cases, while a large majority of new Voyage customers are AI natives and have no prior relationship to MongoDB…

…Within Atlas, Voyage customers roughly doubled quarter- over- quarter for the second consecutive quarter, continuing the encouraging signs of the demand for our AI embedding capabilities…

…When I look at the names of the kind of customers we are getting, whether they are in San Francisco Bay Area, whether they are large enterprise, whether they are in London or Tel Aviv, or Seattle, they tend to be driven by AI workloads. When the team did analysis on where is the referral for our Voyage is coming, as you would have imagined, most of this referral is coming via coding agents. Number one, Claude, and number two, Codex is driving most of the referral traffic for Voyage…

…Coding agents love Voyage. They are recommending us, and we are getting this new customer cohort…

…The awareness is low, that Voyage is actually coming from MongoDB. This customer told me, “Oh, we love Voyage. We are using Voyage.” I said, “You know that is a MongoDB product.” They are like, “Oh, we did not know that.” Okay, then we should now look at Atlas because you have Atlas auto-embeddings.

MongoDB’s management thinks it’s natural for customers whose data are already on MongoDB to build agentic workflows on top of the data; MongoDB’s platform have important features for AI, such as search, vector search, and embeddings, that are built in; management is seeing different industries utilise MongoDB for a range of AI use cases; management is seeing a growing number of AI workloads reach production; MongoDB is starting to see some benefit from AI but it’s still small; the AI workloads management is seeing on MongoDB are mostly for customer-facing workloads; one of MongoDB’s bank customers told management that they think vectors should be integrated into the data layer, which is to MongoDB’s advantage; a media company which is a big vector search customer, had its agents realise that semantic queries need to be within an operational data layer; Eleven Labs, an AI native that continues to scale well with MongoDB, thinks it’s important that vector is being embedded in a database

For customers that already run large part of their data estate on MongoDB, building an agent on top of that data is a natural extension because the data an agent actually needs is live operational data, not a stale copy sitting in a warehouse. Search, vector search, and embeddings are built in, not bolted on, so rather than agents connecting to many separate systems, they connect to one platform. We are seeing this show up across industries in a range of use cases, whether it is retrieval of internal knowledge, customer-facing chatbots and agents, or fraud and identity workflows. It is still early, but we are seeing more of these workloads reach production, such as the Financial Times…

…We have started to see some benefit from AI, even though it is small, but we are excited about the momentum, and we do expect consumption to continue to be consistent with what we have seen during the first half of the year…

… What I’m seeing is initially, say you are a wealth manager at a bank and there are lots and lots of knowledge base articles that you want to vectorize, use our embeddings, and then use as a chatbot for folks that are doing wealth management and talking to clients real-time. That is one very specific example where a particular large bank is using MongoDB. There are also other examples where because there are lots and lots of documents, employee-facing use cases where knowledge base articles so that employees can leverage, do a search, because now search is fully integrated into the operational data and documents get loaded, and then embeddings make the vectorization better. That will be another large enterprise example where we are seeing use cases. But the clarity that I got was that it was almost always, “Hey, we want to use MongoDB where the scale matters on the agents that we are trying to create for our customer-facing activities,” whatever the customer-facing activities are. We are not seeing early traction with, “Hey, I created a co-pilot kind of a thing that appeals to a couple of hundred employees.”…

…These are millions of agents in production that are doing something that is customer-facing, and they would say, “We want to use MongoDB for scale, performance, and of course, run anywhere,” and that’s where you’re using it…

…This large bank told me that based on their testing, they believe that vectors should be integrated fully in the operational data layer, and MongoDB doing that was seen as a huge advantage…

…Even a large media company, which became one of our biggest vector search customer, that was driven by an agent trying to do the semantic query and figuring it out. Okay, if this is an operational data layer, then it just works…

…We are seeing that even in AI native cohort, that vector being part of the database is received really, really well. One of the examples I shared, last quarter, ElevenLabs, which continues to scale nicely with MongoDB, they see that as a huge advantage of vector being embedded.

The Financial Times is using MongoDB for AI-driven discovery; the Financial Times is using Vector Search and Voyage AIto build a hybrid full text and semantic search solution; the Financial Times has used the Voyage-4 and Voyage-4-lite models to significantly cut retrieval costs with minimal performance impact; the time-to-value for the Financial Times in using MongoDB is measured in weeks

It is still early, but we are seeing more of these workloads reach production, such as the Financial Times, which leverages us to power AI-driven discovery, reaching millions of readers with interactive experiences at scale. With Vector Search and Voyage AI, the Financial Times now unifies their operational data and vector embeddings on a single platform, building a hybrid full text and semantic search solution, eliminating the complexity of syncing separate systems and accelerating time to production. By indexing content with the high accuracy Voyage-4 model and serving over 100,000 daily queries on the cost-efficient Voyage-4-lite model, the Financial Times has significantly cut retrieval costs with minimal performance impact. What used to take weeks of manual index monitoring is now finished in a day…

…Like what we saw on my Financial Times use case, that I shared, the time to value for them to leverage Atlas Vector Search and embedding was in weeks, not in months and years, to get AI ready for searches and others that happen.

Frontier AI labs are customers and partners of MongoDB; multiple frontier labs are using Atlas for mission-critical workloads; there’s one frontier lab using MongoDB for inference and chat workloads after poor performance from PostgreSQL, and experiencing 10x faster reads; frontier labs are using MongoDB for research workloads for model development; MongoDB’s relationships with frontier labs are still early, but management is optimistic about the traction; MongoDB’s management recently launched a fully managed MCP (model context protocol) server for developers to connect to MongoDB when using agentic coding platforms; Anthropic’s management have been pointing developers to MongoDB Voyage for embeddings; the frontier lab using MongoDB for inference started with inference, and then expanded usage to other workloads; the frontier lab using MongoDB for inference is having a great experience with Atlas

Moving on to the momentum we are seeing with Frontier Labs, who are both customers and partners for us. Multiple leading labs leverage Atlas for workloads that are mission-critical to how they ship their products. One lab uses us for inference and chat workloads after moving away from PostgreSQL due to performance lags and outages affecting user experience. They migrated their chat memory system onto Atlas in just four weeks and now run at 10x faster reads than PostgreSQL. Beyond that, labs use us for research workloads to store experimental results, evaluation data, and training artifacts for model development. These relationships are still early, and engagement varies lab by lab, but we are energized by the traction we are seeing with them…

…Just recently, we launched a fully managed MCP server, making it easier for developers and agents to connect directly to MongoDB when they are using Claude Code, Codex, and Grok Build, as well as popular coding tools like Cursor and Devin from Cognition. Paul Smith, Chief Commercial Officer at Anthropic, described our technology partnership and recent integration with Claude by noting, “The best AI applications need a strong database, which is why we have long pointed to developers building on Claude to MongoDB Voyage for embeddings. More recently, demand from those developers drove MongoDB to build a new managed MCP server, which has seen fast adoption since launch and now lets developers explore, query, and manage their MongoDB data without ever leaving Claude.”…

…With one of the labs, they started towards the later half of last calendar year with one of the workloads that was running inference on MongoDB and used us as a memory layer. With that lab, our team, what they saw on the Atlas performance for that specific inference, then they said, “Wow, Atlas is performing really well across reads and writes compared to PostgreSQL.” That is what they were using originally. They started then moving just recently in Q2, few other workloads for inference on Atlas. So we had one inference workload that started last year in November, December timeframe, and then they moved another couple of workloads for inference for some other products that they have created in, I want to say this is August, so around June, July timeframe. We are seeing. They told me straight up, this is the technology team, that Atlas has taken all the pain away from an uptime perspective, performance perspective. We do not even think about it. We are now as we create new products, we want to run inference on it.

MongoDB’s management thinks the choice of the data layer made by AI natives will determine whether their product can rapidly scale; some AI natives choose MongoDB from day one, while others migrate to MongoDB after hitting scaling limits when using their prompt-driven development platforms; many of MongoDB’s new customer additions in 2026 Q2 (FY2027 Q2) are AI natives

The final piece of the AI opportunity is AI natives, companies whose data layer determines whether the product can support rapid scale. Some choose us from day one. Others start elsewhere, like prompt-driven development platforms, and migrate to us as they hit scaling limits and real usage arrives…

…We added a record 2,900 net new customers this quarter, and many of them are AI natives.

Fireflies, an AI native building an AI assistant for work, serves more than 20 million users; Fireflies chose MongoDB from day one for its flexible document model, which has provided the foundation for Fireflies’ hypergrowth

Fireflies, a unicorn AI native startup, is building what it calls the number one AI assistant for work, helping people unlock the knowledge buried in their conversations. Fireflies serves more than 20 million users across 1 million+ organizations and has processed over 7 billion meeting minutes. Fireflies chose Atlas from day one for its flexible document model over a rigid relational schema, and today runs more than 40 microservices with change streams powering real-time pipelines for analytics and growth intelligence. That lean, scalable foundation has helped fuel their hypergrowth seamlessly.

Eve, an AI native providing AI solutions for legal work, is using Atlas embedding and Voyage’s reranking API to improve retrieval quality directly in Eve’s RAG (retrieval-augmented generation) layer 

Eve is one of them. A unicorn AI native that automates legal case intake, medical chronologies, and demand letter drafting for plaintiff law firms. Eve uses Atlas embedding and reranking API powered by Voyage AI’s Rerank 2.5 to surface the most relevant evidence from large sets of case documents. This improves retrieval quality directly into Eve’s RAG layer while simplifying the infrastructure needed to build and evolve these AI experiences.

Oracle (NYSE: ORCL)

Oracle had very strong year-on-year revenue growth of 121% for its Cloud Infrastructure business in 2026 Q2 (FY2027 Q1), driven by a strong demand environment

Cloud infrastructure revenue for Q1 was $7.4 billion, up 121%, reflecting strong execution as we brought record levels of new megawatt capacity online, supported by a continued strong demand environment for compute and our database services.

Oracle’s gross margin declined in 2026 Q2 (FY2027 Q1) as expected as it builds out its AI infrastructure business; the buildout has caused Oracle’s free cash flow to be negative; management continues to expect Oracle’s capex to around $70 billion in FY2027; Oracle recently completed a $20 billion at-the-market equity issuance in 2026 Q2 (FY2027 Q1) to help fund its capex; management has previously said that FY2027 and FY2028 are peak capex years for Oracle;  Oracle’s AI infrastructure projects are delivering a free cash flow conversion ratio of 100% to post-tax EBITDA very shortly after they ramp up; management is seeing costs for data centers going up, and so Oracle has to charge more, so there won’t be an impact on gross margins; management thinks operating margin is actually more important for Oracle than gross margin; management expects Oracle’s gross margin to step down in FY2027 and then flatten over the next couple of years

Our gross margin did decline as expected, driven by impacts from ramping up our data centers and the acceleration of infrastructure revenue. However, this was offset in the quarter by lower operating costs and strong operating leverage tied to simplification and efficiency actions…

…Our CapEx for the quarter was $28 billion, leading to negative free cash flow of $5 billion. Our net cash CapEx, so net of prepayments, was $18 billion for the quarter. To note, our CapEx will not be linear throughout the year. We continue to anticipate $90 billion-$95 billion in CapEx for the full year, with not more than $70 billion in net cash CapEx. Lastly, we are quite pleased to announce that we completed our previously disclosed $20 billion at-the-market equity issuance in entirety during the Q1…

…[Question] You have told us that fiscal 2027 and 2028 are peak CapEx years. At the same time, others in the market are spending hundreds of billions of dollars on capacity with seemingly no end in sight. How should we think about Oracle possibly slowing down spending beyond the next two years if others aren’t?…

…What I would say, though, is that each of these projects that we are doing, by nature, is a strong free cash flow generating project. As soon as they ramp up, very shortly thereafter, they are delivering a free cash flow conversion ratio of something like 100% to post-tax EBITDA. In fact, the business by nature is somewhat, quote, “self-funding” at some point, in terms of throwing off a lot of free cash flow…

…Prices in a world where demand exceeds supply, typically prices don’t go down, they do go up. I think the net effect is that obviously things cost more, but then we have to charge more money for them so that we get compensated. We’re doing that across all of these different businesses. We don’t expect this to have an impact on our gross margins…

…Gross margin to me is always an important indicator, probably even more important internally for us to double-check, and I think investors obviously want to double-check. It is something that will change very quickly, for example, if we do not have the right pricing model. When we talk about driving value, though, and driving value for the business, for me, operating margin is probably the ultimate point that we want to follow. Gross margin, like you mentioned, at the moment, there is a number of things going on. We have both the ramp-up in data centers, plus we have an adjustment across the two business models that we have in the business. Software being a much higher gross margin business, but with higher R&D and sales costs below gross margin. Infrastructure being a lower gross margin business, and we have talked about that and we gave the numbers. Clay has given the expectations for quite a bit of that business. Not database, obviously, but the more AI infrastructure and cloud side. That business, by nature, has much lower R&D and sales associated with it, at least in a company like Oracle, where we can effectively gain from all of the R&D that is already been done and being done across the rest of the company. So really, how to watch how we are going to drive value out of the business over time, I think operating margin is really the key metric that we would look at…

…I mentioned in the Q4 that we would expect a step down in gross margins this year. You can see the EPS guidance that we give, though, so you can see what we might expect in terms of operating margin. Over the next couple of years, as we finish the ramp-up, you can reasonably expect that gross margin would flatten, I would say.

Oracle’s remaining performance obligation (RPO) in 2026 Q2 (FY2027 Q1) was up 46% year-on-year to $664 billion (was $638 billion in 2026 Q1); the vast majority of the sequential increases in RPO were from prepay or bring-your-own-hardware contracts, which do not require incremental capital from Oracle, although it still requires capex; the new S$26 billion in RPO will not impact Oracle’s capital expenditure or revenue until FY2028 and beyond; management expects half of Oracle’s RPO to convert to revenue in the next 36 months

our Remaining Performance Obligations, or RPO, increased $26 billion from Q4. There are two things happening here. First, we continued to grow our RPO during the quarter to support future revenues. The vast majority of those new contracts were via prepay or bring your own hardware or similar mechanic, so will not require incremental capital from Oracle. Also, that new RPO will not impact our CapEx or revenues until fiscal 2028 or beyond. Second, we started to see a strong conversion of our RPO into revenues this quarter, driving our cloud infrastructure results…

…I didn’t say, and I don’t think myself nor Hilary said that it doesn’t require additional CapEx. We said it doesn’t require additional cash from Oracle…

…While there clearly are capital expenditures, it does not require Oracle to go out and find additional cash to do it. Now, the question becomes, well, how do you do that? Well, we have a variety of different models. Sometimes it’s working with our suppliers, through different financing arrangements that allows us to pay for the capacity as the customers pay us. That’s one mechanism. Another mechanism is that a customer says, “Hi, I’d like to pay for the hardware, but use your operational ability and your cloud infrastructure technology assets and your data center to go out and actually turn that into an AI cluster.” It’s a different option. A third option is that the customer has been able to raise money. Maybe it is a startup, maybe it is an established company, and says, “Hi, I would like to pay you upfront as a prepayment, and in return, that doesn’t require you to front the cash to go out and spend your dollars on that CapEx.”…

…We now expect around half of our RPO to convert into sales over the next 36 months…

…We closed more than $30 billion of additional AI contracts in Q1 without requiring additional capital from Oracle.

Oracle’s management thinks that AI is an accelerator for packaged applications; management thinks AI agents can perform tasks using an organisation’s established workflows and business rules, and human employees just have to oversee the agents; management thinks Oracle can combine AI with business rules, regulatory compliance, security models, and data modesto enable customers to realise AI’s value; management is confident that the introduction of AI into Oracle’s product suite will deliver significantly faster ROI (return on investment) for customers; management will soon introduce an agentic AI feature that can dramatically reduce SaaS deployment times; management thinks Oracle AI Agent Studio has a very compelling value proposition because customers get to build AI agents on very complex business rules, a highly differentiated security model, and data models 

The introduction of AI is an accelerator, not a replacement, for packaged applications. As such, our decades of experience and expertise running business processes across every industry, in every geography, for organizations of any size, gives us the understanding of how to help them succeed.

Before AI came along, application suites had already proven their effectiveness. Companies had been able to increase their profit margins because end-to-end automation with standardized and efficient business processes proved to be much more effective than one-off custom solutions. But that did require organizations to follow workflows and processes as designed in the system, something that many struggle to achieve consistently across functions, teams, and regions. AI changes this dynamic. Rather than asking every employee to navigate and execute a process exactly as the system expects, AI agents can perform tasks using the organization’s established workflows and business rules. Employees then shift to overseeing agents, resolving exceptions, and applying human judgment where it matters most. By combining applied AI with decades of sophisticated business rules, regulatory compliance, security models, data models, and customer configurations, we enable customers to continuously realize AI’s value while keeping their data secure and their operational guardrails intact…

…We are incredibly confident in the potential for this new paradigm to deliver much more rapid ROI for our customers. At AI World in October, we will unveil a new agentic AI accelerator poised to redefine how customers deploy Oracle applications faster, simpler, and at a dramatically lower cost. Working alongside Oracle and customer teams, AI agents will automate and orchestrate implementation at an unprecedented scale, compressing SaaS deployments from years to months, and months to weeks…

…The next layer is our Oracle Fusion Agentic Applications AI studio, which allows customers and/or partners to build their own AI agents right inside the same platform. That is not a different platform. It is not a different control plane. It is the same control plane and the same platform that our applications are running on, which means that Oracle AI Agent Studio, which allows customers to build their own agents or partners, gets all of the same quarterly updates, gets all of the same security patching, and is available as a complete service to our customers. You are allowing customers to position AI as a UI on top of a very complex set of business rules, on top of a highly differentiated security model, and of course, data models that have evolved for years and years. You put the horizontal applications, the vertical applications, the Oracle AI Agent Studio together, and we think that is very compelling.

Oracle’s customers used embedded AI capabilities more than 150 million times in 2026 Q2 (FY2027 Q1), up 42% sequentially; Oracle’s AI agents executed 3.5 million times in 2026 Q2 (FY2027 Q1), up nearly 100% sequentially; Oracle’s customers have more than 2,300 AI agents in production, up 90% sequentially; AI production usage in Oracle Fusion consumed 900 billion tokens in 2026 Q2 (FY2027 Q1)

Customers used our embedded AI capabilities more than 150 million times during the quarter, with usage growing 42% sequentially. Our AI agents executed more than 3.5 million times in production during the quarter, nearly doubling quarter- over- quarter. Customers have over 2,300 AI agents in production, and that is up 90% quarter- over- quarter. Overall, AI production usage across Fusion alone consumed 900 billion tokens during the quarter. 

Oracle’s management recently announced the general availability of the AI-powered NetSuite Next; more than 10,000 NetSuite customers are already using the NetSuite AI Connector service that lets customers connect NetSuite data to AI assistants; Every Man Jack estimates that the NetSuite AI Connector service will save it $350,000 annually

We are announcing the general availability of our new AI-powered offering called NetSuite Next. This presents an agentic experience that is simpler, more powerful, and is infused with AI across the workflows that customers rely on every day. It is easier to adopt, it is more productive from day one, and it is more valuable as customers grow. Additionally, the NetSuite AI Connector Service, which lets customers securely connect their NetSuite data to leading AI assistants of their choice, including ChatGPT and Claude, is already one of the fastest adopted capabilities in the whole entire history of NetSuite, with more than 10,000 customers already using it. Personal care company Every Man Jack estimates that the service alone will save $350,000 annually and nearly 5,000 hours of work.

Oracle delivered 850 megawatts of AI infrastructure in 2026 Q2 (FY2027 Q1); delivery in FY2027 Q1 is 3x what was delivered in FY2026 Q4, and 73% of the capacity delivered in FY2026; Oracle’s global GPU utilisation is 97.9% (was 97.5% in 2026 Q1); Oracle’s GPUs that came up for renewal in 2026 Q2 (FY2027 Q1), most of which are at least 4 years old, was renewed or resold at prices 20% higher; management sees long useful lives for the AI infrastructure Oracle is building; Oracle’s Abilene, Texas AI data centre has delivered 75% of its total capacity (was 42% in 2026 Q1); customer-acceptance at Abilene has shortened to only 24 hours; OpenAI’s latest GPT-6 Astra model was trained in Abilene; Oracle’s Shackelford, Texas AI data center is progressing well; management finds NVIDIA’s Vera Rubin systems to be performing better than expected, and will deliver them to customers in 2026 Q3 (FY2027 Q2); Oracle’s AI infrastructure investments are pretty diversified; when Oracle builds a data center, the capacity is not all delivered at one go; Oracle’s new data centers in New Mexico and Wisconsin are making good progress; Oracle is deploying Bloom Energy’s fuel cells in New Mexico and management sees it as the most environmentally friendly on-site energy generation technology; management sees the current constraints in AI infrastructure as power and data centers

We delivered 850 megawatts of AI capacity containing more than 300,000 GPUs to customers since the end of Q4. Delivery in Q1 is almost three times what we delivered in all of Q4 and 73% of the total capacity we delivered last fiscal year…

…GPU utilization remains extremely high at 97.9% in Q1. GPU longevity and value continue to impress. Of all the GPUs that came up for renewal in Q1, that capacity was renewed or resold at a 20% premium to prior contracts. The majority of those GPUs are four years or older. We see a long, useful life with increasing value for the AI capacity we’re deploying.

Abilene continues to deliver at an extraordinary pace. We delivered 131,000 GPUs there in Q1, 1.9 times the volume delivered in Q4. Six of the eight campus buildings, representing 618 megawatts and 75% of total capacity, have now been delivered to the customer. Customer acceptance has compressed to only 24 hours, showing that the systems arrive ready for customer workloads. The recently released GPT-6 Astra was trained at our site in Abilene. Shackelford is our next gigawatt-scale campus and is progressing well.

NVIDIA Vera Rubin systems are performing better than expected across hardware quality, manufacturing yield, and performance. We will deliver our first Vera Rubin systems to customers in Q2…

…New Mexico and Wisconsin are very important large sites for us, but I think it is important to have some context. We talk about these sites as being around 1 gigawatt a piece. We just delivered 850 megawatts in Q1. What that means is that, as you can see, neither Shackelford, nor New Mexico, or Wisconsin, or Michigan were delivered in Q1. So we have a large, diverse, broad set of data center developments going on throughout the U.S. and around the world to deliver capacity to customers. Now, some of these sites, like New Mexico and Wisconsin, obviously garner a lot of attention. There is a lot of discussion about them. But I think it’s important people realize that all of our eggs are not in a single basket…

…When these large sites are built, they don’t all come online at once. Let’s say that you have a gigawatt site, and it’s supposed to start delivering, let’s say, in January of a year. It’s not like in January you get a gigawatt of capacity. It’s phased over many quarters…

…New Mexico is an interesting location. We’re making very good progress. In terms of construction, data center is definitely on track. We’re going through the process of acquiring our air permit. And the technology that we’ll be deploying there is Bloom fuel cells, which is by far the most environmentally friendly way that we can do on-site power generation. Has extremely low water consumption, has extremely low emissions compared to really any other way to do on-site generation. We’re very confident that as we continue through this process, we’ll work with the local regulators and community citizens in Doña Ana County, and with everybody else in New Mexico. But we’re just going through the process, and I don’t think that’s rare for large projects like this one.

In Wisconsin, we’re not doing on-site generation. We’re really working with our partners across the board to design and deliver that energy capability through the grid. But again, working through the process, these are complex projects. And again, in Wisconsin, data center delivery is actually very much on track and going well. Working with the Public Service Commission and ATC and We Energies, we’re constantly evolving different aspects of the energy design and delivery plan…

…The environment continually changes. It used to be that the constraints were GPUs and fabs. Then constraints moved to power generation. There’s data center constraints. But the world’s a big place…

…

Oracle has completed its Azure and AWS regional footprint expansions and this gives customers a consistent way to run Oracle AI Database next to their applications in the cloud they choose

We completed our planned Azure and AWS regional footprint expansion, reaching 70 multicloud database regions and 119 availability zones. This gives customers a consistent way to run Oracle AI Database next to their applications and data in the cloud they choose.

Oracle recently expanded its relationship with OpenAI and now offers OpenAI API access; Oracle is now bringing Gemini models to its enterprise applications; Oracle has released new Grok models; management is expanding Oracle’’s opens source model catalog

We expanded our OpenAI relationship to offer OpenAI API access, ChatGPT for work, and Codex through Oracle Marketplace, including GPT-6 Astra. We are bringing Gemini models to Oracle’s enterprise applications, and we released new Grok reasoning, multimodal, and text-to-speech models. We also continue to expand the open source model catalog, including new models from NVIDIA, Qwen, Google, DeepSeek, and others.

With APEXlang, developers can tap on the speed of generative AI for coding, but remove the downsides of difficulty in maintenance; Oracle’s AI Data Platform is now integrated with Codex and Claude Code; Oracle’s AI Data Platform has a few different business models, which includes consuming non-Oracle work

APEXlang is a new technology that represents an APEX application as structured, human-readable application definitions that can be stored in source control, validated, and governed. AI coding agents generate and modify those definitions while the APEX engine continues to provide the security, reliability, and operational controls required for enterprise applications. Developers gain the speed of generative development without the downsides of difficult-to-maintain opaque application code. We are taking the same approach with the Oracle AI Data Platform. AI Data Platform is now integrated with Codex and Claude Code, allowing developers to work with AI Data Platform data, knowledge, and capabilities from the coding environments they already prefer…

…[Question] You mentioned the AI data platform, you guys put it in your press release, so I just want to double-click on that. It seems like an important way in allowing customers to deploy agents against their proprietary data. Can you just help us understand the business model around that? Does it drive incremental consumption of Oracle Database OCI, or do you see that being sort of a standalone software revenue opportunity?

[Answer] You sort of clicked on the answer, really, to the business model is all the above. Certainly, we can run this in a model where we are consuming 100% of non-Oracle work. This is by no means specific to the Oracle Database or the Oracle applications. The AI Data Platform is agnostic and able to pull in and automate the ontologies from any data source. In fact, we have hundreds of data sources that we are doing this automation for today. So, whether it is just pure consumption of AI Data Platform, whether it is used in concert with our applications, or whether it is used in concert with OCI, we are more focused on allowing the customer to make the best choice, or the partner to make the best choice that suits them.

Oracle is already investing in forward-deployed engineers at its customers for a number of its AI products

We are already investing in deploying forward deployed engineers at our customers. That’s true for the AI Data Platform. It’s also true for our Fusion Agentic Studio, as we see them really as a combination platform running on a single control plane in OCI.

---

Disclaimer: The Good Investors is the personal investing blog of two simple guys who are passionate about educating Singaporeans about stock market investing. By using this Site, you specifically agree that none of the information provided constitutes financial, investment, or other professional advice. It is only intended to provide education. Speak with a professional before making important decisions about your money, your professional life, or even your personal life. I have a vested interest in Adobe and MongoDB. Holdings are subject to change at any time.

Earlier this month, I published More Of The Latest Thoughts From American Technology Companies On AI (2026 Q2). In it, I shared commentary in earnings conference calls for the second quarter of 2026, from the leaders of technology companies that I follow or have a vested interest in, on the topic of AI and how the technology could impact their industry and the business world writ large. 

A few more technology companies I’m watching hosted earnings conference calls for 2026’s second quarter after I prepared the article. The leaders of these companies also had insights on AI that I think would be useful to share. This is an ongoing series. For the older commentary:

• 2023 Q1 – here and here
• 2023 Q2 – here and here
• 2023 Q3 – here and here
• 2023 Q4 – here and here
• 2024 Q1 – here and here
• 2024 Q2 – here and here
• 2024 Q3 – here and here
• 2024 Q4 – here, here, and here
• 2025 Q1 – here and here
• 2025 Q2 – here and here
• 2025 Q3 – here, here, and here
• 2025 Q4 – here, here, here, and here
• 2026 Q1 – here, here, here, and here

With that, here are the latest commentary, in no particular order:

Adyen N.V. (OTC: ADYEY)

The complexities faced by merchants are increasing because of AI and fraud

As the industry shifts, so does the complexity of our customers’ operations and the scope of what we solve. Today, merchants are competing for consumer attention while shoppers expect increasingly personalized experiences and have access to instant product and price comparisons. AI is reshaping enterprise operations, exposing the limitations of legacy systems. Fraud is more sophisticated and requires businesses to protect revenue without compromising the customer experience. It is becoming clear that modern businesses require solutions not only at the transaction, but also well before and after it.

Adyen’s management recently introduced Adyen Agentic, which allows merchants to participate in the AI commerce ecosystem; Adyen Agentic has the Agentic Feed module that allows AI agents to have accurate inventory data; Adyen Agentic has the Agentic Cart module that brings an AI agent’s selection into existing checkout, tax, fulfillment, and order management systems; Adyen Agentic is AI platform-agnostic; major payment companies American Express, Visa, and Mastercard are early strategic partners of Adyen Agentic; Adyen Agentic is already live with some merchants; management thinks that trust must be established instantly when AI agents make payments on behalf of consumers; Adyen Agentic’s payment module enables merchants to accept machine-to-machine payments in a safe and secure way; OpenAI is also an Adyen Agentic partner; navigating the agentic commerce landscape is currently what merchants are most focused on

We introduced Adyen Agentic, enabling merchants to integrate once and participate across the rapidly evolving ecosystem of AI commerce…

…Adyen Agentic is another way in which we expand our platform beyond payments. Two modules of the product suite extend our role earlier in the commerce journey: Agentic Feed structures and distributes real-time catalog, pricing, and availability data into conversational commerce environments so AI agents have accurate inventory. Agentic Cart then hooks the agent’s selection back into existing checkout, tax, fulfillment, and order management systems to build a dynamic, purchase-ready order. Because this architecture is AI platform-agnostic, merchants can integrate once and participate across the entire ecosystem. Early participants in Adyen Agentic include strategic partners American Express, Mastercard, Salesforce, and Visa, as well as enterprise retailers ESW, Scheels, Sézane, and SharkNinja…

… As AI agents begin making purchases on behalf of consumers, trust must be established instantly, without a human in the loop. The payment module within the Adyen Agentic product suite is built to enable merchants to accept machine-to-machine payments while maintaining the same high standards of authentication, fraud prevention, and payment performance that defined our platform for the past twenty years…

… I think it is good to point out on the OpenAI relationship that we work with them as an LLM, as we do with all the parties. That is to help our merchants…To be super clear it is an Adyen Agentic partner and also a customer…

…Currently, what’s top of mind is for merchants is “How are you going to help us through the Adyen Agentic threat for them?

Adyen recently won OpenAI as a customer for consumer payments

Recent wins like OpenAI show our ability to solve the most complex operational and financial challenges in these emerging business models…

…If you look at a company like OpenAI working for us, that is just for payments. That is for payments of their consumers…

…On the OpenAI, it is their payment. It is the payments which their clients pay to them, and no forward-looking statements on that.

Orb helps to reduce the complexities involved with usage-based billing and this is important for software companies because these companies are transitioning to AI-driven usage-based billing; management thinks Orb can open up a market for Adyen with very mature companies; management thinks Adyen can onboard AI native companies with Orb

Our acquisition of Orb extends this strategy beyond customer engagement into monetization. AI is transforming how software is consumed, making usage-based billing the default pricing model for many modern businesses. Traditionally, companies relied on fragmented systems for usage metering, billing, and payments, creating operational complexity. By integrating Orb with our payment infrastructure, we replace this fragmentation with a unified monetization engine spanning the entire revenue lifecycle — from usage tracking to settlement. Beyond basic billing, Orb enables merchants to continuously experiment with, optimize, and evolve their pricing strategies using real-time data. For software and AI companies, this eliminates underlying friction and turns billing into a strategic growth driver…

…Orb, if you look at AI native, that is billing and you see that opens up a market for us where we can land very mature companies…

…The reason to work with Orb is that we can onboard AI native companies. So that is separate from that. What you see is that the AI native companies grow very fast, and the billing sits in their infrastructure. So where we take that, you’ll see that in the future.

Nu Holdings (NYSE: NU)

Nu Holdings’ management introduced NuFormer, the company’s foundation model for financial behaviour around a year ago; since NuFormer’s introduction, management has focused on building a single AI platform that will power the entire Nu Holdings business; the work on building the AI platform includes growing Nu Holdings’ GPU fleet, expanding architecture research, and building on a decade of transaction history across more than 100 million customers; management recently updated NuFormer to a hybrid linear attention design, and trained the model with the Muon optimiser; any improvement to NuFormer can instantly upgrade performance across all of the company’s business lines without retraining; the latest generation NuFormer model has much higher context length and training and inference speeds, while having lower costs; Nu Holdings can now achieve the same predictive performance with NuFormer with 20 million fine-tuning data rows that previously required over 400 million, cutting development cycles from weeks to days; NuFormer now teaches nearly every decision Nu Holdings makes; NuFormer was first deployed in the credit portfolio in Brazil before it was deployed in Mexico, then unsecured lending in Brazil, and to core credit models; NuFormer is now being tested in credit cards for SMEs, and for Nu Holdings’ Colombian customers; NuFormer is also being used to handle more than 60% of customer support conversations in Brazil; NuFormer is being used to predict what a customer wants next and allows Nu Holdings to recommend relevant products; NuFormer is used to put relevant campaigns infront of customers most likely to find them useful; NuFormer is more powerful than traditional models, but managemnt is still tracking its performance

About a year ago, we introduced nuFormer, our foundation model for financial behavior. Since then,e we have focused on one objective, building a single AI platform that powers business and customer decisions across Nubank. That work spans every layer of the stack. We increased and upgraded our own GPU fleet, giving us full control of the compute layer. We expanded our architecture research efforts, and we continue building on one of our greatest advantages, more than a decade of transaction history across more than 100 million customers in three countries…

…We recently advanced NuFormer to a hybrid linear attention design, the same architectural approach behind frontier models like Kimi K3 and Qwen3.5, and we trained it with Muon, the same class of optimizer powering today’s most efficient large language models. By decoupling NuFormer’s core backbone from specific downstream decisions, any improvement to the central model can instantly upgrade performance across all our business lines without costly retraining.

The latest generation quadrupled context length, training speed, and inference speed, while reducing the cost of running models in production. As we have scaled pre-training, the base model’s understanding of how our customers behave has become deep enough to change how we build every model on top of it. To give you one example, today we can achieve the same predictive performance with 20 million fine-tuning data rows that previously required over 400 million, cutting development cycles from weeks to days.

The platform now reaches nearly every decision we make. We first deployed NuFormer in our flagship credit portfolio in Brazil. Through 2025, we replicated the model in Mexico, demonstrating that the platform generalizes across markets. During the first half of this year, we extended it to unsecured lending in Brazil and to the next generation of our core credit models. We are now testing it in credit cards for SMEs and for our Colombian customers…

…Today, AI agents handle more than 60% of customer support conversation in Brazil, with customer ratings at or above human parity.

Beyond underwriting and customer support, we are using artificial intelligence to optimize decisions across credit, deposits, and growth, moving from predicting outcomes to determining the actions that maximize value under real-world constraints. Now the same understanding of transactions that predicts credit risk also predicts what a customer wants next. It allows us to recommend the products that maximize long-term customer value, personalize the app experience, and move toward our vision of an AI private banker.

NuFormer is also improving how we grow. As the model learns our representation of how every customer behaves, we use it to put each campaign in front of the customers most likely to find it useful, and more than 100 campaigns have already run this way. 

One AI platform now powers underwriting, customer support, optimization, and growth…

…It certainly is the case that our AI generated models and assisted models are more powerful than traditional logistic regression models. That is incontrovertible. We are tracking them, though, in the exact same way that we would have tracked our historical models. We are looking at the degree of predictability, the variance at the low- end and the high- end of the predictive range, as well as the outcomes across both back-testing as well as forward-testing of that model in production.

Management believes NuFormer can be translated very quickly from Latin America to the USA, but a specifically US-tuned model will take 1-3 years; management’s focus in the USA in the beginning will be to build out Nu Holdings’ dataset to prepare for expansion once management has the same level of confidence as they do in Brazil, Mexico, and Colombia

It will take us some time to build up the same confidence in our credit risk models in the U.S. as we have in Brazil and Mexico and Colombia, where we have been operating for years. The way to think about it is that the platform, the NuFormer platform for credit models and the credit risk expertise that we have in the company, will translate very quickly across the border. But the actual data richness and building the experience of foundational testing and having the models in place that are specifically tooled for the U.S. market will take somewhere between 12 and 30 months, depending on the degree of maturation of those curves. Our priority at the beginning of our entry into the U.S. market, when that happens, will be to test, learn, build out our data set, and then be ready to expand once we have that same level of confidence there that we do in our core markets.

NVIDIA (NASDAQ: NVDA)

NVIDIA’s management sees supply constraints for 2027 (FY2028)

We expect to grow revenue by approximately 70% in fiscal 2028. This is a supply-constrained outlook…

…Customers’ forecasts point to our growth doubling next year. However, as I mentioned earlier, we expect to grow approximately 70% as we are supply-constrained. NVIDIA Compute is fully utilized across every cloud we serve.

Hyperscale revenue grew 13% sequentially in 2026 Q2 (FY2027 Q2); ACIE (AI Clouds, Industrial, and Enterprise) revenue grew 25% sequentially in 2026 Q2 (FY2027 Q2); management expects the AICE segment, which includes neoclouds, to be nearly half of NVIDIA’s data center business; sovereign AI revenue, driven by neoclouds, was up 35% sequentially and tripled year-on-year in 2026 Q2 (FY2027 Q2); management is seeing regional cloud surging everywhere, because of the fungibility and strong economic characteristics of NVIDIA compute; neoclouds are surging everywhere; management has introduced a revenue-sharing structure for neoclouds, where NVIDIA provides a minimum revenue guarantee to give lenders confidence in neocloud projects, in exchange for a portion of neoclouds’ revenues above that floor; NVIDIA is not making loans for the neocloud projects under the revenue-sharing structure; NVIDIA gets paid twice under the revenue-share structure; management thinks the revenue-share structure expands NVIDIA’s addressable market and has the potential to deliver billions in revenue; management expects the AICE segment’s computing needs to be larger over time than the whole of global cloud computing today

Hyperscale revenue of $49 billion grew 13% sequentially, driven by sustained strength in Blackwell…

…ACIE revenue of $40 billion increased 25% sequentially and 138% year-over-year. Growth was driven by neocloud capacity additions to meet the rising demand from enterprises, AI startups, and sovereigns, as well as hyperscalers purchasing capacity to supplement their own build-outs…

…Hyperscalers will remain a major growth driver, but non-hyperscaler growth, our ACIE segment spanning sovereign regional neoclouds, enterprise edge, and air gap data centers will represent roughly half of our data center business…

…In sovereign AI, our business, primarily through the regional neoclouds, grew 35% sequentially and more than tripled year-over-year in Q2…

…Because NVIDIA Compute is productive, fungible, rentable, and durable, regional cloud interest is surging around the world. We helped CoreWeave, Nebius, and Nscale build entire infrastructure businesses, and neoclouds are emerging everywhere. Firebird in Armenia, Cassava Technologies across Africa, GMI Cloud in Taiwan, Yotta and Neysa in India, Firmus in Australia, YTL AI Cloud in Malaysia, pairing local land, power, and operating expertise with our platform. Last month, we announced a partnership with Noetra, Japan’s national AI company, to build an NVIDIA DSX AI factory that will create open models to power AI agents, digital twins, robotics, and physical AI applications. South Korea’s LG and Hyundai Motor Group are partnering with NVIDIA to build and scale AI. In Europe, a record 35 new NVIDIA-powered AI supercomputers were unveiled to advance industry and scientific breakthroughs. 

Neoclouds are seeing strong demand pipelines for many diverse offtakers. Rather than allocating their entire capacity to a single long-term offtake guarantee that lenders typically require to finance a data center independently, we have introduced a revenue-sharing structure. NVIDIA provides a take or pay commitment on a portion of the facility’s capacity, a minimum revenue guarantee that gives lenders the confidence to underwrite the project, and in exchange, we share in a portion of the neocloud’s revenue earned above that floor. Independent capital still underwrites every deal on its own merits. We’re not making loans. In this model, we get paid twice, once on the hardware sale and again through the share of rental revenue, a highly reoccurring stream layered on top of a one-time equipment purchase. Over time, this model can expand our addressable market and create reoccurring usage-linked revenue stream alongside our core platform revenue, with the potential to drive billions in revenue over the medium to long term…

…There is sovereign AI, there are regional AIs, there are neoclouds, there are AI startups at enterprises where we are seeing, which represents about half of our business, and that is growing 100% a year. That part of the world’s computing is likely to be larger over time than even what we are currently experiencing in the cloud.

Half of NVIDIA’s Data Center revenue comes from hyperscalers, and the other half comes from ACIE (AI Clouds, Industrial, and Enterprise) customers, including sovereigns; ACIE customers grew 31% sequentially in 2026 Q1 (FY2027 Q1), with AI Cloud revenue tripling year-on-year; the number of partner data centers in the AI Cloud business exceeding 10 megawatts is now over 80, up nearly 100% year-on-year; Sovereign revenue was up 80% year-on-year in 2026 Q1 (FY2027 Q1); NVIDIA’s AI systems are now in nearly 40 countries

NVIDIA’s management sees a backlog of $2 trillion in the cloud industry; management expects hypercaler capex to be nearly $800 billion in 2026 and $1.3 trillion in 2027

With cloud industry backlog now greater than $2 trillion, CapEx by the top five hyperscalers is expected to reach nearly $800 billion in 2026 and $1.3 trillion in 2027.

NVIDIA and Amazon’s AWS recently expanded their partnership, whereby AWS will be deploying an additional 2 million GPUs, and Vera CPUs, from 2026 Q3 (FY2027 Q3) through to 2028 Q2 (FY2029 Q2); AWS will be serving NVIDIA’s Nemotron family of open source models; Amazon will be adopting NVIDIA’s full physical AI stack for its warehouse robots

Today, we are delighted to announce an expansion of our partnership with AWS. Building on its already vast installed base of NVIDIA Compute, AWS is deploying an additional 2 million GPUs starting this quarter through the second quarter of fiscal 2029, along with Vera CPUs, some integrated with Rubin, others standalone. AWS will serve NVIDIA Nemotron family of open models on Amazon Bedrock and SageMaker. Amazon will also adopt our full physical AI stack, Omniverse, Cosmos, Isaac, and Jetson to power its fleet of warehouse robots.

NVIDIA’s management is seeing the company’s neocloud partners bring capacity online faster and at lower cost; management expects neoclouds to have 8 GW of total capacity by end-2026, up from 3 GW at end-2025

Using NVIDIA DSX reference designs, our neocloud partners are bringing capacity online faster and at lower token cost. They are expected to exit the year with 8 GW in total installed capacity, up from approximately 3 GW at the end of 2025.

NVIDIA’s next-generation GPU system, the Vera Rubin, has brought its revenue opportunity to $40 billion per gigawatt; Vera Rubin delivers 30x higher throughput per megawatt, and 35x lower token cost, compared to Grace Blackwell Ultra systems; NVIDIA started shipping Vera Rubin in Aug 2026; management expects Vera Rubin to be the fastest product ramp in NVIDIA’s history; management expects Vera Rubin to be 20% of Data Center revenue in 2026 Q3 (FY2027 Q3); management thinks the revenue opportunity for future generations of GPU systems on a per gigawatt basis should increase materially over time, as the newer GPU systems become ever more productive

Since Hopper, our revenue opportunity has grown from roughly $18 billion per gigawatt to $25 billion with Blackwell, to $40 billion with Vera Rubin…

…Vera Rubin exemplifies this, delivering 30x higher throughput per megawatt and 35x lower token cost relative to Grace Blackwell Ultra. We commenced production shipments of Vera Rubin earlier this month. Having already received purchase orders from every major hyperscaler, AI cloud, and system OEM, we expect Vera Rubin to mark the fastest product ramp in NVIDIA’s history…

…We see Vera Rubin accounting for about 20% of data center revenue in Q3…

…[Question] As we think about the path even beyond Vera Rubin, we think about Vera Rubin Ultra and so on and so forth. Should we really conceptualize $40 billion goes to $60 billion, $80 billion?

[Answer] Is our goal to put as much compute on a plot of land? Is our goal to put more compute into 1 GW or less? Obviously, we would like the speed of light answer. The perfect answer is actually infinity per gigawatt. If we could literally get $1 trillion of compute into 1 GW and one piece of Land, Power, and Shell, it would be a fantastic outcome. The answer is directionally in that direction. We started in the world of general purpose computing during Moore’s Law. We were probably, pick your favorite number, but I am going to go with something like $5 billion, $3 billion per gigawatt of compute with general purpose computing. Then eventually with Hopper, it was $18 billion. Now Grace Blackwell is $25 billion. Next, Vera Rubin is $40 billion, and after that it is going to be higher. That is excellent. That is fantastic for the industry. It is fantastic for customers. So long as the productivity of it continues to grow, the durability and the fungibility continues to grow, then people are happy to invest in assets that generates revenues, generates profits, and helps them recoup their returns so incredibly fast.

NVIDIA’s Spectrum-X Ethernet product grew 2.6x year-on-year in 2026 Q2 (FY2027 Q2), making NVIDIA the largest and fastest-growing network company globally; there are 5 different types of networking systems needed to run AI data centers

Spectrum-X Ethernet, which grew 2.6x on a year-over-year basis, is already helping us become the largest and fastest-growing network company in the world…

…We just mentioned, each gigawatt of technology and NVIDIA’s revenue exposure in the Hopper timeframe with Hopper plus InfiniBand, and now Vera Rubin and CPU and three types of different networking, because it takes that many types of networking to address the entire world’s data center. Not to mention the scale-in security networking and the scale across multi-campus networking. You could argue five different types of networking systems.

NVIDIA’s ethernet networking product, Spectrum X, is now larger than all ethernet peers combined; NVIDIA’s other networking product, Infiniband, grew 4x year-on-year in 2026 Q1 (FY2027 Q1), driven by XDR technology

NVIDIA’s management sees agentic AI driving demand for CPUs; NVIDIA is in full production for its Vera CPU; the Vera CPU completes agentic tasks 1.8x faster, and provides 5x the bandwidth per watt, compared to other CPUs; management expects the Vera CPU to be deployed by all hyperscalers and major neoclouds and AI labs; management continues to see $20 billion in total CPU revenue in 2026 (FY2027); management expects CPU revenue to more than double in 2027 (FY2028)

Rising adoption of agentic AI is driving an acceleration in demand for data center CPUs…

…Today, we are in full production of our next generation Vera CPU. As a standalone product, Vera expands our TAM even further. Vera completes agentic tasks 1.8x faster on the spec benchmark and provides five times the bandwidth per watt than any other data center CPU. We expect Vera to be deployed by every major hyperscaler, neocloud, AI lab, and system OEM, with shipments already underway to our lead partners, including OCI, SpaceX AI, and starting this quarter, AWS. We continue to see demand for approximately $20 billion in total server CPUs. Based on our customer demand and improving supply outlook, our preliminary expectation is for CPU revenue to more than double in fiscal 2028, positioning us as one of the world’s leading server CPU suppliers.

NVIDIA’s rack-scale LPU (language processing unit) system, Groq 3 LPX, is in full production; Groq 3 LPX produces 4x the number of tokens per second against the next best alternative; management expects to ship Groq 3 LPX in volume in 2026 Q3 (FY2027 Q3)

At Hot Chips earlier this week, we announced that Groq 3 LPX, our first rack-scale LPU system, is in full production and already setting records, demonstrating nearly 4x the number of tokens per second against the next best alternative on our Artificial Analysis benchmark. We expect to ship Groq 3 LPX in volume later this quarter to early adopters. Nebius will be the first.

Global VC funding for AI exceeded $400 billion in 2026 H1; 70% of global VC funding is spent on compute; 20 AI startups have exceeded $1 billion in annualised revenue run rate, up from 13 AI startups in 2025 Q4; AI startups specialising in vertical enterprise software grew the fastest

Global VC funding in AI, roughly 70% of which is spent on compute, exceeded $400 billion in the first half of 2026, surpassing the $265 billion raised in all of 2025. Nearly 20 companies, including Cursor, owned by SpaceX, Figma, and Together AI, now exceed $1 billion in annualized run rate revenue, up from 13 companies in Q4 of last year, with vertical enterprise software logging the fastest growth.

Samsung Electronics is using NVIDIA’s solutions to achieve 20x greater performance in computational lithography

Samsung Electronics is using NVIDIA cuLitho to achieve up to 20x greater performance in computational lithography.

NVIDIA’s management sees the company as a neutral partner to sovereign and neoclouds

We don’t own a cloud ourselves. We are a neutral partner to every sovereign and neocloud.

NVIDIA’s management sees extraordinary compute demand from frontier AI labs, but the frontier AI labs’ balance sheets and credit profiles cannot support their compute needs; NVIDIA has invested $50 billion in frontier AI labs; NVIDIA has partnered with 6 of the world’s leading infrastructure capital providers to establish financing platforms that will raise more than $500 billion of 3rd-party capital to fund the compute needs of the frontier AI labs; NVIDIA recently partnered Softbank Energy for a Portsmouth data center that will exclusively host NVIDIA compute for OpenAI, and NVIDIA is providing credit support; OpenAI has existing and planned commitments for 12 GW of NVIDIA compute; NVIDIA has provided credit support for 2 GW of compute for another frontier AI lab that has also secured substantial amounts of NVIDIA compute independently; management does not see the credit support for OpenAI and other frontier AI labs as circular financing, and instead, sees attractive returns from the support with limited risk; management thinks the frontier AI labs are once-in-a-generation companies that will become the largest technology companies in history; NVIDIA’s compute for supporting the frontier AI labs is fungible and can be used by other customers; management expects NVIDIA-supported demand from frontier AI labs to be 25% of NVIDIA’s business in 2027 (FY2028); management sees the frontier labs having skyrocketing sales and fantastic margins

The frontier AI labs have extraordinary demand for training and inference compute, but they are growing faster than what their balance sheets and credit profiles can support. They have rapidly growing customer demand, yet still lack the decades-long infrastructure contracts and investment-grade financing capacity needed to secure the AI factory infrastructure independently. In other words, their growth is not limited by their technology or customer demand. It is limited by compute. For these companies, more compute means more intelligence, more users, and more revenue. NVIDIA is needed to help power this flywheel. 

First, we have invested nearly $50 billion in the frontier AI labs. This was a meaningful commitment, but it represented a small fraction of our expected free cash flow over the same period. Further, to support the frontier labs infrastructure build-outs, we recently announced partnerships with six of the world’s leading infrastructure capital providers, Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR, to establish financing platforms that will raise over $500 billion of third-party capital. With these partnerships, building on our unique, fungible, and durable computing platform, the AI labs will be able to build and assess AI infrastructure funded by long-term institutional capital at relatively attractive rates.

Last week, we announced that we secured land power shell capacity through our partnership with SoftBank Energy to exclusively host NVIDIA Compute at their Portsmouth campus. The initial deployment, expected to support 4.25 GW of AI factory capacity, will be utilized by OpenAI. Each generation of NVIDIA AI factory systems deployed at PORTS-Pike could represent approximately 1.5 million NVIDIA GPUs, and over 20 years, the site could support multiple upgrade cycles. Here is the essential economic point. The LPS commitment secures a long-lived AI factory site, while the NVIDIA Compute within the data center can be upgraded repeatedly. This project deepens our longstanding partnership with OpenAI. OpenAI has committed to substantial deployments of NVIDIA AI infrastructure through 2030. OpenAI’s existing and planned commitments represent approximately 12 GW of NVIDIA Compute. For another frontier AI lab, we will provide selective credit enhancement for nearly 2 GW of compute. This complements the substantial NVIDIA Compute capacity they have secured independently without NVIDIA’s credit support.

We recognize the scale of this support, and we know some will call this circular financing. We see it differently. We are going through a major computing platform shift, the creation of one of the most important technologies in human history, and these are once-in-a-generation companies. Their technology leadership is proven, and their customer traction and usage are skyrocketing. We expect them to become the largest technology companies in history. We believe these investments, measured against the strength of their demand, the business they create for us, the ecosystem they build on NVIDIA’s platform, and the equity returns on our invested capital will be excellent, and our risk is limited.

The NVIDIA Compute platform is fungible and durable and can be redeployed to support other customers. For context, we expect demand from the AI labs for which we expect to leverage our balance sheet to contribute toward roughly a quarter of our business next year. This remains compute we ship will be consumed by investment-grade customers or those that are backed by one…

…The frontier labs, their sales are skyrocketing. Their margins are fantastic. They are generating profitable tokens. They are only limited by the amount of compute.

NVIDIA earned minimal revenue from China in 2026 Q2 (FY2027 Q2); management does not include China in its outlook

In Q2, we shipped less than 1% of our total data center revenue in Hopper 200 products to customers based in China in accordance with the U.S. government licenses. Current Hopper shipments are dilutive to corporate gross margins, and given ongoing geopolitical uncertainty, there is no China data center compute revenue in our forward outlook.

NVIDIA is extending longer payment terms to customers

Days of sales outstanding increased to 60 days, reflecting extended payment terms for large purchases by certain investment-grade customers to be shipped over multiple quarters.

NVIDIA is experiencing extreme pricing conditions for memory chips; management has decided to absorb memory-related costs, resulting in lower gross margins for NVIDIA; the scarcity in memory chips is driven by the AI build-out; NVIDIA has good relationships with the main global suppliers of memory chips, and is working with them to increase capacity for NVIDIA

GAAP and non-GAAP gross margins were both 75%, largely unchanged from last quarter due to a similar product mix…

…Many of you have expressed concerns regarding our gross margins, as component costs have risen significantly. As you are already aware, we are experiencing extreme pricing conditions in memory. The magnitude of the price increase has exceeded our prior expectations and are headed even higher into next year. As a result, we are resetting expectations today. For Q3, we expect GAAP and non-GAAP gross margins to be 74% ±50 basis points. We expect margins to bottom in Q4 in the 71%-72% range before settling at 72%-73% in fiscal year 2028 as executed price increases take effect in Q1…

…Memory scarcity today is being driven in large part by the AI build-out itself, and unlike a component that simply raises our cost with no offset benefit. Tighter memory supply is a symptom of the same demand surge that is driving our own growth. We have longstanding, deep relationships with all three major memory suppliers, and we are working closely with them to further increase the capacity our roadmap requires.

The amount of compute required by an AI agent is 15-100 times higher than that of a human

The amount of compute necessary for an agent versus a human using it is probably 15 – 100 times, depending on the type of problem you are trying to solve. The amount of compute necessary is just extraordinary.

It’s getting increasingly more difficult to stand-up AI infrastructure; NVIDIA’s management thinks the cost to develop 1 GW AI data center has increased from $30 billion 5 years ago to $60 billion today, but the increase in productivity has also been tremendous; management thinks that the absolute-large sum of capital needed for AI data centers today, coupled with the fact that NVIDIA’s compute can be used across multiple phases of the AI life cycle and models, mean that NVIDIA has an extraordinary advantage; NVIDIA’s management thinks that the return on investment on a $50 billion data center is now less than a year 

It is also the case that you can no longer procure technology per se and stand up this infrastructure. You have got to go secure the Land, Power, and Shell, which oftentimes is a couple, two, three years out. All of the rest of the supply chain necessary to align the construction, the power, the cooling, all of the labor that is necessary…

…Each gigawatt of data center increased from, say, $30 billion about five years ago to now $60 billion today. Of course, the productivity’s tremendous. The performance is incredible in comparison. But you are talking about a $60 billion investment. To the extent that you could use it across multiple phases of the AI life cycle, run every single type of model you can imagine running on it, whether it is diffusion or autoregressive or state space or some hybrid version of that, every version of attention mechanism you can think of, small or large models. The investment that you make will be preserved and useful and productive for a lot longer time. I think our advantage in this new world is really quite extraordinary, and it could explain why it is that our growth is actually accelerating…

…I heard the other day that return on investment capital is now less than a year, and we are talking about $50 billion data centers.

NVIDIA’s management sees any procurement of NVIDIA compute by the hyperscalers to be very profitable for the hyperscalers

Back in this hyperscale space, that is growing incredibly too, right? You know that they now have backlogs of $2 trillion. You know that when they stand up NVIDIA Compute, when that happens, their revenues go up, their earnings contribution go up. Compute is profitable, very profitable today, and Compute directly translates into increased revenues.

NVIDIA’s management sees the company’s computing systems as having a different purpose to the inference-specific AI chips that the rest of the AI ecosystem, including the frontier labs, are developing; management is confident that NVIDIA’s computing systems will remain very valuable for, and widely used by, the frontier labs

[Question] A lot of these investments are designed to help the frontier labs, especially OpenAI and Anthropic, but both of them are designing their own custom chips. In fact, OpenAI just in the last few days spoke about Jalapeño and their claims about being better than Blackwell and so forth. So how are you balancing this dynamic where you want to invest a lot in the ecosystem, but part of that ecosystem wants to develop competitive solutions?

[Answer] We’re building something very different. Whereas many of these XPUs are inference-specific chips for one cloud or one service, NVIDIA is a platform, an entire AI factory platform that spans the entire AI life cycle that you can use in any cloud. It’s in every cloud. You can run anywhere. We’ll help you set it up anywhere. We built something very different. All of the AI services, at some point, are going to want to go around the world, and those data centers won’t necessarily be just built by them. They’re going to run, and I think they’re going to run on NVIDIA all around the world. And of course, I think our technology, I have 100% confidence that our technology will continue to be extraordinary for them and that the economics of using our technology, whether it’s from data processing to training, to post-training, to agentic processing, our technology’s going to be extraordinary for them. They’re going to use it. I’m very confident that they’re going to be customers and partners of ours for a very long time.

NVIDIA’s management sees a need for both open-source and closed models; management sees skyrocketing use for both open-source and closed models; management sees that nearly all open-source models are running on NVIDIA’s computing systems; management sees skyrocketing revenue and great margins for both open-source and closed model providers; management thinks NVIDIA’s position in open-source models is great because of the wide proliferation of CUDA; management thinks open-source models are reaching frontier capabilities, and is vital in cybersecurity; management thinks NVIDIA is the only platform that runs every frontier model; management thinks both open-source and closed models will succeed, and is happy that is the case

The world will need both closed models and open models. Both closed models and open models are skyrocketing in use. I would say nearly all open models run on NVIDIA, and the reason for that is because NVIDIA’s footprint around the world is the highest, and our architecture is the most fungible…

…The frontier labs, their sales are skyrocketing. Their margins are fantastic. They are generating profitable tokens. They are only limited by the amount of compute. That is equally true for open models. Our position in open models is very good because the CUDA ecosystem is literally everywhere.

The open models are also foundational to just about every AI startup and every enterprise company around the world. It is vital to them. The reason for that is because you should rent intelligence, strong intelligence, smart intelligence wherever you can, which is the reason why we rent it, and I encourage my employees to use the cloud service as much as they can. But every major company and surely every country and every startup needs to build their domain-specific, their proprietary AI, their proprietary alpha. The open models reaching frontier levels has made it possible, has enabled them to all do that. One of the areas where frontier models is vital is cybersecurity. You see the number of cybersecurity companies that are enabled by frontier models so that they could have distributed, massively distributed, continuously running autonomous cybersecurity systems to defend. Those companies are emerging…

…I’m fairly certain we’re the only platform that runs every frontier model, whether it’s closed or open. Most of them were built on NVIDIA, so they run great on NVIDIA. So we’re delighted by any model succeeding. So long as models succeed, I’m very happy. Both closed and open models are going to succeed, and they’re both simultaneously driving our sales.

NVIDIA’s management thinks that demand in the AI industry will inflect further upward when recursive self-improvement and AGI (artificial general intelligence) happens with AI models; the inflection will be driven by the proliferation of agents

[Question] There’s recursive self-improvement, which apparently at Anthropic and OpenAI is going very well with AI that improves itself. Even OpenAI said they could hit AGI by the end of this year. With the developments in RSI as well as AGI, what happens to industry demand? Does it inflect further? What does it mean for NVIDIA when those things take place?

[Answer] It’s going to inflect further. Today, the vast majority of AI is prompted by people. I believe that this last month it has crossed. Most AI are now agentic. But in the future, every company will have a whole bunch of agents. We have 40,000 employees, roughly. In the future, we’ll have 400,000 agents, 4 million agents. Those agents are running continuously. They’re running in the background. If you know anybody who builds edge personal AI agents, and they run it on their DGX Spark. I know a lot of people who run it on DGX stations, this incredible workstation that we’ve built, and you can buy it from Dell, and they’re incredible. These AI agents running on a DGX station runs 24/7, because you got stuff for it to do all the time. When the world goes to agentic, fully agentic systems, you are going to have agents running all the time, working with other agents running all the time. Those will be working in the background, improving your company, improving your lives. In a lot of ways, we are kind of recursive at this point. You could argue it is coarse-grained, but every time you run through an agent, it reflects on how it could do a better job next time, and it updates the skill file. The skills document, the markdown, is updated at the end of every single one of them. Next time you run it is going to get better. It is a kind of a loosely coarse grain self-improvement.

NVIDIA’s management thinks AI is now doing useful productive work and generating profitable tokens; management thinks the AI industry will generate more profitable tokens if they had access to more compute

I think the most important thing that matters for the industry is that, one, AI is now doing productive and useful work. Two, AI is generating profitable tokens. Three, if we had more compute, we could generate more profitable tokens, which results in more profit for all of the services. This is the exact phase where we are at.

State of the AI market?

Last year, one of the funnest things to do is just to go figure out where I go for dinner and who I have dinner with, and their stock price doubles the next day.

NVIDIA’s management sees the entire supply chain for the semiconductor industry as being challenged

Our entire supply chain is challenged. Everybody is really running flat out. More capacity is coming online all the time, which is one of the advantages of what’s going to happen this year. It’s not going to come online in an instance in time, but it’s going to come online every day.

Okta (NASDAQ: OKTA)

Okta’s management is seeing the emerging use of AI elevating the importance of identity in a company’s security posture; management thinks securing AI identity is a no-regret investment that all organisations will make, regardless of how the other areas of AI develop

The emerging use of AI by organizations and threat actors alike has further elevated the role identity plays within a company’s security posture. Organizations are accelerating their infrastructure modernization timelines to address this heightened threat environment. We are seeing conversations that begin with securing AI broaden into identity modernization initiatives…

…We don’t have all the answers, but what we do know is that there are some no regrets investments. We know that every customer is going to have to figure out where their agents are. They’re coming from all over the place. They’re going to have to figure out what they can connect to, and they’re going to have to figure out what they can do. No matter what happens at the model layer or the platform layer or the app layer, or if applications build their own agents or they get disrupted with agents, what the companies build themselves, that’s all going to unfold. What model is the best? Is it an open source model? Is it a frontier model, some combination? That’s all going to unfold as it will over the next several years. But the no regrets decision is you have to have this foundation of where are my agents, what can they connect to, and what can they do?

Okta recently acquired Permiso; Permiso detects and resolves threats across human to agentic identities in multi-cloud environments; management will integrate Permiso into Okta’s existing Identity Threat Protection and Identity Security Posture Management solutions; 80% of security breaches are identity-based attacks, but only a small percentage of Okta’s customer base has the most advanced Identity Threat Protection product that is integrated with Permiso; Permiso has 400 native risk detections

We just completed the acquisition of Permiso, a cloud-native identity security platform that detects and mitigates threats across human, non-human, and agentic identities in multi-cloud environments. Permiso Will be integrated into a unified security offering with our existing Identity Threat Protection and Identity Security Posture Management solutions. The combination strengthens Okta’s AI security offerings with enhanced visibility into autonomous agent behaviors and additional runtime controls to ensure secure agentic activity in real time…

…80% of breaches are identity-based attacks. But when you look at our customer base, relatively small percentage have the most advanced Identity Threat Protection product. Identity Threat Protection, we have talked about it for a while, Jonathan. It is very important and very unique, and by the way, very differentiated. None of the other IdPs have this. It not only evaluates session risk at the time of login, but also post-login. Continuously monitors it, looks for risk signals, not only from Okta, but risk signals from the ecosystem, from CrowdStrike and Palo Alto Networks, and takes those all together and can shut down sessions after login. So any company that is running identity without this technology, you are at risk and you are behind, but not everyone has upgraded to it… Permiso, the way to think about Permiso, it is like the next generation of that. So instead of 90 native risk detections, they have 400 native risk detections. So it is a much richer and deeper set of correlative processes and machine learning that can really look at a session deeply across many vectors and many variables and detect risk.

Okta’s management continues to see 3 advantages the company has in securing AI agents for organisations, namely, (1) distribution, (2) product breadth, and (3) neutrality; Okta’s product breadth was a driving force in its recent mult-imillion deal for Okta for AI Agents with a large healthcare company; the large health company was dealing with agentic sprawl, and Okta will give it a single control plane to discover, secure, and govern its agents; Okta will also manage the healthcare company’s entire identity fabric; Okta’s existing distribution scale of being the identity system of record for more than 20,000 customers helped it win a deal for Okta for AI Agents with a global business management consulting firm; Okta’s neutrality helped it win a deal for Okta for AI Agents with a large asset manager; the asset manager had thousands of agents in production from multiple vendors and only Okta’s neutral platform could cover all the agents without vendor lock-in; Okta’s customers see the company as being well-positioned to secure AI agents

Identity is the primary control plane for securing AI, and customers are extending the trusted foundation they already rely on with Okta’s neutral, modern, enterprise-grade identity platform to now cover agents. We continue to build on three unique advantages to help our customers navigate this shift: distribution, product breadth, and neutrality…

…Our product breadth was a key driver in securing a multimillion dollar Okta for AI Agents deal with a Fortune 50 healthcare company. AI was spreading across their organization, and they could not tell where their agents were, what those agents were connected to, and what they could do. Okta will give them a single control plane to discover, secure, and govern those agents, helping them meet strict HIPAA compliance requirements. Okta will manage their entire identity fabric, including agent governance, Privileged Access Management, and identity security, helping to ensure every human, non-human, and agent identity is managed.

Our distribution advantage comes from the reach and trust we’ve built as the identity system of record for more than 20,000 customers. We saw it at work with a global business management consulting firm that was racing to put its own AI agents into production. After considering an in-house build, the firm chose Okta for AI Agents for its single control plane for human and non-human identities, faster deployment, and lower cost of ownership. Okta will carry the agents’ identities through every handoff, binding the agents to the original employee’s delegation with a verifiable record that can satisfy client and regulator requirements.

Our neutrality was critical to an Okta for AI Agents deal with one of the world’s largest asset managers, where AI was rolling out faster than their security team could govern. Thousands of agents from multiple vendors were running in production, creating risks the organization couldn’t consistently see or control. Only Okta’s independent and neutral platform could cover their heterogeneous environment, from employees and devices to AI agents without vendor lock-in. Okta will provide visibility across all agents and enforce least privileged access so every agent gets only what it needs…

…Our place in the ecosystem is super important and super strategic. It is not just me saying that. I think it is in all these customer conversations. I am having many customer conversations. I am flying around meeting these customers that are trying to solve these security challenges in general, but in particular around AI agents. They see us as like the naturally well-positioned to secure this agentic future. We are going after that on all fronts…

Okta for AI Agents lets organisations secure and control every AI agent they deploy; Okta for AI Agents is still too small to show up in Okta’s overall numbers, but management is very optimistic about its future and thinks being the system of record for agents could be the biggest category in cybersecurity; Okta won dozens of Okta for AI Agents deals in 2026 Q2 (FY2027 Q2) and the deals included several million-plus deals; Okta for AI Agents deals continue to be bigger than the average deal size; management thinks the biggest competitor for AI deals is customer-confusion because every vendor is saying they have the answer; management is seeing a big pipeline for Okta for AI Agents and the question now is how fast the pipeline will convert; 81% of the CISOs (chief information security officer) Okta has spoken to are aware of the risks of deploying agents without adequate security; management thinks Microsoft is trying to copy what Okta is doing with Okta for AI Agents; management sees ServiceNow and Salesforce as being more complementary than competitive, with Okta helping ServiceNow build kill switches for AI agents, and helping users log into Salesforce’s Agentforce; management is currently implementing a per-user uplift pricing model for Okta for AI Agents because that’s what customers want; management will change the pricing model for Okta for AI Agents quickly based on customer feedback; there is some consumption limit with Okta for AI Agents’ pricing model; management thinks it’s possible that Okta for AI Agents will become material for Okta by 2028

Customers want to move quickly without compromising on security and control. Okta for AI Agents lets them do both by helping them discover, govern, and protect every agent…

…As exciting as that is, Okta for AI Agents, it is too small to show up in the numbers right now. Going forward, especially over the next couple of years, we are super optimistic. We think this being the system of record for agentic, for agents in the enterprise and being the system of record for agent identity. In the fullness of time, it could be the biggest category of cyber…

…Particularly strong was the 30% of the new bookings were from new products. Okta for AI Agents inside of that bucket, there were dozens of deals in the quarter, including several million-dollar-plus deals, which is super exciting…

…But what I said last time around, the average deal size for AI deals being bigger than the average deal size for the rest of Okta, that still remains the case…

…[Question] It was great to see those early AI security wins. I wanted to touch a little bit more on just competitive dynamics. Could you help us understand that scene right now?

[Answer] The biggest competitor is confusion. We are competing against confusion, so our solution has to be clarity. And customers are confused because there is so much excitement and so much opportunity in AI. It is the natural tendency of every vendor to say, “What we are doing and what we have done in the past is critical to AI. We have the answer. We have the one answer.” And I think that confuses the customer because they have 17 vendor meetings, and every vendor tells them they have the right answer…

…Last quarter, we talked about record pipe. The pipe is even bigger, and there’s more pipeline. The question is how fast it will convert, right? We don’t have four years of history on conversion. When we think about the future, there’s obviously some degree of being prudent about how fast that’s going to convert. But the pipeline’s there…

…81% of the CISOs that we talk to are aware right now that they are exposed with agents deployed in their enterprises where they do not yet have an adequate security platform in place. A very high percentage of our customers know that they have the need…

…[Question] You got Microsoft, Salesforce, ServiceNow. Maybe can you give us some perspective about to what extent do those products overlap and compete with your Okta for AI Agents?

[Answer] I think Microsoft is copying us, which I think they have been for 15 years. I think they’re copying what we’re doing, and they see the value of an agent registry. I think the challenge for them is going to be it’s really hard to be neutral, and it’s really hard to make an agent registry that works as well for Amazon and Google and OpenAI and Anthropic as it does for Azure and Microsoft. But I think they have a similar vision, and it looks at least from their blogs, I don’t know if they have a real product yet, but at least from their blogs, it seems like they are copying us. 

Then, I think ServiceNow and Salesforce are like every vendor. I think they’re coming at the problem from their perspective. ServiceNow is coming at it from a very asset management, workflow management perspective, and we found it very valuable to work with them because we can add a lot of value in that environment. We can really help them sever the connections, the trusted connections between agents and the rest of the ecosystem. We’re at that level of detail. We have the tokens, we have the protocols, so that can really help the control tower from ServiceNow actually come to fruition with a kill switch that can actually kill the connections. That’s been a really valuable partnership. Salesforce, similar, it’s like they’re coming from more of the service and support and platform layer to some degree, but guess what? People log in to Agentforce through Okta, and we can help people securely connect Agentforce to everything else in the ecosystem. Because the more Agentforce agents are connected to data across the ecosystem, the better, and we can help with that.

So everyone’s kind of sticking in their own lane, and lucky for us, our lane is perfect for this world. Our lane is people to technology and then technology to multiple different vendors with multiple different plays in the technology space. We’re very good at that…

…[Question] Just given how quickly agentic AI adoption is happening out there, how are you incorporating the number of AI agents in your deal?

[Answer] Our pricing model is per user. If you want to buy Okta for AI agents, it’s an uplift to your per user charge. The product works across different use cases, so it can be login for the user and the agent, it can be passing the agent credential across the whole chain of command, it can be governing the agent, but the pricing is per user, an extension of the per user price. Now, the first thing everyone says is, “Well, that’s crazy. Seats are going away, and you’ve got to charge per agent.” That all may be true, but the way customers are using agents now and the way they want to buy is per user. One of our advantages is we’re super close to the customers, and as I’m sure this is for sure going to evolve, and as we come up with ways that work for the customer and work for Okta, how to package it and price it differently, we’ll iterate quickly and give them what they want…

…[Question] Is there any consumption limit on those per user pricing?

[Answer] In our products, we haven’t done that much, but we’re starting to add that stuff…

…[Question] Can you talk about how much the raise in the outlook was due to AI?

[Answer] Still immaterial. Still very small. We are very early innings. But like we have talked about here, we are excited about the long-term opportunity. So for FY 2027, we do not think it is going to be material, but 2028 and beyond, if things keep going the way that they are going, then we do think that there is a real possibility for this to be material for the business in the long run. 

Okta’s management thinks the fragmented AI landscape is creating huge opportunities for Okta

The fragmentation of the AI landscape creates significant opportunities for Okta. As enterprises deploy agents across models, clouds, applications, and infrastructure, they need a neutral identity layer that can secure it all.

Anthropic recently named Okta as the first identity provider for enterprise-managed auth for MCP connectors; enterprise-managed auth is generally available; enterprise-managed auth allows organisations to centrally govern how Anthropic’s Claude connects to applications; Okta’s management recently expanded Cross App Access, Okta’s open standard for securing AI, with more integrations; Okta’s management thinks that cybersecurity in the future will take the entire AI industry to work together to secure; management thinks the addition of Anthropic to Cross App Access is a big step forward for the open standard; Anthropic’s release of enterprise-managed auth is the first time an AI agent has been compatible with Cross App Access; 26 top SaaS vendors are supporting Cross App Access

That’s why we partner with industry leaders, including Anthropic, which this quarter named Okta the first identity provider supporting enterprise-managed auth for MCP connectors. Now generally available, enterprise-managed auth enables IT teams to centrally authorize and govern how Claude connects to enterprise applications.

We also recently expanded our work with AWS, Cisco, OpenAI, Databricks, and Snowflake, alongside more than 25 new Cross App Access integrations, providing a standardized way to govern how AI agents connect to a growing ecosystem of enterprise applications and resources…

…I think longer term is something we are working on as well. I think longer term, the entire industry needs to work better together. The industry right now in security, everyone is coming at the customer saying they have the only answer. They can secure agents. They are going to be the one to do it. The reality is it is going to take us all working together. Okta has been working on this. On these calls, the last five or six calls, we have talked about standards and ecosystem. We made a huge step forward in terms of one of the main standards we have been working on, which is the standard we have talked about called Cross App Access. The huge step forward this time is when the biggest AI agent in the world, Claude, is supporting Cross App Access. They released Enterprise-managed authorization, which is the first time an AI agent has been compatible with this protocol…

…Everyone in the resource side of the equation is starting to support it as well. We announced 26 top SaaS vendors are supporting from a resource perspective this protocol.

Customers are trusting Okta to secure AI because of its track record

They share the concern that you just articulated, which is the various players in the space and the venture-funded companies are moving very rapidly, and it’s difficult for them to have confidence in predicting what the future is going to be. And one of the reasons that they come to Okta and talk to Okta is specifically because we are a proven company. We’ve been solving this problem for 17 years for over 20,000 customers, and we’ve earned the trust of those customers and the partners that we work with to solve these problems.

Okta’s management is seeing the number of agents explode inside organisations

We were talking to a company that ended up being a nice Okta for AI agent deal in the quarter. When the evaluation started, we ran our technology, and we detected 50 instances of a Claude agent in the environment. They were thinking about what they wanted to do, then we came back a few weeks, and there was 1,500 Claude agents in the environment. It’s like 50, two weeks, 1,500. These customers are really tangible, the risk that they’re seeing and the way this is coming into their organization. This is catalyzing some of these deals, this onrush of agents.

Okta’s management recently put the infrastructure in place for consumption-based pricing for Agent SSO because they expect the usage by agents to be quite high

I don’t know if you guys saw the announcement we did about supporting Agent SSO in our base edition across the board, as we did it on Monday. Agents are going to log in way more than people. So in that product, we actually have a cap of Agent SSO that we’re actually not going to enforce right away, but we’re putting the framework and the scaffolding in there to have a consumption-based pricing eventually, because it’s very likely that with agents proliferating and how they behave, that the usage is going to be quite high.

Okta’s management thinks the previous association between Privileged Access Management (PAM) and agentic is wrong

I think that this association between PAM and agentic was overemphasized. I think there was this mindset three years ago that agents needed privileged access, and so PAM was going to be the right place to do agents. I think it’s wrong. I think agents do need privileged access, for sure, but it needs to start on a much broader base. PAM is too narrow. PAM had very small number of users and super locked down environments. Agents, the whole dream of agents is that they’re there for everyone. It doesn’t make sense to start your agent journey and figuring out where the agents are and what they can do and what they connect to. It makes no sense to start it from the most locked down thing sitting next to the Oracle Database on a Sun server.

It makes sense to start it from the broad IdP [Identity Provider], whether it’s customers or whether it’s employees, and then start from there and say, “Hey, how can I take this token that was generated for this user and pass it through with traceability and accountability all the way through the chain it needs to go through?” That’s what we’re seeing in the industry. It’s a much better place to start.

Salesforce (NYSE: CRM)

Salesforce is now partnering with Anthropic to launch Claudeforce; Claudeforce is the 1st time Anthropic has been able to accelerate its go-to-market efforts within Claude; Salesforce and Anthropic are big users of each other’s products; Claudeforce has Claude Cowork layered on top of Salesforce to let users unlock trapped value within Salesforce; Claudeforce delivers intelligence within the software companies are already using to run their businesses; Claudefore is powered by a new AI harness Salesforce built called AI Force (AI Force is also powering a lot of other Salesforce AI products); Salesforce is rolling out Claudeforce to the entire company, and will make the product generally available in September; all Salesforce sellers will be demo-ing Claudeforce; all Salesforce customers will be able to buy Claudeforce via an upgrade to Salesforce’s premium editions; management thinks Claudeforce will be a new wave of demand for Salesforce; Salesforce’s management team is using Claudeforce to build agents that improve its own sales process

This is the number one AI in the world, Anthropic, and the number one CRM, Salesforce, coming together for the first time in an incredibly powerful way to build a new product called Claudeforce…

…This is the first time that we’ve really been able to incredibly accelerate our go-to-market efforts within Claude, and we want that for all the other enterprises…

…We’re big users of Salesforce. Salesforce is big users of Claude Code, of Cowork, of other tools. We’ve put products like Claude Tag in Slack already, which is a part of Salesforce…

…Anthropic builds this amazing model. And now the model has this incredible user interface, Cowork. And when you take Cowork and then you’re able to put it right on top of Salesforce, it’s able to bring the data, the applications, the semantics, the agents themselves and build complete applications, a total user interface to let you get all the value out of Salesforce that’s been trapped. So customers have put hundreds of billions of dollars into Salesforce. There is a huge amount of value that can be unleashed through this combination…

…They want intelligence delivered inside the software they’re already using to run their business. That’s what Salesforce, that’s what Claudeforce is going to do. I show them Claudeforce, I show them reasoning across their Salesforce data and workflows. I can show it right inside Cowork. I also show it right inside Slack. One click to deploy and their path to AI transformation just opens up in front of them…

…This is the best of both worlds. It is the number one AI meeting the number one CRM and putting them together. As I said, when you do that, you get this incredible result, a radically different type of interface on top of Salesforce’s entire platform. It is really powered by this incredible new AI harness that you are going to see at Dreamforce, AI Force, and that harness is able to power not only Claudeforce, but a new version of Slack that you are going to be seeing, as well as Coworker, which is something inside our Lightning interface, and other amazing things as well…

…We are rolling it out to the whole enterprise, to the whole company, and then we are going to make it, Patrick, we are going to make it available to GA in September to everyone at Dreamforce. Every single seller of Salesforce will be demoing this product as one of the surfaces…

…Every customer will be able to buy it. They will have to upgrade to our premium editions, and it is going to be a new wave of demand…

…We build our own agents based on Claude, and that agent inspects, they are our deputy CROs for all our businesses, and they inspect the business, the pipeline. They tell us if there is risk in the month or in the quarter. I do one thing that is, Marc, you would love this. I talk, by the way, I do this on Slack. This is one of the surfaces. I do this in, at the time, Cowork, now is Claudeforce. But I also do it in an app that I build with Claude Code. So depending on the moment, I use one of the surfaces. But I do one thing that is very powerful. Every time that I visit a country, and don’t tell anyone, because I don’t want my team to know. But essentially what I do is, okay, I’m visiting Italy. We have an amazing leader there, Vanessa Fortarezza. She’s our Forza della Natura. I said, “Okay, give me all the open opportunities for the month, then send a Slack to every account executive, copying the whole chain all the way to Vanessa, telling them, ‘Look at the opportunity record, look at everything, make it sound like it’s me.’ I review all the emails, all the Slacks before they go.” It all of a sudden sends 50 very customized Slack messages to the AEs, asking them if I can help, if we can do anything. It’s incredible. The engine is very smart. It proposes already ideas to close the deal. When I visit the country, everybody’s mobilized. Everybody has been working over the weekend. Vanessa is crazy running around, and we close deals faster.

Salesforce’s management thinks fears of a SaaSpocalypse are overblown; Salesforce’s net new AOV (annual order volume) in 2026 Q2 (FY2027 Q2) is the strongest in 4 years; Salesforce’s seats grew in the quarter; Salesforce’s attrition is near its lowest level; Salesforce’s pricing power is intact, with A1E (AgentForce 1 Edition) and A4X (AgentForce For Apps) bookings more than doubling sequentially in 2026 Q2 (FY2027 Q2); Salesforce’s contract length terms improved across all segments; agentic use of Salesforce was up 6x via MCP (Model Context Protocol) calls and Claude calls; 9 of the top 10 AI companies use both Salesforce and Slack, and grew their spend 435% year-on-year in 2026 Q2 (FY2027 Q2); frontier models depend on Salesforce and are not replacing the company’s products; management thinks that the combination of non-deterministic AI models with deterministic software creates value at a level never seen before; management has been pleasantly surprised that AI coding tools are making Salesforce’s product better; customers that start the agentic journey with Salesforce are at double the AOV, with a possibility to 3x-4x the AOV; only 5% of sales workers have upgraded to Salesforce’s higher-end editions and they are paying premiums of 60%-80%

I really want to start here at the beginning and talk about the SaaSpocalypse. I think you can see from the results, net new AOV growth, it’s the strongest in 4 years. Skeptics really said that seats would decline, and also that Agentforce sales and service and Slack, all seats grew year-over-year. Skeptics said customers would leave, but attrition was near its lowest level ever. Skeptics said pricing power would erode and A1E and A4X bookings more than doubled quarter-over-quarter. Contract length terms improved across all segments in new business and renewals…

…Agentic use of the platform, it surged 6x, sixfold, via Model Context Protocol calls and Claude calls…

…Nine of the 10 AI companies, like you just heard from Anthropic, that standardized on Salesforce, use Salesforce and Slack. Their spend is up 435% year over year. Frontier models depend on CRM. They make sure that it works, that it all comes together. They do not replace them…

…This is not the SaaSpocalypse. As I said earlier, we have been hearing about this for the last two quarters, these dire predictions about the end of software and how the models eat everything. But none of them have come true for us, and I do not even understand how they could possibly come true. Look, models you have to remember, models are probabilistic systems. They are non-deterministic systems. But our system, those four layers that we talked about with data and apps and semantics and agents, those are deterministic systems. But when you put those two things together, that is a level of value we have never seen before…

…I think that has been the huge shock for us, that these new next-gen tools, the Cursors, the Replit, the Claudes of the world, make our product better. That all of a sudden it can customize the data, the metadata, set up the preferences, do the administration, do things that maybe you needed a bunch of service folks to do. Now all of a sudden, you’re getting going in a month. Maybe it might have taken you, before, 6 months to a year. All of a sudden now you’re fully automated. Your ability to add new functionality and do new things, this has provided this very strong platform…

…Customers that start the agentic journey with Salesforce, on average, they are already at double the AOV with a perspective of nearly quadrupling, 3 x – 4 x the AOV…

…Only 5% of the knowledge workers that use sales and service have upgraded to the higher-end editions. We get a 60% to 80% premium.

AgentForce ARR reached $1.5 billion in 2026 Q2 (FY2027 Q2), up 240% year-on-year (was $1.2 billion in 2026 Q1, up 205% year-on-year); Agentforce and Data 360 reached nearly $4 billion in ARR (annual recurring revenue) in 2026 Q2 (FY2027 Q2), up 210% year-on-year (was $3.4 billion in 2026 Q1, up 200% year-on-year); AgentForce ARR’s growth was driven by Slackbot, Headless, and AI momentum; Salesforce’s help agent is powered by Agentforce and it has surpassed 5 million conversations, with 64% resolved autonomously; Agentforce bookings grew triple-digits in 2026 Q2 (FY2027 Q2); 50% of the bookings came from customers refilling their consumption tank; Salesforce added 2,000 paying Agentforce customers into production in 2026 Q2 (FY2027 Q2), up 70% sequentially; Salesforce

Agentforce and Data 360 annual recurring revenue (“ARR”) reached nearly $3.9 billion, up over 210% Y/Y. Agentforce ARR exceeded $1.5 billion, up over 240% Y/Y. Effective Q2 FY27, Agentforce ARR includes our AI offerings, Slackbot and Headless 360…

…Agentforce ARR reached $1.5 billion, as you heard, driven by Slackbot and Headless launches and continued AI momentum…

…As customer zero, we are putting Agentforce to work across our own business. Salesforce’s help agent has surpassed 5 million customer conversations with 64% resolved autonomously…

…[Question] I wanted to circle back on the strong Agentforce ARR growth, and in particular, what proportion of current Agentforce is coming from pay production customers rather than pilots? And how has the timeline from pilot to paid deployment and refilling credits changed over the last couple of months?

[Answer] The top line numbers are obviously very impressive, but I like to go a little bit deeper under the hood. When you look just at bookings, not just the cumulative ARR, our bookings, Elizabeth, grew triple digit, so we doubled year on year. That was pretty fantastic. 50% of the bookings came from customers refilling the tank. So they consume, they use the Flex Credits, they want more, they raise their hand, we go there…

We added 2,000 paying customers into production. That is 70% more quarter on quarter.

Many of Salesforce’s customers in Europe have not started on their AI transformations

That’s why I was in Europe. Most of them have not started their AI transformations yet. If you talk to most CEOs, they may say they have a project or they have a failed experiment, or they were told to build a model, but it didn’t work out for them. That is not AI transformation.

Salesforce is helping its customers clean up their data so they can better utilise AI

All of a sudden we’ll go, “Whoa, are you really paying attention to your AI? Because your data quality doesn’t seem to be exactly right.” All that data inside Salesforce now provides hundreds of petabytes of critical context so agents can actually understand your business. We’re doing more to help customers clean up their data, to make sure that it’s well-harmonized, to make sure that they have the right level of data cleansing, building the right data warehouses. Then our apps, well, they can be more valuable than they’ve ever been. Because now they don’t just run your business, they’re running all of your agents as well.

Slack is where AI lives and where human users can work together with AI; there are more than 1 million companies on Slack now; companies are all building agentic products directly onto Slack because that is what their customers want; Slackbot is Salesforce’s fastest adopted AI product; Slackbot has 1 million active users 5 months after launch, up 150% sequentially; Slackbot can read across all of a user’s Slack data; management thinks Slack is the best play to work with, and build with, AI; management recently introduced Slack Code for teams of agents and humans to build together; Slack Code and Slackbot make for a powerful multiplayer IDE (integrated development environment); upgrades to premium Slack editions have tripled since the launch of Slackbot; Slackbot is driving 8.1 million annualised hours of productivity gain for Salesforce employees

When you look at Slack today, it is where work happens, and it is where these AI lives and can do things. Not only is it where human work happens, now it is where you can build together, too, with Slack Code. AI is not trapped in a single player chat window. It is a teammate that works across all your people, finding the right agent, the right action on the flow of work, fueled by all the enterprise context that lives in Salesforce, the conversational context in Slack. Companies like Anthropic, OpenAI, Lovable, Replit, on and on and on, I think there is more than 1 million companies on Slack now, are all building their agentic products directly into Slack because Slack is where their customers want them. Slackbot is our fastest adopted AI product ever. Five months after launch, it has 1 million active users, up 150% quarter-over-quarter…

…It has the ability to read across all of your Slack data, and you are going to have insights into your business you just were not able to have before…

…Slack is not just the best place to work with AI, it is also the best place to build with AI. We just introduced Slack Code, a shared space for teams of agents and humans to build together right where they already work. When you combine Slack Code with Slackbot, it becomes an incredibly powerful multiplayer IDE…

…Upgrades to our premium Slack editions have tripled since we launched Slackbot…

…Slackbot is driving 8.1 million hours of annualized productivity gains for our employees.

Athenahealth used Data 360 to bring all product usage and account data into a single source of truth, and used Agentforce to connect all that data to their AI; employees across every level at Athenahealth can now get instant answers from Salesforce data, without any data going to any model; Salesforce has been building this zero data retention solution in 2023

Let me show you what this looks like for Athenahealth. Every time someone needed a case summary or opportunity insight, they had to ask IT to run a report. They used Data 360 to bring together all the product usage and account data from across their systems to create a single source of truth. Then, with AI Force, this new user interface harness that sits on top of our core systems, they have connected all of it to their AI, complete with the context and the permissions that already lived in Salesforce. Now, employees across every level can get instant answers from real Salesforce data, from real Salesforce metadata, and all of the sharing models and security models. All of this is done with zero data retention. That means no data goes to any model. It all stays in your company. We started engineering that in 2023. We have been perfecting it over the last three years, and we audit it, and we are sure that that data is staying with you. Without ever opening a browser, the usage and the value they are getting with Salesforce is going way up.

Replit’s sales team uses Slack; Replit has tripled its number of Agentforce sales seats; Replit is building custom apps that connect directly to Salesforce data and workflows

Look at that great company, Replit. You saw them in the pre-show. It’s an amazing company. Their sales team lives here in Slack with Salesforce as their source of truth. They’ve tripled the number of Agentforce sales seats. They’re building custom apps that connect directly to Salesforce data and workflows.

Uber’s Uber for Business service has a mountain of inbound leads that its employees were unable to reach; Uber used Agentforce to manage the inbound leads and generated 60% more leads within 2 weeks

Uber for Business. They’ve got 30 onboarding specialists and a mountain of inbound leads nobody was ever going to reach. Their rules, their workflows, every deal that’s ever gone through them. They’ve used Salesforce since the start of Uber, and all of that lived inside Salesforce. Now, this is how they sell. They just pointed Agentforce at the pile, got to work, looked at their whole history of their whole company, all the institutional memory, everything that’s inside Uber. Now six weeks to launch, within two weeks, 60% more leads and converting.

Robinhood put AI inside Slack, and now any employee can use Slackbot to surface past decisions and solve problems

Robinhood is a great example of what is happening. Robinhood needed AI adoption across its whole workforce, not just engineers. Vlad put AI inside Slack, the place where the entire business comes together and connects to Salesforce, Google, Okta, and other systems. Now any employee can use Slackbot to surface past decisions and conversations, solve problems without engineering support.

The CEO of Ohalo tried to vibe-code a CRM software but quickly realised how difficult it was; Ohalo decided to go with Salesforce’s CRM since it was already using Slack; Ohalo has built custom workflows, interfaces, and sales tools that integrates with Salesforce; Ohalo’s CEO realised the best way the company can create value is not to build something that everyone else does; Ohalo used Claude to build an Ask HR tool within Slack instead of building its own communications tool; Ohalo was previously using a CRM from a Salesforce competitor, but spent plenty of effort building custom tools on it without success, whereas it was really easy to build custom tools on Salesforce; Ohalao’s CEO thinks AI is destroying the value of verticalised software, but increasing the value of horizontal software such as Salesforce 

[Ohalo CEO] We did a little vibe coding. I did a vibe code to make a CRM product over the weekend. You quickly realize just how much it takes to do maintenance, to do accounts, to do security. There is just much more to it. We realized pretty quickly there are certain, I would say, software applications that are maybe verticalized, where you have to build custom workflows that make sense for your particular business. Then there is platform software, and platform software, we realized pretty quickly, is really what we need to build our entire company around. Salesforce is what we kind of decided to go with on CRM. We are already on Slack. We are not going to go vibe code Slack. We are not going to vibe code CRM. 

Then we build custom workflows, custom interfaces, custom prospecting tools at my sales teams. I have got agents that are out scanning public county databases, trying to find where farmers are planting what crops, what their names are, what their contact info is. I have got agents doing all sorts of prospecting. That interface gives my sales team leads. That integrates with Salesforce. They can then go run the standard workflows that are a little bit more standardized across verticals, and we can build all of our custom workflows in parallel and in concert with the Salesforce tool…

…I am not going to create value as a business by doing something that everyone else does. If everyone else is using Microsoft Excel, I am not going to go build Microsoft Excel. If everyone else is using an ERP tool, we integrate with NetSuite, and I know that you guys have an integration with NetSuite. We are not going to go build an ERP tool, and we are not going to go build Slack. We are not going to build communications and messaging. We are not going to build CRM. We are going to build custom interfaces for doing plant breeding. We are going to build custom interfaces for prospecting customers. We’re going to build custom interfaces for selecting the right product for a farmer based on my satellite and radar imagery that shows me what that farmer’s field is going to look like next season, making an estimation, and then making the right product selections for him. The interface to do that for my sales team, for my customers, that’s where we add value. The actual following up with the customer, connecting with the customer record, making that transaction record, storing it, that needs to be done safely, securely. It needs to be done with the right account settings, the right access. That’s what we’re not going to go build, and that’s really where the partnership works…

…Just internally, we use Slack, right? That’s our primary communication messaging tool. In Slack, we’ve got an Ask HR channel now. So we basically used Claude to set up an Ask HR channel that anyone in my company can ask the HR questions that they would normally spend time calling up the HR people saying, “Hey, I got this question. I got this question.” Ask HR has access to all of our internal HR documents, and it knows who you are, whether or not it can access your personal information. It knows standard benefits, policies, and it can just answer all your questions for you through Ask HR. I don’t need to go build a communications tool to do that. I can leverage Slack, and I can leverage Claude to deliver that internally. It’s another good example of kind of the sort of thing where we can build something custom using our internal benefits, our internal HR policies, but then we can leverage the platforms and Claude to deliver that…

…Ben works at my shop, and Ben used Claude, he used Cursor, and he spoke to your Salesforce guys, stood it up. What took us months of going back and forth trying to customize this other CRM tool and build the application and workflows around it that we needed, we were pulling our hair out. We’re like, “Let’s just build it all from scratch.” That’s when you and I talked. Ben used Claude and Cursor, got into Salesforce, and was able to get everything stood up in under a month. I wasn’t sold until we did it. Then I was like, “Okay, I’ll give Marc credit at this point, and maybe we should send him a check and pay for the product.”…

…When we got in there, and Ben was able to do this in under a month. He told me he did the whole thing in under a month…

…I was probably early on a SaaSpocalypse train, as you know, because I know you listen to my show. And I always said, like, I think SaaS was this temporary phenomenon between the founding of the Internet and the start of AI. But I think that it’s a little more nuanced than that. It’s probably this verticalized SaaS where you’re trying to standardize a vertical on a bunch of workflows, which actually destroys value in that vertical because everyone is now doing the same thing in the same way. So no one can differentiate in that vertical. We’ve actually dumped all our vertical software — verticalized software. We make all of our verticalized SaaS in-house. But the horizontal, the platform, Salesforce, standardizing on Slack. This is really where I think we’ve realized that’s not — that’s actually going to get bolstered and it’s more valuable with all the other capabilities we can now build around it. So I’m definitely sold in that sense. I think there is still a SaaSpocalypse, but with a lowercase S rather than the uppercase S, for verticalized tools, where I think AI really allows you to rebuild something that is unique, which creates value for your business in a vertical.

Salesforce has delivered 7.0 billion AWUs (agentic work units) to-date (was 3.8 billion in 2026 Q1); Salesforce delivered 3.2 billion AWUs in 2026 Q2 (FY2027 Q2), up 97% sequentially; Salesforce has 1 customer, a digital platform company, with 45 million AWUs, up 14x in 6 months; the digital platform company has an activation agent that goes to all the merchants on the platform that are not transacting to activate them

7.0 billion Agentic Work Units (“AWUs”) delivered to date across Agentforce and Slack, with 3.2 billion in Q2, growing 97% quarter-over-quarter (“Q/Q”)…

…Let me tell you one story of some of the agents that are working the enterprise that people don’t see. Big digital platform company here in the U.S. 45 million AWUs, it’s one of our largest. Grew 14 x. We started this six months ago. In Q2, the AWUs consumption grew 14 x, 14-fold. And what it does is, they call it the activation agent. They go to all the merchants that are already registered with this platform, but they are not transacting. They’re not getting money from the merchants. And the agent has 1,500 interactions every day, getting these dormant merchants alive, and then kicking, and then generating revenue.

Salesforce’s management thinks companies do not want to manage their own catalogue of models, and instead, will consume AI through packaged software; management thinks customers do not care about the underlying AI models and that the highest level of value in the AI technology stack is the application layer; Salesforce has its own AI models, and actually has been developing models since 2015

They do not want to also then say, “I have all these models.” Which kind of gets to your question, which is like, you think that these customers then are going to all of a sudden set up and maintain and build the human expertise to be able to build and maintain and train models? No. I believe what they are going to do is consume AI through packaged software. Not a huge statement, just they are consuming their AI through packaged software. You are using Slack, you are consuming AI through packaged software. You are using Salesforce, you are consuming AI through packaged software. At some level, when you use Agentforce Coworker, you are consuming AI through packaged software. When you are using Claudeforce, you are certainly using AI through packaged software. I think that’s the right way to think about it.

I think the model, it’s like, do you even know what chip you’re using? Do you really know what data center you’re on? Do you know what router you’re on, what switch, what type of fiber, what cable you’re using, what interconnect you’re using? At some level, we do get abstracted back to how the way the world used to be, which is we operate at a certain level of value. I believe the highest level of value is at the application layer…

…I did not have one customer really say to me, “Tell me, are you going to be on open source or are you going to be on frontier?” I think one way to think about this is, yes, this is a highly dynamic market. There will be frontier models. There’s going to be open source models, and there’s going to be room for all of the above…

…Salesforce also makes a lot of models too, by the way. We have a lot of our own models that run in our platform. Our platform runs on our own custom models. We’ve been writing models since 2015, and you can find them all on Hugging Face. If you go, you’ll see all the models.

Salesforce’s management is learning that customers want many different ways to pay for AI products, such as on a per user or per agent basis, on a consumption basis, on a usage basis, or on an outcome basis; Salesforce has been flexible with its pricing models for AI products, and management thinks this is a driver of the company’s large deals; management wants to move beyond just outcomes and wants to share in the economics when it generates more revenue for customers

Customers want to buy and want to price in different ways. This is something I have learned really aggressively recently. Some are still buying by user and by agent, and that is important for them. Some of them want it by consumption, and that is important to them. Some of them are just basic usage customers, and they want to pay that way. Some of them even want to pay by outcome, and that outcome could be by a transaction outcome or a business outcome. Because customers want that kind of diversity in pricing, we have created a high level of flexibility, and that, I think, has really expanded our ability to sign very large transactions with our customers…

…We’ve heard so many companies say, “We’re doing outcome pricing. We’re doing this pricing. We’re doing that pricing.” We’re just too big. We have to do it all. We’ve built these new flexible pricing schemes that let our customers choose the price that they want, and we’re going to get more and more into that zone…

…We are moving to not just outcome-based pricing, which is we completed this many phone calls, therefore give us a dollar. We want to be able to say, “No, we improved revenue by this much, so give us $2 because we made you $20, or we made you $40.” I think that that is like the next generation of enterprise software that is more than outcome-based pricing.

Sea Ltd (NYSE: SE)

Shopee’s advertising revenue grew 70% in 2026 Q2, and the take rate increased by more than 90 basis points from a year ago; number of advertisers and their average ad spend increased by 45% and 15% year-on-year, respectively; there are a number of things that helped with the growth of the advertising business, namely, (1) smart voucher, (2) the AI-powered Shopee GMV Max diagnostic tools, (3) the Brand Max feature for audience insights, (4) improvement in the advertising algorithm for matching, powered by the AI-based GR algorithm, and (5) AI-powered personalisation of content; management sees further room to increase the advertising take rate

Ad revenue was up more than 70%, and ad take rate improved by over 90 basis points year-on-year. We continued to make advertising simpler and smarter for sellers. For example, pairing ads with vouchers that are personalized to buyers to increase purchase conversion and improve the efficiency of sellers’ ad spend. Ad adoption and spend continued to improve across our seller base. The number of ad-paying sellers rose around 45%, while average ad spend per seller increased more than 15% year-on-year. Our operational priorities remain consistent, improving price competitiveness, service quality, and our content ecosystem…

…I think there are a few things helping the ad growth. I’m just listing some of the examples. One of the things, smart voucher, which is we combine a personalized voucher from a buyer together with ads, so we enhance the seller’s ad traffic, increasing the purchase conversions. Another example is we have the Shopee GMV Max smart diagnostic tools. So essentially, this AI diagnosis report and tools to help the seller to analyze how can they have better return on the ads. It’s leveraged on the AI capability to analyze the ad performance and drive improvement. We also have an in-depth audience insight for Brand Max. This feature essentially allows more sellers to view the number of shoppers in each stage of their purchase journey. How does the shopper move between stages? This will give them a more robust and algorithm-driven branding solution to capture the buyers better across their life cycles with the seller. On top of that, there’s also quite a fundamental improvement on the algorithm for the ads. Those on how can we match the buyer’s intention to the ad for better. I think that’s where the AI-based algorithm, the GR algorithm helps quite a lot when we come to the matching part. The other part is the content presentation. We’re using fellow AI tools to create better personalized content for the user when they see the ads…

…In the coming quarters, we still see that meaningful potential to increase the ad take rate. Given that many of the tools, many of the algorithms we’re implementing are still in progress, we can see a meaningful optimization potentials, while we are doing more experiments, while we are optimizing algorithms further in the coming quarters.

Sea’s management has improved Monee’s credit risk capabilities through the use of new credit models that are based on the transformer architecture similar to those supporting large language models; the new credit models have lifted approval rates by 10% while maintaining a similar level of risk; management is using more external data sources now to assess new users; management has used AI-built tools to reduce review times of user-submitted income documents by 95% while maintaining accuracy

One key enabler of our credit business growth has been the ongoing advances we have made in our credit risk capabilities. Our latest risk models are pre-trained on a broad set of behavioral and transactional data across our ecosystem using transformer architecture similar to those following today’s large language models. The model learns from the full sequence of a user’s actions over time, capturing richer context around how customers interact with our platform. Recent enhancements to our underwriting models have helped lift approval rates by around 10% when compared to previous models while maintaining a similar level of risk…

…We are also drawing on more external data sources to better assess users who are newer to our ecosystem. For instance, through partnerships with local mobile operators in Indonesia and open finance data in Brazil. 

We have also used AI to build tools to efficiently verify a diverse range of user-submitted income documents across markets, languages, and formats. Review time reduced by around 95% while maintaining a very high level of accuracy, letting us respond to credit limit requests from users almost instantly…

Sea’s management is launching an AI assistant for sellers in a number of markets; the AI assistant is essentially a digital key account manager that sellers can talk to; the AI initiatives of Sea for the Shopee business on the buyers side have led to better conversions

We are launching the AI assistant for sellers in quite a few of market. Essentially, instead of the seller talk to a key account manager, the IM, as we call it, there is a digital IM that they can talk to, which can help them to answer many questions or many analysis they want to do with their shops. This is also 24 hours available, of course, compared to key account manager usually are not available 24 hour by 7…

…On the buyer side, we spend a lot of efforts on both helping the apps have better conversions, which reflecting our ad take rate improvement over times, but also just general conversion for our search recommendations. We’ve been rolling out our new GR algorithm, a generative algorithm for recommendation and search, which give us a meaningful improvement on the conversion rate that we observed. We’re also doing follow-up work on AIGC on content. If you look at our platforms, we have a lot more contents can be generated by AI now, which can be used to do a personalized targeting for our buyers to improve the conversion as well.

Tencent (OTC: TCEHY)

Tencent’s management is seeing its existing businesses grow partly because of AI enablement; the growth of the existing businesses provide financial support for Tencent’s AI initiatives; management believes that they can upgrade Weixin for the AI era in a cost-efficient way and accelerate the growth and monetisation of the Weixin ecosystem; management is leveraging Tencent’s Hunyuan model for AI teammate creation and code review in the games business; Tencent’s AI Marketing Plus automated campaign solution was upgraded to better support closed-loop WeChat Minishop and Mini Drama advertisers; management scaled up the parameters of Tencent’s advertising AI recommendation system which led to better advertising conversion rates

Tencent’s existing businesses are growing solidly due to intrinsic modes and AI enablement. As discussed early this year, our modes arise from factors including network effect, depth, and value added along supply chain, IP, low tick rates, regulatory requirements, and private data. In addition to these modes, we’re further deploying AI to boost returns in areas including WeChat, games, and advertising. As a result, our existing businesses provide a very strong financial support for our new AI initiatives…

…As we upgrade Weixin for the AI era, we can do it in a cost-efficient way, and we’re confident that AI will over time accelerate the growth and thus the monetization of the entire Weixin ecosystem, generating attractive return for us…

… In games, we are leveraging Hunyuan for AI teammate creation and code review for games, including our flagship game, Peacekeeper Elite…

… In terms of production, the Delta Force team have integrated AI across multiple workflows, including using data agents for performance analysis and the Hunyuan 3D model for asset generation…

…We upgraded AI Marketing Plus end-to-end execution capabilities to better support closed-loop WeChat Minishop and Mini Drama advertisers. For example, AI Marketing Plus now enables WeChat Minishop owners to automatically select products for promotion, generate product-relevant ad creatives, and then run smart bidding to buy inventory for those creatives. We significantly scaled up the parameters of our advertising AI recommendation system to capture user interest with greater granularity and thus improve ad conversion rates. 

Tencent’s new foundation model, Hunyuan 3, has leading cost performance; Hunyuan 3’s production version has substantially better performance than the preview version; Hunyuan 3 had notable improvement in task completion rates and reducing hallucination and errors, because it was trained with feedback loops from Tencent’s product teams; management thinks Hunyuan 3 has strong agentic capabilities and product experience, and thus has advantages for coding, office work, financial modelling, and front-end design; Hunyua 3’s production version has 6x higher average daily token usage compared to the preview version; Hunyuan 3 is consistently among the top 3 models on OpenRouter by token usage; management sees Hunyuan 3 as an important stepping stone for developing frontier models with efficient cost-performance in the future; management has integrated Hunyuan into WorkBuddy, Yuanbao, the games business, and Weixin; the integration of Hunyuan into Tencent’s products is providing valuable feedback for model-training which enables faster model iteration and sustained performance gains; management is accelerating the improvement of the Hunyuan family of models, and is training Hunyuan 4, which is expected to be released later in 2026; management is confident that Hunyuan will reach frontier levels; management wants to develop a frontier model because they think it will help with unit economics, feature innovation, and allow Tencent to capture more value; Hunyuan 3 is a small model but is widely used and has certain characteristics that match or beat much larger models; Hunyuan 3 is focused on use cases rather than beating benchmarks, so it’s more useful than many larger models; management thinks products using Hunyuan 4 will become even more useful once Hunyuan 4 is released; management already has plans for Hunyuan 5; even when Tencent’s models are at frontier levels, management still expects the company to have multiple models for different needs

We’ve made significant progress in constructing a robust foundation, including a substantially improved Hunyuan 3 foundation model with leading cost performance…

…The release of Hunyuan 3’s full production version is very successful, showing a substantial step-up in performance compared to the Hunyuan 3 preview version. Leveraging the feedback loop from product teams to improve the quality and diversity of data used for post-training, and by scaling up reinforcement learning, Hunyuan 3 achieved a notable improvement in task completion rates and meaningful reduction in hallucination and error rates. The improvements in Hunyuan 3’s capabilities are most evident in its agentic capabilities and product experience. The model’s performance step-up across reasoning, agentic, and long contest tasks delivered clear advantages for use cases such as coding, office work, financial modeling, and front-end design…

…The approximately 6x increase in average daily tokens usage of Hunyuan 3 compared to the preview version across all channels during the pay period…

…Hunyuan 3 consistently ranks among the top 3 models globally on OpenRouter based on token usage. Hunyuan 3’s production version has performed well and will serve as a stepping stone toward the Hunyuan family of models, achieving state-of-the-art capabilities in the future while providing users with the cost-performance efficiency that they need today.

We have been integrating Hunyuan into our products, making great impact. For WorkBuddy, Hunyuan can facilitate complex agent workflows with higher task success rates and reduced time to completion. For Yuanbao, Hunyuan delivers leading execution quality in information retrieval, data processing, document workflows, and everyday decision-making. In games, we are leveraging Hunyuan for AI teammate creation and code review for games, including our flagship game, Peacekeeper Elite. In Weixin, we deployed Hunyuan for powering the AI assistant in official accounts and the developer tools for mini-programs. At the same time, product integration is making Hunyuan better by continuously feeding real-world product usage and domain feedback into model training. Our model product co-design approach allows Hunyuan to validate model accuracy and identify and work on edge cases, enabling faster model iteration and sustained performance gains…

…We are accelerating the improvement of our model. We are scaling more powerful reinforcement learning to substantially upgrade models after pre-training is done. We are in the process of upgrading multimodal capabilities. More importantly, we are training a larger parameter model, Hunyuan 4, which we expect to release later this year. By accelerating the technical iteration and pushing the boundaries of model intelligence, we are confident Hunyuan’s capabilities will reach state-of-the-art level…

…The rationale behind investing in our own foundation model is that we can achieve better unit economics, more innovative features, and more exposure to the value of intelligence through co-design across our applications, our model, and our compute infrastructure, especially at this early stage of AI diffusion…

…Hunyuan 3 is a very small model, even in today’s terms, but it’s actually very widely used. I think there are a number of characteristics of Hunyuan 3, which is, it actually has the capability of matching or beating much larger models… 

…It’s actually focused on use cases rather than just benchmark beating. As a result, in real life, it has become much more useful than a lot of models of the same size or even bigger size…

…When Hunyuan 4 comes around, the products that would be using Hunyuan 4 would actually become even more powerful and even more useful than what they are today, and that would actually provide a very significant lift for the products that it’s powering. I think that’s the path, and Hunyuan 4 is only another stop. Then we’ll be upgrading to Hunyuan 5. As we continue to progress, we will be approaching SOTA, and at some point in time, we’ll definitely be able to reach SOTA. Once we are there, we would also have a lot of models of different sizes that will be able to solve different kinds of user problems at the different level of model and cost efficiency. At the same time, we have multiple models that can be used for co-design with our different products, and that would help us to make the products feature-rich and help to make the products powerful as well as the speed of execution will be fast.

Tencent’s management believes they will generate a significant return from building AI-native new businesses for the company; management thinks Tencent could earn a significant return from its capital expenditure immediately by renting out compute, but they believe even better long-term returns can be earned by building frontier models and building applications on top of the models; management can actually resell compute at a 30% profit today compared to what Tencent paid for them a few months ago; management thinks the downside for the AI investments are protected; management is currently investing prudently in AI through Tencent’s operating profit; management will step up the investments if they see opportunity to generate outsized returns; the AI-drag on Tencent’s operating profit is also very dynamic in terms of where it its allocated

We believe we will generate significant return in building a large and valuable AI-native new business for Tencent…

…Given the surge in demand and therefore rental pricing for compute, we could recover the depreciation almost immediately by renting the compute out to third parties as many neo cloud businesses are doing. We would then achieve a decent return in an immediate timeframe. However, in reality we are playing a different game or executing a larger strategy in that we are allocating a very substantial proportion of the new compute to building our own models to state-of-the-art status, and also to deploying, popularizing, and bringing our own AI applications to market leadership in China. Our belief is that by providing the superior intelligence that we can achieve through state-of-the-art models, through market-leading AI applications, that superior intelligence, we can then convert into superior economic returns over the longer term, for example, by selling tokens through the WorkBuddy application….

…Today, if we can actually allocate the compute toward leasing on the Tencent Cloud would actually generate a lot more revenue and would generate significant return from the CapEx. As a matter of fact, for the prepayment and for some of the compute orders that we had made just a couple of months ago, today, we can actually sell that at more than 30% profit compared to the price that we paid just a few months ago…

…The AI investments we’re making are mostly in AI infrastructure, and in the worst case, which we do not believe that would happen, we can choose to rent that infrastructure out at cost recovery or even better prices via Tencent Cloud if needed…

…[Question] The new AI product drag rose from roughly RMB 8.8 billion in first quarter to about RMB 10.5 billion this quarter. Can you walk us through how you manage that investment?

[Answer] it is actually very dynamic. I think we would be investing prudently until the point that we actually see breakout opportunity, then we may step up the investment. I think that is essentially the way we look at it, right? It will be a certain percentage of our profit, but if clearly we see that if we step up the pedal, it would actually generate a lot of returns, then we may step the pedal…

…It is also the case that we dynamically reprioritize the spend within the budget or within the envelope. If you look at where the RMB 8 billion in the first quarter flowed in terms of user acquisition spending and so forth, and which products it supported, versus where the RMB 10.5 billion in the second quarter flowed, there was actually a very big change because we identified that WorkBuddy was breaking out. Therefore, we aggressively prioritized WorkBuddy while deprioritizing some of the other products in that new AI product portfolio.

Tencent’s WorkBuddy and CodeBuddy services have breakout user growth and are the clear leading office productivity services in China; WorkBuddy allows users to orchestrate multiple agents for complex work, and can be controlled by Weixin; WorkBuddy has access to over 70,000 skills from Tencent Cloud Skillhub; WorkBuddy has high retention rates and users are willing to pay; WorkBuddy is attracting a growing developer community because Weixin Pay is embedded within, and this enables payouts to developers when their skills are called; management believes there are substantial opportunities to be unlocked in the productivity market; management believes Tencent’s AI productivity products will have attractive economics through growth in paying users and reduction in token costs; management sees WorkBuddy as a new way for Tencent to monetise existing enterprise relationships; management sees WorkBuddy as a new platform which acts as an orchestrator of different models and skills to perform work economically; management sees Tencent’s Hunyuan as one of the models provided by WorkBuddy, but it would not be the only model; management primarily monetises WorkBuddy today through subscriptions, which creates a time lag between cash receipts and revenue; management is seeing a substantial ramp in WorkBuddy cash receipts

Our AI office productivity workspace, WorkBuddy, and coding tool, CodeBuddy, are achieving breakout success in terms of capability and user growth. They are the clear leading office productivity service in China based on monthly interactions.

WorkBuddy serves as a one-stop-shop workspace that orchestrates multiple agents to handle complex work from end to end. Users can remotely control WorkBuddy via Weixin and WeCom, as well as WPZ, and access to over 70,000 skills from Tencent Cloud SkillHub. Besides the rapid user adoption of WorkBuddy, it is also achieving high retention rates and high willingness to pay among users as it directly contribute to users’ productivity. It also attracts growing and more vibrant developer community by embedding skill pay and Weixin Pay inside task flows to enable payouts for developers when their skills are called. This progress supports our view that there are substantial opportunities to be unlocked in the productivity market, including coding and existing office work scenarios…

…Over time, product economics will be attractive as enhanced premium benefits accelerate paying user growth, while we can reduce token costs through agent efficiency, inference efficiency, and model optimization.

Given Tencent applications such as Weixin, WeCom, and Tencent Meeting are already widely used by enterprises, WorkBuddy provides a new way for us to monetize our enterprise relationships…

…[Question] Whether WorkBuddy, in your mind, is a piece of enterprise software that sits next to Tencent Docs, or is this a new platform play that essentially becomes a marketplace for AI in the future?

[Answer] I think it is indeed a new platform. It’s a very flexible workspace for agentic AI. The core purpose is actually it will solve all the productivity needs of office workers and of all kinds of people who engage in their own businesses, right? One-person companies and the like. And below that, there will be a harness which actually helps the users to make use of the capability of different models to solve the agentic problems of the users. And over time, there will be many models serving the users through WorkBuddy. There will be many skills developed over time by all kinds of different developers. And the purpose is actually solving productivity problems, and then the platform itself would make use of all kinds of different tools and models available to do that. Then, of course, we are the orchestrator. So we can actually choose the right model and choose the right skills to help users solve the problems. And we choose that to make sure that the work is done perfectly, but at the same time, it will be done also very economically,..

…But at the same time, if you can actually solve a lot of the user problems, right, and it’s quite effective, then Hunyuan would actually be one of the main models within WorkBuddy, but it would not be the only model…

…The majority of the WorkBuddy spending by users is on subscriptions. And so similar to games and some of our other businesses, there’s a lengthy time lag between the cash receipts coming to us from the users and those cash receipts translating into reported revenue. But we are seeing a substantial ramp in the cash receipts today, and that will translate into reported revenue growth for Tencent Cloud as we move through the year.

Tencent’s management recently released a prototype of Xiaowei, an agentic AI within Weixin that leverages Weixin’s assets and that is powered by a Weixin-customised model; Xiaowei will be rolled out with a phased approach as management upgrades its capabilities over time; when mobile appeared, the QQ ecosystem was magnified by Weixin, and management believes something similar can happen for the Weixin ecosystem with the addition of Xiaowei; management believes Xiaowei will be part of agent-to-agent interactions within the Weixin ecosystem

We recently released a prototype of Xiaowei, which delivers an embedded and context-aware agentic AI experience within Weixin, leveraging Weixin’s social graph, knowledge graph, merchant reach, and payment functionality. Xiaowei is powered by the Weixin customized model, WeLM, built with a focus on user privacy, Weixin-specific use cases, and cost efficiency. Xiaowei can help users navigate and derive insights from Weixin’s diverse content universe in a personalized and efficient manner. Xiaowei can also leverage Weixin’s unique mini-program ecosystem to help users discover products, make purchase decisions, and place orders, laying the groundwork for an agent-to-agent transaction loop…

…Xiaowei will be rolled out to broader user base in a phased approach as we work on several core initiatives to elevate the user experience. These include upgrading Xiaowei’s dialogue, memory, and recommendation capabilities, expanding service and content integrations, scaling our AI infrastructure, and upgrading our harness to support a significantly larger user base…

…If you imagine the time when QQ was a communication and social tool in the PC stage, then when we get into the mobile age, Weixin appears and Weixin essentially, the ecosystem magnified QQ’s value by more than 10x, because it is enabled in the mobile age and it becomes mobile first. When we look at AI, we believe there is another huge opportunity for the Weixin ecosystem to be first enabled by AI, and over time, it will be AI first application and ecosystem. When that happens, users would have a lot of great experiences…

…I think we are envisioning a future in which a lot of users would be executing their instructions and over time transactions via Xiaowei and via agents. In the past, if you think about the Weixin ecosystem is users interacting with content, interacting with mini programs themselves. In the future, if they can actually send a complex instruction to an agent, then an agent can actually start helping the user to execute transactions. A lot of the mini programs, a lot of the merchants would actually also have agents, which over time can interact with the agent of the users.

Within Domestic Games, management has integrated AI into Delta Force

On domestic games, “Delta Force” achieved lifetime high average DAU in the second quarter, driven by the Burst Fest campaign, the game’s first professional esports final, and a global 20 versus 20 tournament. In terms of production, the Delta Force team have integrated AI across multiple workflows, including using data agents for performance analysis and the Hunyuan 3D model for asset generation. 

The Marketing Services segment’s revenue was up 22% year-on-year in 2026 Q2, driven by higher eCPM and impressions; Tencent’s AI Marketing Plus automated campaign solution was upgraded to better support closed-loop WeChat Minishop and Mini Drama advertisers; management scaled up the parameters of Tencent’s advertising AI recommendation system which led to better advertising conversion rates

For marketing services, revenue grew 22% year-on-year to RMB 44 billion, driven by higher eCPM and impressions. Most major categories increased their marketing spending with us, including e-commerce, internet services, and local services. We upgraded AI Marketing Plus end-to-end execution capabilities to better support closed-loop WeChat Minishop and Mini Drama advertisers. For example, AI Marketing Plus now enables WeChat Minishop owners to automatically select products for promotion, generate product-relevant ad creatives, and then run smart bidding to buy inventory for those creatives. We significantly scaled up the parameters of our advertising AI recommendation system to capture user interest with greater granularity and thus improve ad conversion rates.

Within the Fintech and Business Services segment, the cloud revenue within the Business Services sub-segment had low-20s percentage year-on-year increase in revenue in 2026 Q2, up from high-teens in 2026 Q1, driven by AI-related demand; Tencent’s international cloud business grew rapidly in 2026 Q2, driven by AI; in the international cloud business, CodeBuddy is enabling Tencent to conduct customer cloud migrations over to Tencent Cloud faster than in the past; token prices in China are low, but token manufacturing costs are also low, so Tencent Cloud’s AI business can have positive gross margin even with low token prices; Tencent Cloud’s new AI businesses have similar gross margins for paying users with the overall Tencent Cloud; management sees the overall pricing environment in China’s cloud market to be easier than in the past, even as memory supply costs have increased

Within business services, while we’re still working through capacity constraints, our cloud revenue growth rate accelerated from high teens percentage year-on-year in the first quarter to low 20s percentage in the second quarter, benefiting from AI-related demand, international expansion, and increased usage and pricing for general cloud services. AI-related demand translated into increased revenue across GPU rental, Model-as-a-Service, and WorkBuddy and CodeBuddy token usage. Our international cloud business expanded rapidly. Using skills developed with CodeBuddy is enabling us to conduct customer cloud migrations over to Tencent Cloud faster than we could in the past, for example, on behalf of a leading telecom company in Indonesia…

…It is true that domestic token prices are low, but the domestic token manufacturing costs are also extremely low. I think much lower than widely perceived or externally estimated. The token business, it can be positive gross margin at these low token prices because the cost is low. If you look at the gross margin for the paying users of WorkBuddy, or you look at the gross margin for our models of service, then the gross margins today are already comparable to the gross margins for Tencent Cloud overall. Of course, WorkBuddy in aggregate has a lower gross margin because there is a proportion of free users whom we are subsidizing to drive market share and market growth. On the paying users, we are generating a pretty good gross margin right now…

…It is true also that the China cloud market is price competitive. But that environment has changed a great deal in the last several months as the input costs, particularly for memory, have gone up. We have been increasing the prices we charge to our customers. We increased prices across the board in May for Tencent Cloud, and beyond those headline price increases, we have also been more substantially reducing discounts. The overall pricing environment in cloud in China is not as difficult as it has been in the past.

Tencent’s operating capital expenditure in 2026 Q2 was up 190% year-on-year and up 66% sequentially because of accelerated investments in AI infrastructure; free cash flow was negative in 2026 Q2, and down year-on-year and sequentially; management now sees Tencent’s capex as having 2 components, one for the existing business, and one for the new AI-native businesses; the upfront investments in compute are needed to get the AI-native businesses kickstarted; the primary use of Tencent’s capex is for training the next generation of Hunyuan models and the secondary use is for inference for the use of Tencent’s 1st party models and 3rd party models; management expects to have sufficient compute capacity toward the end of 2026 and early-2027 to rent out GPUs, but management thinks the production of tokens by WorkBoddy has better economic value; management sees the AI capex as a one-off thing for 2026 and 2027 and does not expect big capex spending by Tencent every year

Operating CapEx was RMB 51.8 billion, up 190% year-on-year or 66% quarter-on-quarter as we accelerated investments in AI infrastructure to support Hunyuan Model enhancements with PAPI and coPAPI inference needs, Huaxin AI initiatives, and development of AI capabilities across our products and services, as well as to meet growing external demand for our cloud services. Non-operating CapEx was RMB 1 billion. Free cash flow was -RMB 13.8 billion, reflecting large AI infrastructure CapEx and AI-related prepayments, as well as seasonally lower games gross receipts. Excluding the prepayments for compute procurement, free cash flow would have been RMB 37.6 billion…

…I would say the CapEx will be divided into two parts too, right? One part is really in relation to our existing business, which you can just like in the past, right, you can just say, oh, this is the free cash flow in which we generate operating cash flow, and there is a CapEx in relation to that. That part of the business still very cash flow generative. Then there is another set of CapEx which is related to the new AI native business, which is essentially a lump sum that we need to invest in order to get our compute for model training, in order to prepare for inference needs, and in order to also order some more for building our AI compute and AI Cloud business…

…The reason we are actually investing in all these compute is that we need that in order to essentially get the business kick-started…

…The immediate primary use case for the CapEx is for training bigger and better Hunyuan models in the coming months… An important secondary use case is providing inference for the use of Hunyuan models as well as DeepSeek and other models behind WorkBuddy…

…Then toward the end of the year and into next year, we will also have sufficient GPU ASIC capacity to step up in terms of Tencent Cloud renting out bare metal GPU or providing Model-as-a-Service. But within those opportunities, renting out GPU Model-as-a-Service and then token production for WorkBuddy, we think that it is token production for WorkBuddy that carries the most enduring economic value to us, and that’s why we’re prioritizing it today…

…When we look at the CapEx that we allocate for building the AI native business, it is more of a sort of a lump sum that we are going to be investing this year and next year. I think one should not assume that it will be sort of new every year, because the model-building part is more of a fixed cost that you actually have to get enough compute, but it will not be every year you have to invest more.

Tencent’s management thinks on-device inference will only happen over the long run, but there’s a high likelihood of it happening based on a study of computing history; but for this scenario to happen, the AI model architecture itself will be important

In terms of on-device inference, I think it would, number one, be happening maybe step by step and it will be only over the long run that most of the inference will be happening on device. But I think at some point in time, it is not hard to imagine some kind of inference will be actually happening on device, and some inference will be happening in the cloud. Over time, as the on-device compute becomes more and more powerful and as the model becomes more and more efficient, you will have more inference happening on people’s devices.

I think that would be going back to the normal state of the computer industry. If you think about the computer industry as well as the smartphone industry, most of the compute, which is CPU, actually happens on device. The cloud actually only is responsible for a small part of the compute. In this initial phase of AI infrastructure, most of the compute, because it has to be very powerful. The problem of getting enough compute on device, getting it cheap enough, and also getting it power-efficient enough has not happened yet. So that is why everything happens on the cloud. There will be a time in which more and more GPU capability will be put into everybody’s phone and computer. When that happens, then more and more inference will be happening on the device, and there will be going back to the time when it is actually the software, it is actually the model that becomes much more important. The return for running models and the return on running applications will be higher because the compute CapEx will be not just borne by the model company, but it will be borne across the ecosystem. I think that would definitely happen at some point in time and we are building and preparing for that.

Veeva Systems (NASDAQ: VEEV)

Veeva’s management thinks Vault AI increases the value of the company’s applications; in August, management released more standard Vault AI agents, expanded existing Vault AI agents, and improved Vault AI’s ability to develop custom agents

Vault AI increases the value of our core applications…

…In early August, we crossed a key milestone for Vault AI with the release of additional standard agents, the expansion of existing agents, and more advanced custom agent development capabilities.

Veeva’s management thinks Veeva Falcon helps open a new big market for agentic labor; development of Falcon agents for clinical, regulatory, and safety is progressing quickly; Veeva is working with 5 early adopters of Falcon; management expects early adopters of Falcon to go live in 2026 (FY2027), and for the first top 20 biopharma to go live in 2027 H1 (FY2028 H1); the early adopters of Falcon are sponsors (meaning the entity behind a drug undergoing trials) and management’s current priority with Falcon are the sponsors; management acquired Copli in June 2026 and launched Falcon MLR (Medical, Legal, and Regulatory) for automated review of commercial content; customer demand for Falcon MLR is high; management thinks Falcon MLR can remove at least 70% of manual labour in the MLR review process over the next 5 years; management thinks Falcon is addressing things that are high on biopharmas’ priority lists; management sees very high customer interest in Falcon, with Falcon possibly flying off shelves if it was already available; management sees Veeva’s speed of development of Falcon as the bottleneck; Falcon is a completely new area for Veeva, but fits really well with the company’s existing products; there’s no need for a lengthy or complicated implementation process with Falcon; management thinks Falcon will have a relatively easy selling cycle because the product is not sold to IT departments; the buyers of Falcon are mostly same people Veeva has been selling to for the past decade; management expects Falcon to have similar gross margins to Veeva’s traditional software products because the agents will be working mostly with deterministic software; management thinks Falcon will still be a great business even if AI models do not become cheaper over time because Falcon is pushing a lot of work into the deterministic software layer; the labour Falcon will be replacing will be for both internal and outsourced work, but not outsourced work to CROs (contract research organisations); the pricing model for Falcon is still unclear to management, but it may be enterprise license agreements that escalate over time; in early tests, Falcon outperformed humans

Veeva Falcon represents a big new market for agentic labor…

…Development of Falcon agentic labor for clinical, regulatory, and safety is progressing quickly. We are working with five early adopters to test and refine our agents using real customer data, with more early adopters on the way. We expect to have the first early adopters go live this year and the first top 20 biopharma go live in the first half of next year…

…Our early adopters are with sponsors. Now, we have had some interest with service providers. I say that carefully, interest. They’re interested, but we really haven’t engaged heavily there yet because you have to be focused when you start working with your first customers. Service providers will have similar needs to sponsors, but not the same. We’re focusing on the sponsors first, and I fully expect over time that this will be useful for outsourced service providers. We have to work on the sponsors first… 

…In June, we acquired Copli and launched Veeva Falcon MLR, our agentic solution to automate the review of commercial content. Execution is going very well, and customer demand is high. Over the next five years, we believe Falcon MLR can eliminate at least 70% of the manual labor in the MLR review process across the biopharma industry…

…I think Falcon has a lot of interest right now because it’s very clear that that’s high priority. Quick cost savings and compliance and efficiency. That’s high on everybody’s priority. I think there’s a lot of interest in Falcon. We’re the rate limiter right now. We have to get that product ready, start working with the early adopters, but interest in Falcon is very high…

…If we had our early adopters live and successful right now, I don’t want it be hyperbole, but Falcon would be flying off the shelf if that was the case…

…It is a major change for Veeva. Falcon is agentic labor. That’s something different than we’ve done before. We’ve done cloud software, data consulting. Now we have this fourth thing, agentic labor. It is transforming the discussion. There’s two different things you could do with Veeva. You can do some agentic labor, you can do core applications. That was never the case before. The important point is Veeva, it fits very well. It’s a structural advantage for Veeva to both have the agentic labor across multiple areas in life sciences, and have the core applications across multiple of those areas in life sciences…

…One thing to know is there’s not an extensive Falcon implementation. There’s not a data mapping from one system to the other. There’s not a cut-over process. There’s not ETL to do. This implementation, the full value is faster with Falcon…

…it’s actually going to be an easier selling cycle because IT is really not involved in the agentic labor. That’s not something they’re involved in. Because it’s not like that. If you’re selling a solution to safety, this is about the budget of the safety team. It’s really the head of the unit, the business unit, and the head of the operations of that business unit… 

…We have not hit the case for Falcon where we’re selling into a buyer that we are not selling into, because we’re always selling into the business side with our business applications…

…These areas where we’re doing Falcon, they were ready, I would say, on the average, 60% of business sell, and those are people that we’ve been selling into for 10 years…

…I don’t really want to make predictions on Falcon because it’s early, but in general, I don’t think we’re going to have a gross margin problem. I think the gross margins will be roughly similar to our software. Here’s why. When we really go deep into Falcon and we have what we call Falcon copies, where we have the real customer data that we’re testing the agents with and developing the agents with, we know what’s going on. More and more of that work goes into the deterministic software…

…Even if that wouldn’t happen [referring to lower prices for AI models], I think Falcon would be a great business because we’re pushing a lot of things into the deterministic layer…

…[Question] Do you believe that the work will be shifting to Falcon is something they were outsourcing to other partners such as CROs?

[Answer] In terms of where the labor is done or where it will be displaced, I think there will be a combination of internal and outsourced, although generally not the CROs…

…[Question] Any incremental color on the pricing of the Falcon products?

[Answer] They want predictability for that because for one thing, they get that predictability when they either hire or outsource labor. It is quite predictable, and it is better for them. It is actually better for us too. W hat gets in the way of that a little bit is, well, Falcon is quite early now. It can do certain things, but it cannot do the things that it will do three years from now. How do you have a fixed price when your capabilities are rapidly improving? I think with some of our customers, we will end up having enterprise license agreements, enterprise subscription agreement for the labor based on the size of their company or their function, but it will probably escalate over time. It will be lower in the beginning when Falcon is less mature. If you want a teenage Falcon, it costs you X, and if you want a Falcon that is 25 years old, it costs you a bit more…

…In some early test runs, it is like, Wow, we tested this against the humans, what the humans did, and Falcon is already better than what the humans did.

Veeva’s business consulting and services business is improving with AI

In business consulting, AI makes process design, organizational structure, and change management more critical than ever. In services, AI-enabled offerings that deeply understand our products increase speed and value and drive customer demand. The innovations in our talent processes we started last year are already showing early signs of success that will lead to long-term excellence. We are excited about the future of AI-enabled consulting and services. 

In Commercial Cloud, Ostro is growing rapidly; management is looking to expand Ostro from commercial into medical and other areas; management thinks conversational AI that is 100% compliant is a strategic area; management thinks Ostro can grow far beyond its starting point; Ostro is a recent acquisition by Veeva, and it provides conversational AI for brands to provide patients and doctors with immediate, compliant answers

Ostro is really growing rapidly, both in customer and brand acquisition and in our longer-term product vision as we look to expand Ostro from commercial into medical and other use cases. Precision AI, conversational AI that is 100% compliant, is a very strategic area. We believe Ostro, like Crossix, can grow far beyond its starting point provided we deliver product excellence and customer success for all customers and brands. 

In Commercial Cloud, Crossix had a strong quarter in Measurement and Audiences; Crossix has been winning new customers and expanding with top 20 biopharmas; management thinks AI is a tailwind for Crossix

Crossix also delivered another strong quarter of growth in Measurement and Audiences, strengthening our leadership position with new customer wins and top 20 biopharma expansions. AI creates new and effective forms of digital marketing, which creates new opportunities for Measurement and Audiences. 

Aspen is Veeva’s new horizontal CRM (customer relationship management) product announced in early-August; Aspen is a CRM platform built for AI; Aspen has a few early adopters; Aspen is on track for availability later in 2026; management has very high aspirations for Aspen and thinks the product is addressing a clear market need; it’s very early days for Aspen; management is treating Aspen as a startup inside Veeva; Veeva is currently investing at a small scale into Aspen; Aspen’s users are currently being charged at $50 per user per month, and it comes with overage fees; management is willing to change Aspen’s pricing model based on customer feedback; management thinks Aspen’s advantages over existing CRM products are (1) price predictability, (2) dependability, (3) scalability, and (4) having a better system, data model, and business logic

In early August, we announced Aspen CRM, our next-generation enterprise CRM platform built for AI…

…We are learning from a few early adopters and remain on track for planned availability later this year for a broader group of early customers.  

We are just getting started with Aspen, but I am fully convinced that it addresses a clear market need with disruptive innovation that has real potential for greatness…

… Aspen, I think it is very early…

…This is a startup inside of Veeva. It is moving very rapidly. It is on 90-day plans…

…Your question was about how to size the investment as well. The investment is very small on the Veeva scale. It is not something that Brian, our CFO, notices, really, on the Veeva scale, because you have to keep that very small when you are working with early customers and you are iterating an early product…

…Aspen is taking a different approach that may or may not prove effective. It is to say, well, it is a different approach, a different technical stack, a different approach there, and a different pricing approach. It is just much more simple. You get your productivity, $50 a user a month…

…Then there is, of course, usage overage. Okay, let us say you buy five users. It is $50 a month, and you put a terabyte of data in there for some reason. Well, okay. Well, that is not anything that anybody thought about. There will be overage charges that you will pay monthly on the overage. It is a mix. I would say we are shooting for mostly predictable, because at the end of the day, large businesses would really want mostly predictable. You have to have this escape hatch to say, Yeah, I cannot use unlimited compute, because that does not make sense. Now we are also going to listen to our early customers, and we are a very customer-friendly company, and if there is a better way to do it, we will certainly do that…

…[Question] What are the specific customer problems that you’re aiming to solve with Aspen, versus what some of the existing horizontal CRM platforms struggle with today?

[Answer] One is price. Price being unpredictable, getting out of control. That would be one. The other one would be just dependability of the vendors, that you can really count on the vendor to be on your side. Scalability of the vendor. Sometimes they want something that really works for a small company but can scale up to a very large. Now in the market, you have to pick, like, do I want something that works for a small company, or do I get something that’s too big for me now, but it can scale up? Other things are just like data entry. The existing CRM systems really, if you get into them, okay, they require a heck of a lot of data entry. Most of that with AI doesn’t need to be done anymore. And then I just think there’s this other fundamental thing of better CRM system, and that’s just the details of a fundamentally better data model, better business logic, better just details like how do you handle multi-currency? How do you handle forecasting? How do you handle implementation so that you can get the CRM you want for your company in three months, rather than getting half of what you want in three years?

Veeva’s management does not appear to be willing to squeeze the company’s margins to invest in AI initiatives; 

[Question] As we think about you guys ramping on Falcon, ramping on Aspen, ramping on some Vault agents, Vault AI, should we expect maybe an uptick in terms of sales investments, R&D investments?

[Answer] We are very excited about Falcon and the path that it can be on. Y ou have also seen us over time consistently think about both growth and profitability. It is not different entering a new market like Falcon. Maybe the dynamics of the market are very slightly different, but it is the same overall approach that we are taking there. We scale investment as we scale revenue. There is certainly nothing material that I would call out for this fiscal year. It is all factored into the guidance that we have updated for FY 2027.

Veeva’s management thinks the current costs of AI models is not sustainable and that prices will become cheaper over time

We use the non-deterministic models, the anthropic models, et cetera, when we need to. A lot of this value is going into the agent. Then I believe everybody knows that the cost of these models are going to go down, whether they’re with better hardware or open weight models or et cetera. The current cost of the models is not sustainable, not based on what we’re doing, but based on this notion of what software development is doing. Eating up 50% of the tokens in the world and hundreds of billions of dollars. Somebody’s going to build a better mousetrap for that over time, and that’ll compress the prices.

Veeva’s management is seeing customers want to buy AI solutions from existing vendors, and not from new, unproven vendors

[Question] An emerging trend that we’ve been seeing across large healthcare and life sciences organizations within the industry this year is one in which customers want to embrace AI, but they don’t want to take the risk on a new and unproven entrant that offers AI for only a specific point solution or a niche use case, as they don’t have the time to evaluate hundreds of new vendors. It sounds like they’d rather consume AI from the incumbent platform vendors that they’re already deeply embedded with. One, are you getting this sort of same feedback from your customers?

[Answer] Some months ago, when we first introduced Falcon, I had customers come up to me personally, we were at an event and they said, Oh, thank goodness you’re announcing that because that’s I didn’t want to evaluate all these small vendors. ” Now that you have an offering, that helps me not have to go and look at all these small vendors. Yes, it’s absolutely what customers want.

---

Disclaimer: The Good Investors is the personal investing blog of two simple guys who are passionate about educating Singaporeans about stock market investing. By using this Site, you specifically agree that none of the information provided constitutes financial, investment, or other professional advice. It is only intended to provide education. Speak with a professional before making important decisions about your money, your professional life, or even your personal life. I have a vested interest in Adyen, Ahabet (parent of Google), Amazon (parent of Amazon Web Services), Microsoft, MongoDB, Nu Holdings, Okta, Salesforce, Sea, Tencent, and Veeva Systems. Holdings are subject to change at any time.

Last week, I published The Latest Thoughts From American Technology Companies On AI (2026 Q2). In it, I shared commentary in earnings conference calls for the second quarter of 2026, from the leaders of US-listed technology companies that I follow or have a vested interest in, on the topic of AI and how the technology could impact their industry and the business world writ large. 

A few more technology companies I’m watching hosted earnings conference calls for 2026’s second quarter after I prepared the article. The leaders of these companies also had insights on AI that I think would be useful to share. This is an ongoing series. For the older commentary:

• 2023 Q1 – here and here
• 2023 Q2 – here and here
• 2023 Q3 – here and here
• 2023 Q4 – here and here
• 2024 Q1 – here and here
• 2024 Q2 – here and here
• 2024 Q3 – here and here
• 2024 Q4 – here, here, and here
• 2025 Q1 – here and here
• 2025 Q2 – here and here
• 2025 Q3 – here, here, and here
• 2025 Q4 – here, here, here, and here
• 2026 Q1 – here, here, here, and here

With that, here are the latest commentary, in no particular order:

Airbnb (NASDAQ: ABNB)

Airbnb’s management has rebuilt the company into an AI-native one; management thinks AI is the best thing to happen to Airbnb, as it has allowed the company to reduce the time from concept to launch by as much as 60%, and increase the features and improvements shipped in 2026 H1 by 80% compared to 2025 H1; AI has accelerated improvements in Airbnb across search, sign-up, checkout, and payments, leading to more conversion of traffic into bookings; management has introduced AI-generated listing highlights; management is using AI to surface review highlights; management will introduce AI home comparisons later in 2026; management sees AI as an existential risk to Airbnb but it ended up being a good thing for Airbnb, especially after the hire of CTO Ahmad Al-Dahle, and now management thinks Airbnb is one of the most AI-native companies around outside of the frontier labs and hyperscalers; management thinks the AI-native nature of Airbnb will enable it to enter many new businesses within and outside of travel; management thinks being AI-native will not affect Airbnb’s cost structure; Airbnb does not need major capital investments to become AI-native; management expects Airbnb’s inference cost to be much lower compared to incremental revenue that can be generated; Airbnb has not been tokenmaxxing, as management has been thoughtful about the outputs of the company’s token-usage; in Airbnb’s updated guidance, management expects higher expenses because of AI spending, but the company’s margins are also expected to increase

We’ve rebuilt Airbnb from the ground up to be an AI-native company…

…AI is transforming how we execute and build products. Said simply, AI is the best thing to ever happen to Airbnb. Today, we’re building, testing, and iterating faster than we could just a year ago. Across some of our key initiatives, we’ve reduced the time from concept to launch by as much as 60%. Compared to the same six months last year, we’ve increased the number of features and improvements we shipped this year by nearly 80%. The acceleration from AI allowed us to make hundreds of improvements across Airbnb for hosts and guests…

…I’ve talked in past quarters about Project Y, which is our innovation blueprint, where hundreds of improvements compound over time. AI is accelerating this work across search, sign-up, checkout, and payments. By reducing friction across the guest journey, we are converting more traffic into bookings, and that’s become one of the biggest drivers of our growth. We improved search and discovery, making it easier for guests to find and book the right home, hotel, service, or experience, and it’s meaningfully improving conversion.

We also introduced AI-generated listing highlights so guests can quickly understand the key details about a home. We also launched AI-powered review highlights, surfacing what guest reviews say about a home’s location, amenities, and more. Later this year, we’ll introduce AI home comparison, allowing guests to compare homes side by side before booking…

…Last year, I told our company that AI is an existential risk to us. It was the only existential risk to this company. Now, policy is a risk, but it’s not an existential risk. It’s a risk that we will manage forever. The existential risk to everyone was AI. Is AI good for you? Is AI bad for you? I think the moment of truth happened this year. Moment of truth happened. First, we hired our CTO, Ahmad Al-Dahle. He was the leader of Meta Llama models. He came in, I think we went from a company that was a middle-of-the-pack company for AI to a leader in AI, at least amongst companies that are not frontier labs or hyperscalers. I think we are amongst the most AI-native companies now in all of Silicon Valley. I think because of that, this allows us to go into many new businesses in travel and eventually beyond travel that we might not have been able to go into except for the opportunity that AI affords us…

…[Question] With AI helping drive faster velocity of innovations, I was hoping to get your perspective on how this transition to AI native could impact your product costs and if there are any operational adjustments you’re making to help minimize the impact.

[Answer] It won’t affect us that much… We do not need to make any major capital investments. We are not buying up a whole bunch of GPUs. Second, the inference costs of Airbnb are de minimis relative to the ROI of our business model. Right? We’re not in the business of information where we’re trying to monetize. Our transactions are very high-dollar transactions, if AI can just increase our conversion rate just a little bit, the inference cost is so outweighed by the amount of money we make on that increased ROI. I think that what you’re seeing is the cost of tokens to develop products and the inference costs to run the models pales in comparison to the incremental revenue we generate and the incremental output or throughput we’re seeing…

…We’re not so-called token maxing, which I think is this thing where I think all these CEOs at the beginning of the year have this mandate. “I want to see everyone use AI,” with a vanity of have people use as many tokens as possible. Luckily, I have a great technology leader in Ahmed, we’ve been really, really rigorous and thoughtful about it’s not about how many tokens you use, it’s about the throughput of your product and the quality of your product designs and how much you’re shipping…

…In the updated guidance that we provided, it obviously does assume a material increase in terms of the AI spend over the course of the year. I would note that, yes, we are expanding margins while absorbing that increased cost.

AI is making it easier for property owners to become Airbnb hosts; AI is helping Airbnb improve the pricing information and hosting tools provided to hosts; management thinks Airbnb’s AI tools will for pricing will be very powerful

AI is also making it easier to host. We know that as hosts are more successful when they have the right price, the right insights, and the right tools, and AI is helping us improve all three. We made it easier for hosts to set competitive prices and get more bookings. We also gave hosts more actionable insights to help them improve their listings and increase their earning potential. We’re rolling out AI tools that help new hosts get started faster and better understand their pricing and earning opportunities…

…Most people don’t have a single price. They have different prices for every day, and the best way to price your home, like a hotel, is to have different prices on different days, and for those prices to be dynamically changed. I think that it’s very possible that, right now, hotels have very sophisticated pricing management. They have entire teams of people doing that. I don’t think anyone is going to be better than AI at doing this. I think that our models are going to be very, very powerful, and I hope in the future, hotels can even use that.

Airbnb’s management thinks AI is transforming the company’s customer support; Airbnb’s AI assistant is now available in more than 50 languages; the AI assistant has resolved 45% of issues without a human agent, and at much faster times; management will introduce AI voice support later in 2026; the AI assistant helped Airbnb reduce customer support costs per booking by 16% year-on-year in 2026 Q2; management expects the AI assistant and AI voice support to further reduce Airbnb’s customer support costs per booking

AI is also transforming customer support. Our AI assistant is now available in more than 50 languages. Nearly 45% of issues that start with our AI assistant are now resolved without a human agent, while delivering much faster resolution times. Later this year, we will begin introducing AI voice support, extending the experience to phone call…

…In Q2, customer support costs per booking declined about 16% year-over-year, driven in part by improvements by our AI assistant. We expect those costs to continue to decline as our AI assistant resolves more and more issues, and of course, as we bring it to voice.

Airbnb’s management is able to use AI to understand whether a user wants homes or hotels

We have really, really good personalization, and we know now with our personalization, and really driven by AI, whether someone wants to see just homes, just hotels, or both.

Airbnb’s management started testing AI Search recently on a tiny percentage of the company’s overall traffic; the test results for AI Search have so far been extraordinary; AI Search will initially be turned on only by a toggle, and management expects Airbnb to take months to instill a new habit among users; management sees AI Search as having 4 major natures, namely, (1) the ability for users to search with natural language, (2) users receiving responses in natural language, (3) visual titles, and (4) AI-generated personalised highlights; management envisions AI Search to be conversational, visual, and personalised; management expects AI Search to meaningfully improve Airbnb’s conversion rates

on AI search. Good news, we are beginning to put it in test this month. That test is going to be a very small % of our traffic, and based on those results, we are going to then begin to expand it to more traffic over the course of this year. I just want to point out that the tests that I’ve seen, I think, are extraordinary. They’re really great. That being said, we just have to remember that people come to Airbnb, often, most people, a handful of times a year, and they have an expectation that they see a search box with a location. It’s going to take some time, months and months, to retrain the customer. The way we’re initially going to roll it out is the default is going to still be the core search. Above, you’ll see a toggle. Once you’ve turned the toggle on, you’re going to be able to try the new AI search. We’ll have to see how it converts. I think for people who toggle it on, it’s going to convert very, very well. We don’t want to impose that on everyone…

…AI search, you can actually think about as really three or four major features. One is the search input. I can type in natural language whatever I want. The second thing is it can essentially respond to you in natural language. Rather than just saying, “300 search results,” it can respond to you in natural language. Then the title. The titles could actually be AI-generated, and they can be conversational, as if you’re reading a chatbot, but much more visual. Then you get to the product description page, and the highlights are AI-generated in real time, personalized to you. You go down the page, you have a question, you can ask the PDP through AI. You see the entire journey, not just AI search, is going to be powered by AI. What this will feel like is it’s going to feel as, or almost as conversational as a chatbot. Hopefully less chatty, in fewer words, because we think travel’s more visual. Very personalized. What this will mean is much higher conversion rates.

Airbnb’s management thinks that a lot of AI companies have not figured out how to make money on the consumer side because massive capital expenditure is needed, and inference costs are high

We are coming on the near four-year anniversary in three months of ChatGPT. In the nearly four years, almost all the actual business that’s been generated is on the enterprise…

…Part of the reason why is a lot of companies have not figured out how to make money on the consumer side. Why is this? Because the inference cost is not cheap, and there’s huge capital expenditures.

Arista Networks (NYSE: ANET)

Arista Networks’ Etherlinks switches now has more than 100 cumulative customers, up from just 4-5 in 2024

Our AI fabrics momentum with Etherlink switches now exceeds 100 cumulative customers from the initial 4 to 5 customers I spoke of in 2024.

The maximum possible scale for an AI network depends on the number of tiers and the number of ports; increasing tiers and ports is expensive and eats power; Arista Networks’ 7800 AI Spine allows users to achieve high scale without adding tiers; the 7800 AI spine has an important use case in scale-across; management sees scale-across switching and routing as a $15 billion to $20 billion market by 2030; Arista Networks’ scale-across solution can provide near instantaneous recovery in the event of transient congestion, packet loss or a physical failure of AI clusters independent of their geographical location and management expects this scale-across use case to be 30% of Arista Networks’ overall AI target of at least $3.6 billion in 2026; Arista Networks’ management continues to see 3 types of AI fabrics for the company to participate in, and they are scale-up, scale-out, and scale-across; management thinks competing whitebox solutions are seen more in simple scale-up or scale-out use cases where software requirements are low; Arista Networks’ solutions stand out for massive training and inference clusters; Arista Networks’ scale-across technology is mostly used by AI titans at the moment, along with some neoclouds

The maximum possible scale for an AI network generally depends on two things: the number of tiers in the network and the number of ports per device, often known as Radix. Increasing the tiers and ports is expensive and power hungry. Our customers deploy and often chose the Arista flagship 7800 AI Spine to achieve that high scale without adding additional tiers. 

Scale-Across is an important application. The scarcity of compute capacity, physical space and gigawatts of power mandate that the AI infrastructure must be designed thoughtfully. The Arista 7800 platform continues to be the flagship spine for distributed scale across applications, providing traffic isolation, contextual routing and security. The Scale-Across switching and routing TAM is forecasted to be roughly $15 billion to $20 billion in 2030, and Arista is well poised in this segment. 

Our Scale-Across AI innovations deliver programmable and deterministic routing, SRv6 multi-plane forwarding, multi-tenancy and traffic engineering as well as load balancing across the regions. We are capable of providing near instantaneous recovery in the event of transient congestion, packet loss or a physical failure of AI clusters independent of their geographical location. This Scale-Across use case is expected to be approximately 30% of our overall AI target of at least $3.6 billion in 2026…

…There’s 3 types of AI fabrics that are critical for us to participate in, scale-up, scale-out, and scale-across…

…White box is certainly a tactical solution that we tend to see more in use cases that are simple, scale-up or scale-out where the actual amount of software and system requirements are low. But when you look at traditional network topologies, you can slow down the job completion significantly if you go with a box-by-box approach. And so this — as you rightly point out, the Etherlink system-wide portfolio with all of the features Ken alluded to for reliability, MRC, SRv6, traffic engineering, synchronizing elephant flows, low latency, this massive training and inference does put more pressure on the combination of our hardware and software, and we feel very well recognized and ready to achieve that…

…[Question] Where we are with the scale-across build-outs. Like is it concentrated to maybe a couple of the cloud titans? Or can we see a world where this extends out to some of the Neocloud customers as well?

[Answer] Our scale-across is dominated by the cloud and AI titans. But I see no reason why it wouldn’t apply to — and we’re already seeing it apply to several Neoclouds. So I would say less in the enterprise and definitely more in the Neoclouds and titans.

There are 3 important recent innovations in AI networks, namely, Smart System Upgrade (SSU), Multipath Reliable Connection (MRC), and Segment Routing v.6 (SRv6); SSU is a recent Arista Networks innovation that enables customers to upgrade switch software without any disruption; frequent upgrades of software is important because AI is uncovering security vulnerabilities and creating tools to exploit the weaknesses; maximising XPU utilisation requires MRC, which enables senders to spray a single XPU-to-XPU flow across many paths through an AI fabric; SRv6 enables senders to control which paths an XPU-to-XPU flow will take in MRC; SRv6 is not a new innovation, but using it to load balance an AI fabric is; Arista Networks’ EOS (Extensible Operating System) provides a unified operating system, so it can support SRv6 from scale-out to scale-across; Arista Networks uses explicit SRv6 probes for multipath and multiplane monitoring for reliable accelerator communication

I have never witnessed the combination of rapid innovation and scale deployment that we are seeing in AI networks. I’d like to call your attention to 3 innovations: SSU, MRC and SRv6…

…SSU is Arista’s Smart System Upgrade, the ability to upgrade switch software without any disruption. Frequent upgrades are a hard reality today, especially as AI both uncovers security vulnerabilities and creates tools to exploit them. While many competing systems require a full reboot to address these issues, leading to expensive and disruptive downtime, Arista’s EOS handles these upgrades seamlessly. We ensure our customers stay secure without sacrificing even a single minute of valuable XPU cycles…

…To maximize XPU utilization, you need MRC or Multipath Reliable Connection. See in first-generation AI networks, every packet on an XPU to XPU flow has to take the same path. That means if 2 flows hash to the same link, they both run at half speed. MRC enables senders to spray a single flow across many paths through the fabric, where receivers reassemble any data that arrives out of order, eliminating the performance hit from fabric hash collisions. But how is the sender supposed to control which paths the flow will use? And that’s where the third innovation comes in.

SRv6 or Segment Routing, it’s not new, but using it to load balance an AI fabric, that’s the game changer. The sender tags each packet with a stack of SRv6 segment IDs, dictating the exact path the packet will take. The system then uses real-time congestion signaling to dynamically shift packets away from hotspots. Because Arista EOS provides a single unified operating system, we support this SRv6 intelligence all the way from the scale-out fabric to the long-distance scale across routing. It gives our customers the combination of high quality, top performance and operational simplicity that Arista is known for…

…Arista’s multipath and multiplane monitoring with explicit SRv6 probes ensures that reliable performance for accelerator communication.

Arista Networks’ management sees the semiconductor industry’s supply chain challenges persisting, but the company has made solid progress in hardening its supply chain; management thinks the industry’s supply chain problem will last until 2028; as part of the improvements made to Arista Networks’ supply chain, the company has tripled its purchase commitments to $9.7 billion from a year ago (was $8.9 billion in 2026 Q1); Arista Networks’ purchase commitments are mostly for chips for new products and AI deployments; management has secured Arista Networks’ memory supply for 2026 and for parts of 2027; management has built capacity in a 12-month window for PCBs (printed circuit boards) and optics; management has established a supply chain for liquid cooling for the next generation of AI infrastructure; management has increased Arista Networks’ manufacturing and distribution capacity; management thinks the semiconductor industry is in a constant state of scarcity across 3 dimensions, namely, power, space, and compute, and getting all 3 solved simultaneously is very difficult; the supply chain challenges in the semiconductor industry has not increased management’s visibility with customers

While the industry-wide supply tightness and rising component costs persist, Arista has taken individual and aggressive proactive steps. Arista is making solid progress here in addressing our tight supply chain…

…The industry is going to have a 2-year problem. I don’t think we get out of it as an industry until 2028…

…Arista is leaning in with our increased multiyear purchase commitments, now almost tripling from a year ago at $3.6 billion to approximately $9.7 billion by the end of Q2 2026…

…Our purchase commitments at the end of the quarter were $9.7 billion, up from $8.9 billion at the end of Q1. As mentioned in prior quarters, this expected activity mostly represents purchases for chips related to new products and AI deployments…

…We’ve secured multiyear agreements with leading vendors of strategic components, qualified new suppliers in key areas to limit risk and built out supply chains for next-gen AI technologies. Our capacity has been increased in both manufacturing and distribution, and we’ve negotiated better component delivery terms to drive up both factory efficiency and capital deployment. Relationships with our strategic silicon vendors continue to be strong with really excellent collaboration in both supply chain and technical engagements. Our memory supply has been secured for 2026, and we have extended visibility well into 2027 across DDR4, DDR5 and NAND memory. And importantly, we’ve increased our resiliency through optionality and expanded vendor qualification. For PCBs and optics, we’re now able to build capacity in a 12-month window and have strengthened our engagement and commitments from key suppliers. We’ve improved our lead times and inventory management of thousands of component SKUs, improving subcomponent pipelining and multisourcing and providing increased flexibility with reduced inventory risk. In a new area, we’ve now established a liquid cooling supply chain capable of driving and delivering the next generation of AI infrastructure. This includes cold plate, quick disconnect and tubing vendors with capacity agreements for cutting-edge new AI technology.

And to match our capacity with customer demand, we’ve increased both our manufacturing and distribution capacity. We have now 3 contract manufacturers and 3 distribution facilities, providing geographic diversity in the U.S., in Asia and in Mexico…

…We’re going to be in a constant state of scarcity for power, space and compute. Getting all those 3 to lock in, whether you’re a cloud titan, an AI titan or a Neocloud is going to be very, very difficult…

…[Question] With the tightness in the overall supply environment and your demand that you’re seeing, I’m wondering how far out your visibility with customers is extending?

[Answer] We still have the same kind of 2 quarters of visibility that we referred to.

Arista Networks’ ramp of new products and new use cases in AI is contributing to growth in deferred revenue

Our total deferred revenue balance was approximately $6.9 billion, up from $6.2 billion in the prior quarter. The majority of the deferred revenue balance is product related. Our product deferred revenue increased approximately $600 million sequentially versus last quarter. We remain in a period of ramping our new products, winning new customers and expanding new use cases, including AI. These trends have resulted in increased customer-specific acceptance clauses and an increase in the volatility of our product deferred revenue balances.

Arista Networks’ management has raised its revenue guidance for 2026 to growth of 40% (previous guidance was for growth of 28%), but kept its AI fabrics guidance unchanged at $3.5 billion; management thinks the AI fabrics revenue will increase, but is unsure by how much; Arista Networks’ business performance for 2026 will be gated by supply

Reflecting our strong momentum, we are raising our 2026 fiscal year outlook to 40% revenue growth, equating to approximately $12.6 billion. Within this guide, our 2026 campus revenue goal is at least $1.25 billion and our AI fabrics goal is at least $3.5 billion…

…[Question] You took up the calendar year ’26 revenue guidance substantially from last quarter. But if I remember correctly, you got the AI target unchanged and only touch the campus. Can you kind of explain kind of the thought process there?

[Answer] If you had to ask me whether our AI number or campus number will go up, I think Chantelle, Todd, Ken and I absolutely believe it will. The question is not whether it will go up. The question is what is that number? And that I would like to reserve that $1.1 billion question to, well, it depends on how we ship. If we ship more front-end AI or we ship more WiFi or wired or Etherlink switches or routing. So I’d like to give our customers the priority and Todd’s team the flexibility to ship what we can. And that’s why we’re not holding ourselves to a number…

…Where we are at this time of year, Jayshree and I are guiding based on what we’re confident we can get the supply for. And if the supply was to release a little more, there is an opportunity to do better in the year.

Arista Networks’ management thinks it’s a bad idea for there to be a variety of different technologies for networking in scale-up use cases; management has decided for Arista Networks to commit to an open CPO (co-packaged optics) technology; management thinks the industry will still stick with pluggable optics and copper in the near future, but CPO will appear in 2028 and 2029

Let me take the scale-up use case, which we are less prevalent in. I think there’s very much a philosophy there on copper if you can, optics if you must. So I think you’re going to see a lot of copper in that 2-meter, 3-meter distance, well within a rack, that type of thing and the importance of pluggable optics. But in some cases, there is a number of instances of proprietary implementations of traditional co-packaged optics that’s been floating around. Arista is not a fan of 5 different proprietary implementations. They’re going to have and we’re going to commit and Andy and the team have been working hard at this to really solve one, which is an open CPO.

And we don’t think open CPO is going to happen overnight. But the idea here is to use socketed optical engines, pigtail fibers and allow these modules to be fully pretested. And whether they started on the board or nearby, the idea is to have a truly open interface that can operate with multiple vendors and multiple switch configurations. And sometimes those open CPOs are often called NPO too, [ nearby ] optics, right? So we’re big fans of that. It’s very early stages. It probably comes into examples and trials next year. And probably from our perspective and the industry perspective with all the supply chain shortages, majority of the world will still remain pluggable optics and copper, but there will be some amount of co-packaged optics in 2028 and 2029.

Arista Networks’ EOS (Extensible Operating System) has the lowest security vulnerabilities; management thinks regardless of what model is being used, the network must have robust security; management thinks there’s to be a lot of diversity in agents and models; management thinks the more diversity there is in models and AI infrastructure providers, the better it is for Arista Networks

I just want to acknowledge Ken’s architectural design of EOS and how we have the lowest vulnerabilities and we’re absolutely committed to building that rock-solid foundation…

…In terms of our own approach to models, every day, it’s in the news, there’s a China model, whether it’s a DeepSeek or whatever. But from our standpoint, the network must be robust no matter what the model. There’s going to be a lot of enterprise AI agents. There’s going to be a lot of diversity of models…

…we are extremely excited about the growth and diversity in the industry. The more diversity there is among models and model makers and between AI infrastructure providers, the better for us.

Arista Networks’ management thinks AI workloads will be shifting to inference eventually

It’s going to be training at some point, but it’s going to move to inference.

Arista Networks’ management expects more large customers to become 10%-customers for the company, and join the ranks of  Meta Platforms and Microsoft

[Question] Can you revisit your thoughts on adding a 10%? I think in the past, you’ve talked about 1 or maybe even 2 additional 10% plus customers.

[Answer] we’re going to increase the number by $1 billion or more. And undoubtedly, we remain very committed to our 2 longest partners, Microsoft and Meta. And I fully expect there to be 1, maybe 2 10% customers.

Arista Networks’ management sees the AI accelerators market as currently being dominated by NVIDIA; in the NVIDIA-dominated AI accelerators market, the scale-up networking technology is typically provided by NVIDIA; Arista Networks does better in scale-out or scale-across networking technology with NVIDIA accelerators; management is excited about non-NVIDIA AI accelerators, such as AMD’s MI series and Google’s TPUs; management is working with non-NVIDIA AI accelerators to build both scale-up and scale-out networking technologies; management thinks it will take time for Arista Networks to win more share of networking technologies in NVIDIA AI accelerators, but the wins will happen; management thinks NVIDIA’s Infiniband networking technology is no longer widely discussed

We live in an NVIDIA world. And I think we all can safely say it’s a high percentage of the GPUs we connect to. So there’s really 2 use cases there. One is where NVIDIA provides the full vertical stack and usually, that’s an NVLink. So there’s very little participation from Arista or anybody else’s scale up. And generally, it Arista does better there in the scale-out or scale-across domain.

The second is the non-NVIDIA accelerators, and you’ve heard me talk about our enthusiasm with the MI series from AMD. We’re excited about the Google TPUs. Increasingly, we see that as a formidable training processor. And we’re very excited about the range of inference accelerators as well. And we have a number of partners who are working with — you can imagine, many of our customers are building their own in-house.

So now parsing your question a little bit. In that sector, where it’s non-NVIDIA, Arista will be excited and will be working more closely, both in the scale-up and scale-out, in some cases, to build custom racks with their custom processors so that we can better tune our network with the behavior for their inference or training engine. In the NVIDIA cases, it’s going to take a bit longer. I feel a little bit like 2, 3 years ago when we were talking about InfiniBand and now we don’t mention InfiniBand, but it took 2, 3 years to move to Ethernet. So it will take time to go from a proprietary scale-up that’s been around a long time with NVLink to these other alternatives, even if Ethernet is really good. But I expect us to do much better there.

Arista Networks’ management thinks customers value the company’s EOS (Extensible Operating System) for its operational excellence, AI features, and combination with the NetDL diagnostics layer; with the current scarcity in the supply chain, customers are turning more to Arista Networks’ integrated networking technology that comes with quality hardware and the EOS, as opposed to going with white boxes (generic networking hardware) or blue boxes (Arista hardware with generic software)

The 3 things they value greatly with EOS is operational excellence. They don’t have to put a lot of staff. Most of these Neoclouds don’t have staff. And even if they’re paying a little more for the CapEx, it more than makes up for the operating cost that they would incur. So that’s a huge piece. The second is the AI features themselves. You heard Ken talk about a few of them. There’s a tremendous depth and breadth to EOS that they can’t recreate in a white box or they can figure out how it is. And the third is reliability and vulnerability, which is becoming top and center. You can’t put these things in the middle and have them blow up. So — and our combination of both EOS and NetDL, our diagnostics layer, which plays in both the hardware and software has been very compelling. So we continue to see that while it’s interesting to talk about these open NOSes that more and more customers want either EOS itself in its entirety or a hybrid combination of open NOSes and EOS…

…[Question] You’ve talked in the past about kind of new customers who have come in who have maybe wanted to start down the road of blue box and then discovered with the complexity that they actually need to go down the road of EOS. Can you just talk about kind of the latest trends that you’re seeing there, particularly as you continue to add customers?

[Answer] I think in this in this current scarcity of supply chain and people and needing to deploy AI fast, Arista is really winning out with the system-wide approach on EOS and good hardware. 

Cloudflare (NYSE: NET)

Cloudflare’s management sees the company providing the new kind of cloud that’s needed for an agentic future; management thinks Cloudflare’s Workers Developer Platform is the fastest, most secure, and most cost-effective place to build and deploy agents; management sees Cloudflare occupying the pole position to lead the internet’s agentic AI phase; management is seeing a lot of code from vibe-coding platforms going to Cloudflare; management thinks agents need ephemeral, low-cost places to do their work; management thinks the ideal case is for an organisation’s agents to be running in many different places, and for the network to become the computer; management thinks that if every knowledge worker ran an agent that ran in a container, there would not be enough CPUs to power that; Cloudflare has a container-less version called Isolates that is more lightweight; management is not interested in simply renting out commodity AI compute because that is not an attractive business

It’s clear that the agentic future needs a new kind of cloud. Developers are flocking to Cloudflare because our Workers Developer Platform gives them what they need to build that agentic future. We’re the fastest, we’re the most secure, and the most cost-effective place to build, deploy, and scale agents and the code they generate…

…It’s the same one that has propelled Cloudflare into the pole position to lead the next phase of the Internet in the age of agentic AI…

…Cloudflare Workers is turning out to just be the perfect platform for building agents and agentic workloads. It’s extremely lightweight. You only get charged for when it’s actually doing work. You can spin things up and spin them down very, very quickly. It has become the go-to place for sophisticated developers to be able to launch code…

…If you look at companies like Lovable and Replit and Base44 at Wix and others, that a lot of times that code is actually getting deployed and where the preferred target of that code is going is actually to Cloudflare…

…What I think that we’re seeing is what agents need is the ability to have very ephemeral, low-cost places to create code, do inference, access the network, coalesce information, store some things, and pack all of these things together. It’s not simply the inference that matters. The real key is how do you orchestrate all of those pieces together. In the ideal case, your agent doesn’t run just in one place. It runs in many places. It might be operating in literally different places around the world where it has to access information. It has to get different things. You want, effectively, for the entire network to, as the old Sun saying was, become the computer. The network is the computer…

…I think that’s exactly what agents need. They need the network to be the computer…

…If every knowledge worker on Earth ran an agent that was running in a container, we don’t have enough CPU to actually power that. We’d need to increase the amount of CPU that’s accessed by many orders, many times. What we’ve done, which is a new version, which is much lighter weight, in terms of our sandboxing technology, which we called Isolates, that gives you the ability and the scale to actually deploy and run code in a way that is going to keep up with the demands that agents have…

…If you’re selling what is just commodity compute, if you’re basically letting an AI company use your balance sheet and your credit rating in order to buy servers that are the same as everybody else’s servers, then that’s just not attractive business for us…

…I don’t think we want to get into the business of renting servers, because over time, it’s a commodity business and it’s not very attractive.

A leading digital native media company signed a deal with Cloudflare to combat aggressive scraping, despite competitive pressure to use its incumbent hyperscaler

A leading digital native media company expanded their relationship with Cloudflare, signing a five-year $31.8 million contract for application services and Zero Trust. To combat aggressive scraping and accelerate global performance, this customer chose Cloudflare for our best-in-breed edge capabilities and operational velocity. Despite competitive pressure to consolidate spend with their incumbent hyperscaler, this customer’s long-term commitment is a proof point that when performance and security are non-negotiable, enterprises choose Cloudflare’s unified platform.

A Global 2000 European technology company expanded its relationship with Cloudflare for future AI workloads; the European technology company is using Cloudflare to replace 5 incumbent point solutions, with up to 7 solutions targeted

A Global 2000 European technology company expanded their relationship with Cloudflare, signing a three-year, $11 million contract for application services and Zero Trust with our developer platform seeded for future AI workloads. After years of acquisitions resulted in a fragmented IT footprint, this customer chose Cloudflare to eliminate a stack of five incumbent legacy point solutions with up to seven targeted on their long-term roadmap in favor of our single unified platform as the foundation for their entire organization to build on.

A rapidly-growing generative AI company signed a deal with Cloudflare for the Developer Platform; the generative AI company pulls a lot of videos and images, so its egress fees would have been way too high if it was using hyperscalers’ cloud computing platforms

A rapidly growing generative AI company signed a one-year, $7.5 million pool of funds contract for our developer platform. This customer’s workloads pull an enormous volume of images and video. At that scale, a hyperscaler’s egress tax would break the economics and create vendor lock-in, limiting their choice of inference tools and GPUs. Their engineering team evaluated multiple providers and chose Cloudflare as the only one that pairs a zero egress model with the reliability, scale, and comprehensive capabilities of an enterprise-grade platform. By structuring this as a pool of funds deal, the customer can solve their immediate storage needs while retaining the flexibility to expand across our entire developer platform.

A rapidly-growing technology company in the Asia Pacific region expanded its relationship with Cloudflare after signing a 2-year deal in 2026 Q1; the company chose Cloudflare over a hyperscaler

A rapidly growing technology company in APAC expanded their relationship with Cloudflare, signing a one-year, $4 million pool of funds contract for our Workers Developer Platform. This deal accelerates a powerful partnership, building on an $8.7 million application services contract signed just last quarter. In only one year, this customer has standardized on Cloudflare end-to-end, from application security and Zero Trust to now our developer platform, directing every request through a Cloudflare Worker and using KV and Durable Objects as the routing and tenant configuration layer for their entire platform. They chose Cloudflare over their incumbent hyperscaler to avoid added latency, proving the flywheel of our unified offering.

A Fortune 1000 company expanded its relationship with Cloudflare and is now adopting Workers Developer Platform; the Fortune 1000 company is using Cloudflare to replace multiple products

A Fortune 1,000 technology company expanded their relationship with Cloudflare, signing an 18-month, $15.9 million contract for application services and our Workers Developer Platform. This customer serves hundreds of thousands of businesses, which requires an architecture that can act as their global front door for security and performance without adding latency. By standardizing on Cloudflare over legacy alternatives, they eliminated multi-product complexity and secured long-term operational predictability as they build an AI-first customer platform.

A leading technology company signed a deal with Cloudflare for the Workers Developer Platform; the leading technology company needed an elastic, secure container infrastructure for its agentic workloads, and chose Cloudflare over hyperscalers

A leading technology company expanded their relationship with Cloudflare, signing a one-year, $6 million pool of funds contract for our Workers Developer Platform. As this customer scales their new AI agent capabilities, they needed an elastic, secure container infrastructure that could scale with their rapid growth and ship new capabilities in weeks, not quarters. They chose to build on Cloudflare over legacy hyperscalers and point solution competitors because of our built-in threat intelligence that actively prevents compute abuse, rapid pace of innovation, and the ability to deliver FedRAMP compliance. This win also shows how the most sophisticated AI builders are increasingly selecting Cloudflare as the agent cloud of the future.

In 2026 Q2, Cloudflare’s management saw agentic traffic on the internet exceed those of human users for the first time; management is seeing unabated growth in requests from AI agents; management sees the Internet shifting towards machine-to-machine traffic, and thinks Cloudflare is well-positioned for the paradigm shift; management recently announced Monetization Gateway, which allows customers to sell any resource behind Cloudflare; management thinks Monetization Gateway will empower the next set of business models over the Internet; management recently announced Wallets, which allow buyers to pay autonomously through agents; management recently announced cloudflare.pay, which allow merchants and buyers to identify themselves with AI agents; management recently announced a research pilot with OpenAI to figure out a sustainable ecosystem for content creators and AI companies; management has more to announce in the coming months on the sustainable ecosystem; management initially thought that agentic traffic would surpass human traffic only in 2027, but the event happened faster than expected; management expects agentic traffic to be 1000x higher than human traffic in 5 years, meaning human traffic will become a rounding error; management expects some of the agentic traffic to be malicious in nature, such as hackers or AI companies trying to take content; in the case of AI companies taking content, Cloudflare will block such malicious traffic; management recognises that some companies will want traffic from AI companies, so Cloudflare has to be more efficient at supporting this; management is seeing that the give to get of agentic traffic is different with human traffic; management expects agents to pay only fractions of a penny for every request, but there are still unanswered questions on who foots the various bills for the traffic; management thinks the business model of the internet for the next 27 years will be very different from the Google-defined model of the past 27 years; Cloudflare handles 500 million requests per second on its network and management thinks 1%-10% of the requests could be monetised through a micro transaction; the amount of micro transactions to be processed is many orders of magnitude higher than what traditional payment networks can handle; management is thinking of even letting free users of Cloudflare monetise through micro transactions, because this would really accelerate Cloudflare’s business

For the first time in human history, in Q2, more than 50% of the traffic flowing across Cloudflare’s network was not human. The number of requests on our network from AI agents continues to grow unabated. With the web shifting from human-driven browsing to AI answer engines and agent-driven commerce, we are witnessing a fundamental rewrite of the Internet for machine to machine traffic. Cloudflare is positioned at the center of this paradigm shift, building the scalable infrastructure, the controls, the developer tools, and the payment rails to power the agentic Internet…

…During these, we unveiled the key building blocks for a two-sided agentic marketplace. Monetization Gateway allows our customers to sell any resource behind Cloudflare, whether it’s a web page, an API, a dataset, or an MCP tool. This will empower new business models that will define the next generation of the Internet. In addition, we announced Wallets, which will offer a way for buyers to pay autonomously through their agents, and cloudflare.pay, which will provide merchants and buyers an agent-friendly means to identify themselves and establish trust. Not only are we building the foundational elements for agentic commerce to succeed, we also believe AI companies and content owners should thrive together. That’s why we recently announced a first-of-its-kind research pilot with OpenAI that we believe may help pave the way to a sustainable ecosystem of content creators and AI companies. Over the coming months, we’ll announce more ways that AI companies, content creators, and businesses large and small can thrive together. The business model of the Internet is changing, and there is no company better positioned to define its future than Cloudflare…

…I was asked in the end of 2025, in November of 2025, when I thought that non-human traffic would pass human traffic. We pulled all the data, we ran all the numbers, and we were pretty confident that it was going to be the second half of 2027. I was asked the same question again in March of 2026, and we did the same exercise, and we were surprised to see that it had moved up, that it would cross in the first half of 2027. I was quite surprised when in May of this year, our team came to me and said, “You won’t believe it, but non-human traffic has now passed human traffic online.” To give you a sense of how this trend is playing out, and with the big caveat that I have called it wrong at every point along the way, if the current trends continue, we think in five years, non-human traffic will be as much as 1,000 times as much as human traffic. In other words, humans will be a rounding error on the Internet, not because human traffic goes down, but that’s just how fast we’re seeing non-human traffic grow…

…Some of that non-human traffic, it’s malicious. That could be malicious like it’s hackers or bad guys. It could also be it’s malicious from the perspective of a particular customer’s business model, where it’s traffic that is maybe an AI company trying to take the content from a media company that relies on advertising. In those cases, we block that traffic, and we don’t charge the customers anything more for blocking that traffic because we think that that’s the right thing for us to be doing and delivering, and that’s part of being a security company. At the same time, though, there are some people who want that traffic, and so we’re doing everything we can not only to serve that, but to make it as efficient as possible to serve it. If we’re going to have 1,000 times as much traffic online, we’ve got to get a lot more efficient, and companies like Cloudflare are critical to be able to support that for customers, whether they’re large and small…

…It’s clear that as agents are accessing all of these sites and the volume that they’re accessing them on, the sort of give to get that you have with human traffic is different…

…Things like Cloudflare.pay, that’s us setting the foundation to be able to say, how do we charge agents some, again, what will be a very, very small fee, fractions of a penny for every request that goes through, but for the requests that pass through that traffic. Somebody has to pay for the bandwidth, somebody has to pay for the server, somebody has to pay for the people doing the work to create the content.

…I think that the business model of the Internet for the last 27 years has been largely defined by advertising and really defined by Google. I think the business model of the next 27 years of the Internet is going to be very different…

…We handle, let’s say, about half a billion requests per second through Cloudflare’s network. We roughly estimate that somewhere between 1% and 10% of those you could monetize through some sort of a micro transaction. Again, these would be tiny fractions of pennies. That means that you, day one on launching something like this, you’d need to be able to support, call it 10 million financial transactions per second, and be able to scale up to call it 100 million financial transactions per second. To give you some sense, Visa, and again, these are from memory, but Visa, which is the largest payments network in the world, at peak during the holidays, handles about 20,000 transactions per second. You have to build something that’s three orders of magnitude bigger than Visa in order to pull this off…

…One way to think of this is a lot of that is our free customers. What if we made it less than free? What if being part of Cloudflare, we actually sent you money for being part of us because we were generating through a series of micro-transactions, largely serving agents. I think that then just continues to accelerate the flywheel across all of our business.

More than 80% of the major AI companies are Cloudflare customers

Over 80% of the major AI companies are Cloudflare customers and rely on us.

Cloudflare’s management is hearing from customers that their key concern with deploying AI is security; the lack of Zero Trust policies for agents from security vendors is causing organisations to cancel deals

The number one thing that’s causing our phone to ring from big companies is them saying, “Listen, we know we have to do AI, but we need to do it more securely.”…

…I was just in London a few weeks ago, meeting with a large government agency there that was well down the track with one of the sort of first generation Zero Trust companies to implement that across a big chunk of the U.K. government. We sort of started talking about agents and what their plan was and how they were thinking about it. Very quickly, it became clear that the vendor that they were considering really hadn’t thought about this, whereas it’s been core because of the fact we have a developer platform to how we did things. They literally canceled the RFP and are now reevaluating this with a sort of agents first approach.

Cloudflare’s management is seeing the company’s customers increasingly move their business models away from a typical SaaS model into one with consumption-based structures

As we said at Investor Day, their business model is also going to evolve. It’s moving away from a purely ratable SaaS model towards a much more diversified mix of pool of funds, consumption-based structures, and also what we call T-shirt sizing. As the business accelerates, that also leads to more and more customers burning through their T-shirt sizing faster. Pool of funds are getting consumed faster and getting renewed.

Cloudflare’s management thinks the GPU utilisation of the hyperscalers are really low; management thinks Cloudflare has been able to get as much as 10x the utilisation of every capex dollar; management sees Cloudflare as being in  

We know what the typical kind of inference loads are at the hyperscalers, and their GPU utilization. It’s super low, and that’s not the hyperscalers’ fault. It’s that what they’re selling is just a box, and it’s up to the customers in order to actually maximize the Customers I think don’t have the diversity of traffic or the ability to really schedule things or get the highest possible utilization…

…If we can, and in some cases, get 10 times as much utilization out of every CapEx dollar. 

Coupang (NYSE: CPNG)

Coupang’s management expects the company’s margin to continue expanding because AI improves discovery, personalization and service for customers, and productivity and cost to serve for Coupang

The long-term margin drivers keep compounding. Automation continues to improve productivity across our fulfillment and logistics network, and margin-accretive offerings like advertising and FLC are still early in their scale. And AI raises the ceiling on both. We think of AI as a multiplier and what it multiplies is a set of assets we’ve been building for 15 years, the physical network, operating data from billions of orders picked, packed and delivered and direct relationships with tens of millions of customers. Applied to the customer experience, AI improves discovery, personalization and service. Applied to operations, it compounds productivity and lowers the cost to serve. And applied to margin-accretive offerings, it raises the returns for the merchants and brands who are using them, which expands the addressable opportunity itself. 

Coupang’s management thinks that the winning experience in agentic commerce has yet to emerge; Coupang is exploring agentic commerce; management thinks that Coupang is in the best position to win, whatever form agentic commerce takes

On the agentic AI part that you’ve brought up specifically, we think this is still a work in progress. We think the industry’s direction or the — it’s not clear that the winning experience has emerged but we’re investing. We are investing in teams and the research, as you mentioned, to explore it while being thoughtful about it and investing with the same discipline that we do and all the other initiatives that we have on the exploration front.

Whatever form agentic shopping takes, we believe we’ll be in the best position to provide the winning experience which we believe will combine AI with all the other aspects of customer experience to provide a complete and seamless buying experience that customers trust. And to build a complete and seamless buying experience, you need more than AI. AI is one input, but there are many other assets that will be part of it. And we believe we should be in a position to provide the best of all worlds.on that front. 

Datadog (NASDAQ: DDOG)

Datadog’s management is seeing its base of customers, from startups to large enterprises, adopt AI; the customers’ adoption of AI is accelerating their usage of the cloud and Datadog; Datadog has more than 750 AI customers in 2026 Q2; all 10 of the largest AI companies are Datadog customers; management is seeing AI activity growing across its non-AI customer base; the number of MCP tool calls has quadrupled sequentially again in 2026 Q2, and is up 22x from 2025 Q4; the 750 AI customers includes AI startups and hyperscalers; 31 customers in the AI native cohort now spend more than $1 million annually (was 22 in 2026 Q1), with 8 spending more than $10 million annually (was 5 in 2026 Q1); the hyperscalers are different companies from the AI labs mentioned in Point 14

Our broad base of customers, from the most nimble startups to the largest and most established enterprises, are all adopting AI. We think this is accelerating their usage of cloud and modern technologies, as well as their usage of the Datadog platform to observe, secure, and act on their cloud and AI workloads…

…As of Q2, over 750 AI customers use Datadog to monitor and improve their tech stacks. When we look at the largest companies driving AI, all 10 of the top 10 AI leaders are Datadog customers. Beyond AI natives, we see AI activity growing across our broader customer base. We are also seeing signs of rapid growth in agentic activity with a number of MCP tool calls quadrupling again quarter-over-quarter and growing more than 22x when compared to Q4 2025…

…This 750-strong customer group includes a broad range of AI startups as it has in the past, but now also includes hyperscalers using Datadog for in-house AI labs. In Q2, this includes 31 customers spending more than $1 million annually, of which eight customers spent more than $10 million annually…

…[Question] If I’m thinking about the two seven-figure AI labs that you landed this quarter, and then going to David’s commentary around winning some of these in-house AI labs with the hyperscalers, are those one and the same here?

[Answer] These are different customers. The ones we mentioned on the new lands are new lands.

Datadog’s management recently expanded Bits AI to automate the DevOps loop; for the DevOps loop, Bits AI can now create and maintain monitors, identify root causes quickly and recommend and implement fixes, follow safety guardrails, continuously learn, and detect symptomatic behaviors early; management recently announced Bits AI Products to improve the development loop; for the development loop, Bits AI can now analyse code changes and run end-to-end checks, generate code fixes, and automate synthetic test generation and maintenance; Bits Security Analyst now works with Datadog Cloud SIEMs (Security Information and Event Management), and users can gain insights no matter which SIEM they use; management is not worried about Bits AI using automation to reduce the volume of activity that drives Datadog consumption, because they think if Bits AI can provide great value to customers, it will yield a great outcome for the company; customers who use Bits AI tend to use more of Datadog’s products; management has significantly expanded the use cases of Bits AI and the new use cases are seeing a lot of adoption; management has a new model with AI credits for Bits AI; management thinks Bits AI could eventually become an AI SOC (Security Operations Center) platform; see Point 14 for AI labs using Bits AI

We expanded Bits AI to accelerate and automate the DevOps loop. This is the loop that goes from detection to investigation to remediation that engineers go through each time something breaks. At DASH, we announced a lot of new Bits capabilities for the DevOps loop. Bits can now create and maintain monitors, identify root causes within minutes of a negative signal, recommend and implement fixes, follow guardrails to add safety and controls, continuously learn and improve from prior incidents, and detect symptomatic behaviors early to repair infrastructure issues before they escalate…

…We announced Bits AI products to address the development loop. This is the loop that goes from coding to delivery to evaluation that developers navigate to get code to production. For this loop, Bits Release now acts as an AI release validation agent, analyzing the impact of code changes, running end-to-end checks, and verifying production rollouts. Bits Code generates code fixes, guaranteeing every fix and reproduction behavior, Bits Testing also automates synthetic test generation and maintenance…

…We expanded Bits Security Analyst to run on known Datadog Cloud SIEMs so customers can benefit from the smarts and the learnings of a broad data set regardless of which SIEM they deploy…

…[Question] Just thinking that Bits AI is going to be increasingly automating activities that historically has created observability workflows, I’m curious and really wonder, how do you ensure that greater automation that might be driven by Bits AI doesn’t eventually reduce the volume of activity that traditionally drove Datadog consumption?

[Answer] If we provide more value, as I was saying earlier in the call, we sell more software by helping customers make more money or save money or both. I think if we can automate more and let them do more, we’ll provide more value. That’s as simple as that. I think the future of observability is not just observing, it’s fixing. It’s not waking up people in the middle of the night because something broke, but fixing it for them. It’s not letting people do damage control on a security incident because an attacker is in. It’s preventing the attacker from getting in to start with by auto-remediating issues. We’re very busy building all of that. We’re super confident that this will yield great business outcomes for us in the end. That’s what we see from customers in the market. When they use Bits AI, they use more of our product. They deploy more of it. They create more dashboards and alerts and everything else. They have more users inside of our product. It’s not a zero sum game…

…Bits AI used to be fairly specific. It used to be dedicated to alerts. Like, Bits AI would pick up an alert and would run an investigation for you. Now the surface of contact is a lot wider with the customer. Bits AI, you can access it through chat. You can, of course, still do the investigations, and we’ve done quite a bit more there. You can have Bits AI manage your monitoring and manage your detection for you. You can have it code for you. You can have it generate managed tests. There’s all sorts of different use cases that we built into it that broaden the surface of contact, and we see a lot of adoption across all of those different areas. We also are changing the way we package it. We have a new model with AI credits that we’re rolling out just because the surface of contact is so much wider now than the specific feature…

…[Question] I want to follow up on the questions around Bits, which sound super interesting. I guess longer term, as you try to push deeper also into the security side of things, could this evolve into a broader AI SOC automation kind of platform?

[Answer] That’s definitely, we’re taking moves towards that, right? Initially we built the SIEM first for that, then we built the agent into the SIEM. Our Bits Security Analyst. Now we’ve actually separated the agent from our SIEM so customers can use it with other SIEMs. We do that because the agent performs just so well, and it’s been such a differentiator when we pitch the SIEM that we think we’re limiting our sales market-wise if we just go after customers that want to re-platform their SIEM, and it can have a much broader appeal as an AI SOC.

Datadog’s management recently expanded the products under its Datadog for AI bucket; Data Observability enables companies to trust the data being used by AI; Bits Data Analysis answers business questions with data content; Agent Console provides visibility into agentic use; Agent Observability identifies agentic quality and cost issues; Bits Evals automates repetitive parts of agentic development

We expanded Datadog for AI, our products that observe, secure, and optimize the AI stack from end to end. Data Observability enables companies to trust the data being used by AI with lineage quality monitoring and jobs monitoring. Bits Data Analysis uses a rich data context to accurately answer business questions. Agent Console provides visibility into AI agent usage, cost, and effectiveness. In Agent Observability, our patterns capability automatically clusters user interactions into behavior groups to identify quality or cost issues. Bits Evals handles the repetitive parts of the agent development loop in order to improve the outcomes of agents.

Datadog has Bits Database Optimizer within Database Monitoring to simulates and evaluates the impact of AI-generated changes; Datadog has Infinite Cardinality Metrics within Custom Metrics Data, which allow users to answer complex questions without incurring extra costs; Infinite Cardinality Metrics solves a long-standing source of frustration with customers, where they unpredictably get higher bills because they send more data or more fine-grained data; Infinite Cardinality Metrics has gotten great feedback so far, but it’s still early

Within Database Monitoring, Bits Database Optimizer now automatically simulates and evaluates the impact of AI-generated changes in order to optimize slow queries… 

…For custom metrics data, we introduced Infinite Cardinality Metrics, which allow our users to answer arbitrarily complex questions as they generate larger amounts of data with AI agents without incurring any extra costs…

…In terms of Infinite Cardinality, I would say it’s been one of the longest-standing source of frustration for customers, when sometimes they send more data or they send more fine-grained tags with their data, and they get some unpredictability on the bills because of that, because it increases the cardinality of the data we’re getting. We’ve solved that from a technical perspective and from a commercial perspective by packaging our metrics a little bit differently. We think it’s particularly important and relevant as customers are building more applications with AI and as they want to send basically more tags and more information and ask more complex questions and get more fine-grained answers to those questions. That fits well within their plans, basically. We’ve got great feedback on that so far, but it’s still early.

Datadog’s management recently launched products to protect users against AI-powered attacks; AI Guard Agent Discovery finds and maps all known and unknown custom agents; AI Guard For Custom uses real-time observability data to block attacks; AI Guard For Coding Agents blocks malicious skills and packages in code; management thinks that cybersecurity in the AI age requires a complete change in how security products work, where the security needs to be a lot closer to the application and infrastructure, and this plays into Datadog’s strengths; management is now seeing new classes of security issues popping up weekly; management is building these insights into Datadog’s security product

We launched a number of innovations to secure the AI stack and defend against a new class of AI-powered attacks. AI Guard Agent Discovery finds and maps every known and unknown custom agent so security teams can see what is protected and what is not. AI Guard For Custom Agents provides runtime protections to block attacks that can only be detected with real-time observability data. AI Guard For Coding Agents applies the same deep observability to block malicious skills and packages in code. We also announced Runtime Prioritization Engine to cut vulnerability noise by over 95%…

…There’s a complete switch in the way the security products need to work. You can’t wait basically for putting humans in the loop. You can’t have the typical path when you have 12 or 15 different products that are going to aggregate signal, then you put that signal into a system to aggregate, to prioritize them for humans, then humans review them when they can. You need to integrate everything a lot more. You need to operate a lot closer to the application and to the infrastructure. You need to have AI agents solve the issues first. It’s a complete reveal for most of the industry, I think it plays into our approach, which is to have an integrated platform and have all of the different data streams come directly from observability straight into the security agent, and have all that be integrated from end to end. Obviously, this is a field that’s moving very fast. We see new classes of issues pretty much every week at this point. We are quite busy building that up, we think it displays into our strength and into where we are basically already are, and we’re building for our security product.

Example of 7-figure land deals with 2 neuro AI labs; the AI labs will use Datadog for visibility across their training infrastructure so they can train their models faster; the AI labs are using Bits AI to rapidly build their observability stack; the AI labs are different companies from the hyperscalers mentioned in Point 2; the AI labs are using Datadog for observing the training of their models; AI model training is a new business area for Datadog

We landed seven-figure annualized deals with two neuro labs. These AI labs are rapidly scaling their AI model training workloads and preparing for major product launches. By deploying observability using Datadog, they gain visibility across their training infrastructure and GPU fleets and can iterate faster on their AI models. They are also using Bits AI to rapidly build monitors, dashboards, and alerts for deep observability context…

…[Question] If I’m thinking about the two seven-figure AI labs that you landed this quarter, and then going to David’s commentary around winning some of these in-house AI labs with the hyperscalers, are those one and the same here?

[Answer] These are different customers. The ones we mentioned on the new lands are new lands. These are companies that didn’t exist a few years ago. What’s interesting about them on the use case there is that very often we land customers when they go into production, they release products, and they start serving their customers. In this case, these are customers we’re getting as they are training models, they’re using us to observe and improve and optimize the training of the models. That’s an exciting new area that was not really a business area for us a couple of years ago.

Example of a 7-figure expansion deal with a large health insurance company; the health insurer is using Datadog to protect PII (personally identifiable information) across dozens of business units; the health insurer is using Bits AI Investigation to speed up incident resolution and reduce expensive escalations; the health insurer is expanding to 19 Datadog products

We signed a seven-figure annualized expansion for an eight-figure annualized deal with a Fortune 100 health insurance company. This customer’s biggest pain point is to deliver great experience to their members throughout their care while protecting PII across dozens of business units. Datadog’s HIPAA compliance and PII handling in RUM, Log Management, and Cloud SIEM allowed us to differentiate and win over competitive solutions. Bits AI investigation is already speeding up incident resolution and reducing expensive escalations. This customer will expand to 19 Datadog products.

Example of a 9-figure renewal deal with a leading AI company (most likely referring to OpenAI); the AI company is a long-time, and very large, customer of Datadog; the AI company is using 17 Datadog products for visibility on production workloads at very large scale; the AI company will have a user reduction starting in 2026 Q3, which management has incorporated into guidance; management cannot share much about the reduction in usage; the renewal with the AI company appears to cover many products it was previously using

We signed a nine-figure renewal with a leading AI company. This longtime, very large customer uses 17 Datadog products to enable unified visibility on production workloads at a very large scale, albeit with a user reduction starting in Q3, which we considered in our guidance…

…We don’t want to comment too much on any specific customer, because we also don’t really control what’s happening with any specific customer…

…It’s a longtime customer who uses many of our products, but there’s not a lot more we can share…

…I think there’s a lot of continuity in that renewal. It’s one way to put it.

Datadog’s management continues to believe that digital transformation, cloud migration, and AI adoption are long-term growth drivers of the company’s business; AI is already a tailwind for Datadog because it drives cloud consumption and thus more use of Datadog’s platform (see Point 2 for Datadog’s AI customers, and adoption of AI among customers); next-generation AI is introducing new complexity and observability challenges and Datadog is solving these problems with its Datadog for AI products; Datadog’s access to large volume of data has enabled management to build the second version of Toto, Datadog’s foundational model for time series forecasting; Toto version 2 has state-of-the-art performance on key benchmarks; Toto version 2 has true scalability for time series models, and management is targeting the same improvement path language models have followed since 2020; management is working on world models and larger dedicated models to power Bits AI; Datadog has acquired Adaptive ML to accelerate Datadog’s work in Toto, and larger models; management thinks inference will eventually be the dominant AI workload and that there’s opportunity for Datadog at every layer of the inference stack; Datadog’s GPU Monitoring and Agent Monitoring products are doing well; management thinks the AI-related focus of customers will change over time, as customers were busy prototyping with AI in 2025 and the focus in the last few months have switched to AI costs; Datadog just had 2 new AI labs use the company’s products for observing AI training and this follows on from 2026 Q1 when Datadog landed a hyperscaler customer who wanted to use Datadog for observing AI training; management is unsure if the agentic-led renaissance of CPU usage (the CPU-to-GPU ratio is much higher for agentic use cases than for model training) has led to the acceleration seen in the consumption of Datadog’s infrastructure products; management has seen an explosive in usage of Datadog’s AI-related monitoring products in recent quarters, and the usage is coming from both non-AI and AI-natives

There is no change to our overall view that digital transformation and cloud migration are long-term secular growth drivers for our business. We now have an additional growth driver with AI as we help our customers deliver value with this transformative new technology. We are tremendously excited about our opportunities in AI…

…AI is a tailwind for Datadog today as cloud consumption grows and drives more use of our platform…

…Next-gen AI introduces new complexity and observability challenges. We are addressing this with what we call Datadog for AI to observe and secure the AI stack from end to end. This includes GPU Monitoring, Agent Observability, Agent Console, Data Observability, AI Guard, and many other products…

…Our AI research team and our large volume of rich data using critical workflows enable us to conduct groundbreaking research. We have shown some of our work already with the second version of our time series model, Toto, in May. Toto version 2 was exciting for two reasons. First, we’ve shown it to be state-of-the-art on key benchmarks. More importantly, we’ve demonstrated for the first time true scalability for time series models, allowing us to target the same improvement path language models have followed since 2020. Now beyond Toto, we are working on larger and more ambitious dedicated models, post-training models to power Bits AI and bringing other modalities beyond time series data into world models that we think can lead to a step change in capabilities for our customers. We plan to accelerate these research efforts with the acquisitions of Adaptive ML, which will close in June…

…There’s opportunity at every layer of the stack in inference. We do think at the end of the day, inference will be the dominant workload. That anytime you train, you probably will want to infer more than you train, as a rule of thumb. We see opportunity at the low level, when it comes to the infrastructure, the GPUs, and the consumption you have there. There’s opportunities at the very top end, when you measure what the agents are doing and whether you’re getting the right outcomes or whether you’re getting the right alignment. There’s opportunities at every layer in between, just looking at the LLM itself, just looking at the tool calls and the applications that are being called by the agents…

…We mentioned our GPU Monitoring product is actually getting quite a bit of usage in a number of neuro labs and very AI-first types of customers. We’re also seeing an explosion of volume in our agent monitoring product, we’re well-positioned there. We think this market is going to change quite a bit. The preoccupations of customers, they also change over time. For example, last year, our customers were mostly trying to validate correctness and validate that they were getting some form of outcome that it could then scale up. I would say three to six months ago the focus has moved quite a bit towards cost. Customers were spending a lot on AI, and they were wondering how to optimize cost…

…When we have a concern with customers, that’s the one thing they kept mentioning is, “Hey, how can you help me rein in my AI costs? This is growing very fast, and I don’t have any control on it, and I don’t know whether I’m reaching the right outcomes with that.”…

…[Question] If I’m thinking about the two seven-figure AI labs that you landed this quarter, and then going to David’s commentary around winning some of these in-house AI labs with the hyperscalers, are those one and the same here?

[Answer] These are different customers. The ones we mentioned on the new lands are new lands. These are companies that didn’t exist a few years ago. What’s interesting about them on the use case there is that very often we land customers when they go into production, they release products, and they start serving their customers. In this case, these are customers we’re getting as they are training models, they’re using us to observe and improve and optimize the training of the models. That’s an exciting new area that was not really a business area for us a couple of years ago. We’ve seen a number of new proof points around that. In addition to that, we’ve mentioned in previous calls, we’ve also landed the AI Lab or super intelligence labs of a number of hyperscalers. I would say the workloads are similar in that it’s largely training of the models, the customers are a bit different…

…[Question] CPUs have had a renaissance lately driven by agentic AI. Would be great to hear your thoughts on this topic if it can be an incremental growth driver for your infrastructure monitoring.

[Answer] We do see an acceleration of consumption of our infrastructure products in general. At a high level, we do see that across the customer base. I don’t know that if we see specifically the CPUs that get attached to GPUs in the new build-out. I think a lot of it has more to do with the fact that the AI agents are largely spending a good amount of their time, like sometimes the majority of their time, coding tools. Tools are just applications that already existed, and those applications typically run on CPUs, so we see quite a bit of that…

…[Question] Did you see rising demand for the AI monitoring tool, particularly with open source tools being deployed across enterprises?

[Answer] There used to be very little volume a year ago. It started growing quite a bit into the second half of last year. Now it’s been very rapidly accelerating over the past couple of quarters. We’ve seen an explosion, basically, of the volume we’re getting there. We get more usage from different kinds of companies, so we definitely see that. We see it also across traditional companies and some more recent AI natives. We see a little bit of both.

Datadog’s management sees the company helping customers save a lot of money on AI initiatives

What we do for our customers today, especially as they keep adopting AI, is we help them save a lot of the money they would spend on building, running operations or running AI agents.

Datadog’s management thinks it’s great for enterprises that there are many model options to choose from; management thinks a multi-model world is great for Datadog because having options (1) leads to more complexity for enterprises, and Datadog helps customers deal with complexity, and (2) means customers are likely to do a lot more training of AI models on their own, and observing AI training is a new opportunity for Datadog; management has long held the view that AI will be a multi-model environment

[Question] I’d love to just get your thoughts on the impact of diversification of AI model usage in your customers and what you’re seeing there.

[Answer] We think it’s great. There’s a lot more options for customers to choose from in general. That opens up a lot of doors and opportunities for them. It also creates a lot of complexity, and we’re here to help deal with that complexity. For us, these are great opportunities. By the way, we’ve had that thesis since the early days of AI that we would not just end up with one or two big AI companies and everybody using them, the same way we didn’t just end up with one or two big cloud companies and everybody just using software from them. The ecosystems are very rich. There are lots of providers. There are very large providers. There are smaller providers, and everything in between… We think the same is going to happen in AI. We think also that the multiplication of models, and open source models in particular, opens the door to customers doing a lot more training on their own. That’s a new market for us.

MercadoLibre (NASDAQ: MELI)

Mercado Ads’ AI Advisor reaches sellers on Whatsapp and can automatically analyse campaigns and deliver budget and ROA (return on advertising) recommendations; AI Advisor now reaches tens of thousands of sellers per month, and 1/3 of sellers engaging with AI Advisor makes a change to a campaign; the number of sellers using Mercado Ads’ AI-powered budget orchestrator was up 63% sequentially in 2026 Q2; Mercado Ads has a new AI-powered search architecture that surfaces more relevant ads and increase click-through-rates

Mercado Ads’ AI Advisor reaches sellers on WhatsApp, analyzes campaigns in real time and delivers fully automated budget and ROAS recommendations with zero human intervention. It has scaled from small pilots to tens of thousands of sellers per month, and one in three sellers who engage with it go on to make a change to a campaign directly within the conversation. The number of sellers adopting our AI-powered budget orchestrator, which helps sellers manage spend across multiple campaigns, grew 63% QoQ in Q2’26. These initiatives lower barriers to entry, and enable sellers to invest more in advertising. We also improved how we rank ads on product pages, and our new AI-powered search architecture helps us to surface more relevant ads and increase click-through-rates. 

Handwritten code is now the exception at MercadoLibre; in 2026 Q2, MercadoLibre’s code submissions were up 110% year-on-year, merged code was up 100%, deployments were up 75%, and rollbacks fell; agents at MercadoLibre have autonomously reviewed more than 0.5 million code submissions and migrated 9,000 services; AI agents are accelerating MercadoLibre’s product roadmap; MercadoLibre has 20,000 developers who are using AI

Today, human-written code has become the exception, and productivity has soared as a result. In Q2’26, code submissions were up 110% YoY, merged code was up more than 100%, and deployments were up nearly 75%, all while rollbacks fell YoY. Agents are doing genuinely autonomous work, having independently reviewed more than half a million code submissions and migrated roughly 9,000 services to newer platforms over the past year. This is a structural gain in productivity that is accelerating our product roadmap…

…We have 20,000 developers that are using AI. 

In 2026 Q2, MercadoLibre’s management completed the rollout of its AI-powered marketplace search architecture across its largest sites; management used different models for different countries, but they are delivering comparable uplifts in conversions and click-through rates at lower cost; even when MercadoLibre used more expensive 3rd-party models, the uplifts in conversion and ads click-through-rate more than offset the cost of using the models; sellers representing 50% of MercadoLibre’s GMV are using Seller Assistant, and DAUs (daily active users) were up 21% month-on-month in June 2026; the Seller Assistant resolved more than half of the requests in 2026 Q2 without any human intervention; MercadoLibre has a Shopping Assistant but it’s still under early testing, although management is very excited about the early results; MercadoLibre used to have 10,000 customer service reps 4 years ago, but today the reps have declined to 7,000 even though the business is 3x larger, because 90% of customer service interactions are now done without human intervention

In Q2’26, we completed the rollout of our AI-powered marketplace search architecture across our five largest sites. We deployed this architecture with different models in certain countries and they are achieving comparable results – uplifts in conversion and click-through-rates – at a lower cost. In all cases – even where we use more expensive models – the uplifts in conversion and ads click-through-rate more than offset the cost of using third-party LLMs. Engagement with our Seller Assistant continues to rise, with sellers representing almost half of our GMV using it, interactions per seller rising, and DAU growing 21% MoM in June. As engagement rises, our assistant is resolving a growing share of interactions: in Q2’26, more than half of the requests on our Help Portal were solved by the assistant without any human intervention…

…Our shopping assistant, we are just AB testing that one, nothing to really share in terms of engagement and results. We’re very excited with the early results we’re seeing on the shopping assistant, which is on live for some consumers in the marketplace…

…Four years ago, we used to have 10,000 reps in customer service. Today, we have 7,000 reps, even though the business grew by 3x in that period of time. That’s because 90% of the interactions are done without a human participating on the issue.

MercadoLibre’s management has built technology to maximise results while lowering cost when using AI models; management has obtained better commercial terms with AI suppliers; management has lowered MercadoLibre’s cost per token on a year-on-year and sequential basis in 2026 Q2; MercadoLibre’s AI investment is split between cost of goods sold and product development; MercadoLibre’s AI investment grew in dollar-terms in 2026 Q2, but product development as a percentage of revenue declined year-on-year; MercadoLibre was able to grow its business even when slowing down the pace of hiring engineers, because of AI-driven productivity gains; 2026 is the first year in many years where MercadoLibre is not growing its engineering team

We have built technology that analyzes usage patterns across teams, helping us improve our mix of models, reduce inefficiencies in how they’re used, and refine the tools built on top of them – all aimed at maximizing results at a lower cost. This, combined with better commercial terms negotiated with AI suppliers, has resulted in a reduction of cost per token QoQ and YoY. AI investment grew roughly $80mn YoY in Q2’26, split between cost of goods sold and Product Development. Despite this, Product Development expenses fell from 8.4% of net revenue in Q2’25 to 7.2% in Q2’26, as the productivity gains described above allow us to grow without adding engineers at the pace we once would have…

…AI is definitely contributing to cost efficiency. 2026 is probably the first year in many years in which we are not growing our engineering team. That’s also coming from the fact that AI is driving developer productivity up consistently.

Shopify (NASDAQ: SHOP)

Shopify’s management sees Catalog as a source of truth for AI product discovery; Catalog has over 1 billion products in its search index, and structures product data so any AI partner’s agents can access it directly; management believes Catalog will be among Shopify’s most important assets in the future; AI searches powered by Catalog converted at twice the rate of those using scraped data in 2026 Q2; management sees Catalog as the discovery engine for AI; dozens of retailers and platforms have adopted UCP (Universal Commerce Protocol) since its launch in early-2026; every Shopify merchant is UCP-ready, such that agents can access their product data and check out; every Shopify merchant’s products are automatically listed in Catalog; any developer can access UCP and the Catalog API; Catalog is built with Shop sign-in, so agents can recognize returning buyers and surface personalized recommendations, and this is a feature no other catalog API can do; management is now building taste-driven attributes into Catalog

Catalog, which you can think of as the authoritative source of truth for AI product discovery of the world’s best products and best brands. For nearly two years, we’ve been investing in the search index, ensuring over a billion products and 20 years of commerce experience is distilled for agents. It structures merchants’ product data so every and any AI partner can access it directly, giving agents the ability to discover, understand, and recommend our merchants’ products. Let me say this, Catalog will be one of Shopify’s most important assets for years to come, here’s why.

We’re seeing that AI searches powered by Catalog converted twice the rate of those using scraped data. That is because with Catalog, merchants’ products show up complete, accurate, and with the right context when someone is ready to buy. Put simply, Catalog is the discovery engine for the future, Shopify built it and owns it…

…We introduced UCP at the start of 2026, already industry players across the commerce stack and beyond are converging on this unified protocol with dozens of retailers and platforms adopting it to date…

…Every Shopify merchant is UCP-ready. Agents and builders can access their product data, create carts, and even check out using the protocol. Everything flows through Shopify, so their checkout logic and fulfillment rules are perfectly preserved. Their products are also automatically listed in Catalog. Every builder can now access UCP and the Catalog API across millions of merchants, so they can build commerce experiences with the same infrastructure as our major AI partners. We built our Catalog the way only Shopify could, integrated with Shop sign-in so agents can recognize returning buyers and surface personalized recommendations based on their purchase history. No other Catalog API can do this…

…I don’t know if you tuned in to Editions.dev two weeks ago, but we talked about these taste-driven attributes that we’re now building in to Catalog things like, is it formal enough for a wedding? Does it have breathable fabric? What’s the wrinkle tendency? Are these sneakers suitable for an endurance run? These sort of taste-driven attributes are things you can only get with Catalog.

Shopify’s management recently rolled out a new agentic section in the Shopify admin, which provides cross-channel attribution for agentic selling; agentic commerce volume is still small for Shopify, but it’s growing rapidly, with AI-driven traffic and orders to Shopify stores tripling year-on-year in 2026 Q2 (AI-driven traffic to Shopify stores was up 8x year-on-year in 2026 Q1 and orders from AI-powered searches were up 13x year-on-year); new buyer orders from agentic channels are coming in at nearly 2x the rate of other channels; traditional search is still the largest source of traffic for Shopify merchants and it’s growing, up 1.3x in the last 2 years; management sees AI as a complement to search for commerce; management is seeing that AI search has been particularly helpful for smaller merchants that form the long tail of commerce; 75% of Shopify’s AI-attributed orders in 2026 Q2 came from outside its top 100 categories; the phenomenon of long tail merchants doing well in AI search can be attributed to AI agents using Shopify’s catalog to find the product that works best for the consumer; half of all AI-referred sessions are landing directly on product description pages, 2.5x more than with traditional search, and is a strong tailwind for Shopify merchants; agentic transactions have the same economics for Shopify merchants as online store transactions, with no extra fees; management thinks Shopify merchants will benefit disproportionately from agentic channels; conversion from AI search is 80% higher than traditional organic search; management is seeing that large retailers are being pushed to figure out their agentic strategy, and this is where Shopify can help 

In May of this year, we rolled out our new agentic section in the admin, the first cross-channel attribution for agentic selling. Merchants can manage AI channels, they can track performance, and they can get specific recommendations on what to improve, all from a single interface. While the volume from agentic commerce is still small relative to our massive GMV, the growth trends are impressive. Both AI-driven traffic and also orders to Shopify stores tripled year-over-year in the second quarter. New buyer orders are coming in at nearly twice the rate of other channels…

…Search remains one of our largest sources of buyer traffic to our merchants, and it’s still growing. Traditional search sessions are up 1.3x over the past two years, holding roughly a third of all storefront sessions. That is AI as a complement to search rather than a substitute for it…

…Early indications show that AI search has been particularly helpful to some of the smaller brands that form the long tail of commerce. These are brands that also happen to make up the majority of Shopify’s merchant base, smaller businesses with specialized products built for a particular customer. We saw that AI search was starting to disproportionately benefit the long tail in 2025, and that trend has continued, with 75% of AI-attributed orders in the second quarter coming from outside our top 100 categories in Q2.

The explanation is simple. While search engines rank by popularity against a handful of keywords, AI agents make multiple calls into Shopify’s Catalog, working with richer, structured data to match products with the buyer’s specific intent rather than just keywords. When a buyer asks an AI assistant for the best car seat that fits three across a sedan, traditional search focuses on the keyword car seat. An agent, however, understands the actual need, the dimensions, the vehicle type, and the fact that they need three. It searches across all of those constraints at once to find the product that actually works, not just the one that ranks highest. In this world, relevancy reigns. Specific products made for a specific buyer do particularly well…

…Buyer shopping journeys are being compressed as half of all AI-referred sessions are landing directly on a product description page. That is 2.5x more than what we see with traditional search. This is a serious tailwind for our merchants and in turn for us at Shopify…

…Agentic transactions carry the exact same economics as an online store transaction. There’s no new fees. There’s no separate pricing…

…I mentioned a car seat that fits three across a sedan. These are real Shopify products discovered because an AI agent understood what the buyer actually needed, That’s a structural advantage for these small, specialized, independent businesses. That’s our base. That’s our sweet spot. We think that these trends suggest that merchants on Shopify will disproportionately benefit from this new surface area…

…When you zoom it even further, conversion from AI search runs nearly 80% higher than traditional organic search as well…

…On the merchant side, yeah, I meet these very large retailers and the executive teams there literally every single week. Every one of them is being pushed to figure out what their agentic strategy is. By coming to Shopify, we take their agentic strategy off their plate.

Daily active merchants using Sidekick was up 3.6x year-on-year in 2026 Q2, with daily sessions up 4.8x; in 2026 Q2, Sidekick handled 34 million conversations and created 36,000 custom apps, up from 12,000 in 2026 Q1; Sidekick’s personalised guidance for new merchants led to an 8% increase in merchants reaching 5 orders within 15 days; Sidekick can now access data and take action on 3rd-party apps without the merchant leaving Sidekick; merchants of all sizes are adopting Sidekick; the way new merchants and established merchants use Sidekick is different; most of Shopify’s AI costs from merchant use of Sidekick appear in the Subscription Solutions’ segment’s gross profit; management was able to hold gross margin for the Subscription Solutions segment steady even as Sidekick usage scaled; management believes that investments in Sidekick will translate into more merchants joining Shopify and more merchants achieving greater success; power users of Sidekick are pushing it into areas such as advanced design, content, SEO etc.

In the second quarter, daily active merchants using Sidekick were up 3.6x year-on-year, and daily sessions were up 4.8x. It handled nearly 34 million conversations, and it was used to create more than 36,000 custom apps, up from 12,000 in Q1…

…Sidekick’s personalized guidance for new merchants during onboarding led to an 8% increase in merchants reaching five orders within 15 days…

…It can now access data and take action through extensions to third-party apps like Klaviyo without the merchant ever leaving Sidekick. Adoption is widespread across merchants of all sizes…

…In a merchant’s first 30 days, roughly half of their conversations with Sidekick are about store setup, design, and theme configuration. For merchants five years in, that drops to about 8%, while analytics and reporting climbs past 40% as they use Sidekick as their intelligence layer to interrogate their own data and make better decisions…

…As a reminder, the vast majority of AI costs related to merchant use of Sidekick appear in subscription solutions gross profit. We were able to hold gross margins at a relatively consistent level quarter-over-quarter while Sidekick usage scaled, which reflects some cost efficiencies and support, as well as our ability to continue providing merchants unique AI solutions like Sidekick while diligently managing cost. We are big believers in Sidekick and the value that it can deliver to merchants. We believe these types of investments in our platform will translate into more merchants joining the platform, and those merchants having even greater success. That translates to more gross profit for us, but more importantly, it is helping our merchants accelerate their businesses…

…We’re already seeing power users, if I can use that term, of Sidekick, pushing it way further into things like advanced design or content or SEO or even product creation.

Shopify’s management has built connectors to agents from 3rd-party platforms, so Shopify merchants can build on Shopify however they choose; Shopify has integrations across vibe coding platforms for developers to build in

We’ve built connectors to agents including Claude, ChatGPT, Perplexity, Manus, Replit, and Vercel with our AI Toolkit. Our merchants can build on Shopify however they choose…

…Our integrations across vibe coding platforms like Lovable, AI chat agents, and CLI IDEs show Shopify’s commitment to meeting builders where they are, however they choose to get there.

Shopify’s management thinks the company has moved from reflexive use of AI to having AI being a source of leverage; management wants to be thoughtful about AI costs and will use the appropriate models for different use cases; management thinks widespread adoption of AI tooling already is and will continue to yield benefits for Shopify; most of Shopify’s internal AI spend goes into R&D, and R&D as a percentage of total revenue has been declining; Shopify has built distilled models by teaching a large frontier model about specific use cases; the distilled models run faster, are less costly, and can even be better than large frontier models at narrow tasks

We’ve moved from a place of just reflexive use of AI to a place of AI leverage. Our AI philosophy is straightforward: maximum leverage paired with thoughtful cost management. We use the best model for the job, frontier intelligence where it matters, less expensive models where it doesn’t. We believe widespread adoption of AI tooling already is and will continue to yield benefits in the quality of our output…

…R&D, the majority of our internal AI spend is allocated here, so you’ve seen a modest uptick in year-over-year growth. Overall, we’ve driven substantial leverage in R&D as a percentage of total revenue and will continue to be disciplined in managing this spend…

…We now have a number of distilled models where we take a teacher model, usually a big frontier model, then teach it a specific use case to a smaller model, which results in much faster, less costly, and actually sometimes even better at the narrow task.

---

Disclaimer: The Good Investors is the personal investing blog of two simple guys who are passionate about educating Singaporeans about stock market investing. By using this Site, you specifically agree that none of the information provided constitutes financial, investment, or other professional advice. It is only intended to provide education. Speak with a professional before making important decisions about your money, your professional life, or even your personal life. I have a vested interest in Alphabet (parent of Google), Coupang, Datadog, MercadoLibre, Meta Platforms, Microsoft, and Shopify. Holdings are subject to change at any time.

The way I see it, artificial intelligence (or AI), really leapt into the zeitgeist in late-2022 or early-2023 with the public introduction of DALL-E2 and ChatGPT. Since then, developments in AI have progressed at a breathtaking pace.

We’re thick in the action of the latest earnings season for the US stock market – for the second quarter of 2026 – and I thought it would be useful to collate some of the interesting commentary I’ve come across in earnings conference calls, from the leaders of technology companies that I follow or have a vested interest in, on the topic of AI and how the technology could impact their industry and the business world writ large. This is an ongoing series. For the older commentary:

• 2023 Q1 – here and here
• 2023 Q2 – here and here
• 2023 Q3 – here and here
• 2023 Q4 – here and here
• 2024 Q1 – here and here
• 2024 Q2 – here and here
• 2024 Q3 – here and here
• 2024 Q4 – here, here, and here
• 2025 Q1 – here and here
• 2025 Q2 – here and here
• 2025 Q3 – here, here, and here
• 2025 Q4 – here, here, here, and here
• 2026 Q1 – here, here, here, and here

With that, here are the latest commentary, in no particular order:

Alphabet (NASDAQ: GOOG)

Alphabet’s management recently announced new AI models, Gemini 3.6 Flash, 3.5 Flash-Lite, and  3.5 Flash Cyber; the Gemini Flash models are Alphabet’s workhorse models and management is seeing lots of demand for them because of their performance and cost; Gemini 3.5 Flash Cyber has the same performance as other larger, frontier cyber models; Alphabet is testing Gemini 3.5 Pro; Alphabet has started pre-training for Gemini 4 and management is excited by its progress; developers and enterprises have strong token usage of Alphabet’s models; 9 million developers are building with Alphabet’s models each month; Alphabet’s model APIs (application programming interfaces) processed 22 billion tokens per minute in 2026 Q2 (was 16 billion tokens per minute in 2026 Q1); management launched the Omni video model in May and it has led to a 40% increase in daily users creating videos on the Gemini app; Alphabet’s Gemma family of small open models that can run locally on devices, have been downloaded over 900 million times, with the latest Gemma 4 been downloaded 300 million times since launching in April; the Gemini app now has 950 million monthly active users (was 750 million in 2025 Q4), with daily users up 3x from a year ago; Gemini Spark, the personal AI agent within the Gemini app, is now available globally; the Gemini Flash models are great for enterprise needs;

Yesterday, we announced new models, Gemini 3.6 Flash and 3.5 Flash-Lite, which are cost-effective and highly efficient. We are seeing tons of demand for our workhorse Gemini Flash series because it hits the sweet spot of performance and cost. We also launched Gemini 3.5 Flash Cyber, which I’m really excited about. Paired with our CodeMender agent, it finds and fixes vulnerabilities and delivers performance at the frontier comparable to far bigger cyber models. Gemini 3.5 Pro is currently in testing, and our team is already building the next generation of models. We have started our most ambitious pre-training run yet for Gemini 4 and are excited by the progress we are seeing at the frontier.

Demand for our models is translating to strong token usage across developers and enterprise customers, and we continue to be supply constrained, a sign of momentum and rapid adoption. More than 9 million developers are building each month with our models across our APIs and key developer products. Our model APIs are now processing approximately 22 billion tokens per minute. That’s up from 16 billion just a quarter ago.

This quarter, we launched Omni. It allows users to create anything from any input, starting with video. Since launching at I/O in May, there’s been a 40% increase in daily active users creating videos on the Gemini app. Our Gemma family of open models, small enough to run on local devices, are hugely popular. These models have been downloaded over 900 million times, and our latest Gemma 4 models have been downloaded over 300 million times since launching in April…

…The Gemini app, which now has 950 million monthly active users with daily active users tripling in the last year. Users love new agentic features like Daily Brief and our personalized agent, Gemini Spark, which is now available in the U.S. and internationally…

…If you think about an area like customer service, you need really good voice quality, you need live streaming, you need the ability to reason on that. If you’re a professional services firm, you need high-quality summarization, content generation, et cetera. Flash does very well on all of that. 

Antigravity, Alphabet’s 1st-party agentic coding solution, has more than 2.4 million weekly active users; Antigravity has helped a team in Google Chrome compress a 2-year delivery timeline into 3 months; 83% of Alphabet’s sales team are using Gemini-powered tools every weeks, and achieving a 20% higher win rate; Gemini-powered agentic solutions are addressing 75% of support queries

Our agentic development platform, Antigravity, allows anyone to build in the agent-first era. It has more than 2.4 million weekly active users. Antigravity is a powerful tool for users and enterprises, and it’s completely accelerated how we build internally. As just one example, a team in Chrome is now on track to accelerate delivery by 8 times, compressing a 2-year timeline into 3 months through model-driven refactoring…

…83% of our sales team uses Gemini-assisted tools weekly, driving up to a 20% higher win rate when using customized pitch narratives. Our ads customer support teams use Gemini-powered agentic solutions that now autonomously address 75% of support queries, freeing them to solve ur customers’ most complex challenges. 

AI continues to be expansionary for Google Search; Google Search usage hit an all-time high during the recent World Cup; management recently combined AI Overviews and AI Mode into a seamless Search experience; management is infusing Google Search with frontier capabilities such as AI agents; Alphabet’s AI features are driving higher usage of Google Search; AI Mode in Search has surpassed 1 billion monthly active users since launching in October 2025 and is driving incremental growth in overall Search queries; Alphabet’s AI features are now sending billions of clicks per week to websites; Alphabet has lowered the cost of AI Mode responses to its lowest level since launch despite introducing more AI capabilities; management is encouraged with the monetisation on even commercial queries that show AI Overviews; users are asking more detailed questions in AI Overviews and AI Mode, leading to more opportunities for more relevant ads; management is deploying and testing new advertising formats in AI Mode, including Text Ads, Direct Offers, and Highlighted Answers; Highlighted Answers is showing early user traction

With Search, AI continues to drive an expansionary moment with new experiences resonating with users and driving growth in queries. As a big football fan, I was particularly excited to see Search usage hit an all-time high during the World Cup this year…

…We recently brought together AI Overviews and AI Mode into one seamless Search experience that combines our frontier capabilities with the best of the web. We are continuing to incorporate more frontier capabilities into Search with agents, personal intelligence, and notebooks. Our AI-powered features are driving increased Search usage. Since expanding AI Mode globally last October, we have surpassed 1 billion monthly active users. Just like AI Overviews, AI Mode is driving an incremental increase in Search queries overall, and we are now sending billions of clicks to websites every week through AI features in Search…

…Thanks to our engineering and hardware optimizations, this quarter, we reduced the cost of AI Mode responses to its lowest level since launch, even as we have brought more advanced AI capabilities…

…We continue to be encouraged with monetization performance on queries that show AI Overviews, even as we’ve expanded AI Overviews to more commercial queries. Across AI Overviews and AI Mode, people are asking more specific and detailed questions, providing opportunities for more relevant ads. Within AI Mode, we continue to test and deploy a range of new ad formats. For Text Ads, we improve performance by adding contextual site links based on the conversation. Direct Offers is gaining momentum, with partners like IHG Hotels & Resorts soon surfacing special offers during trip planning. In Highlighted Answers, our latest experience, placing clearly marked sponsored links inside list responses is showing early user traction. 

Alphabet’s management recently launched Ask Youtube, which is powered by Gemini and allows users to ask complex questions about Youtube videos; early engagement with Ask Youtube is encouraging, with 140 million users in June 2026; management is widening the Ask experience to Youtube’s search experience; Youtube introduced creator shows and custom sponsorships, where AI is used to dynamically surface videos tailored to a brand’s desired moment; Kate Spade worked with Youtube creators and saw a 3.25% brand lift in purchase intent; Youtube’s direct response growth is fueled by brands’ usage of Demand Gen; Arc’teryx used Demand Gen to achieve a 70% better return on advertising spend

We’re also bringing the power of conversational AI directly into the YouTube experience. Ask YouTube uses our Gemini models to let people ask complex questions about individual videos, get quick takeaways, and jump straight to moments in those videos. The early engagement is encouraging. More than 140 million users engaged with Ask YouTube on the Watch page in June 2026. We are bringing that Ask experience to the wider search experience on YouTube…

…Over 550 million watched them on their televisions. This made the FIFA World Cup 2026 the most viewed World Cup in YouTube history. Advertisers connected with fans through FIFA channel takeovers, Gemini-powered soccer-themed sponsorships, and game day mastheads… 

… More broadly, on YouTube, we introduced an exclusive slate of creator shows making it easier for brands to tap into the fandom of creators. We launched custom sponsorships to put brands at the heart of the world’s biggest moments as they unfold on YouTube, using AI to dynamically surface videos tailored to a brand’s desired moment. Brands continue to partner with YouTube creators to engage new audiences. Kate Spade reached Gen Z through a first-of-its-kind YouTube creator campaign and partnered with creators Ellie Thumann and Hannah Meloche. By leveraging multiple format videos on YouTube, the brand drove a 3.25% brand lift in purchase intent… 

SMBs continue to fuel our direct response growth through campaigns like Demand Gen. They leverage our AI to scale visual storytelling across YouTube, Shorts, and now Google Maps. To attract high-value shoppers across North America and Europe, outdoor brand Arc’teryx used Demand Gen to achieve a 70% better return on their ad spend compared to other paid channels.

Google Cloud’s momentum is driven by Alphabet’s vertical integration of the entire AI technology stack; the Gemini models are a key driver of Google Cloud’s growth and they are integrated into all of Google Cloud’s products; Google Cloud is seeing strong, diversified demand across products, customers, geographies, and industries; Google Cloud is enjoying growth from winning new customers (customer acquisition doubled year-on-year in 2026 Q2), expanding with existing customers (existing customers are exceeding commitments by 50% in 2026 Q2, an acceleration from 2026 Q1), and partners (Google Cloud Marketplace transactions were up 7x year-on-year in 2026 Q2); the Gemini Enterprise platform is seeing rapid adoption, with its Agent Development Kit reaching nearly 70 million total downloads in 2026 Q2; nearly 90% of the Fortune 100 companies are using Gemini Enterprise; businesses are using Gemini to build custom agents, automate processes, and more; 500 Google Cloud customers processed over 1 trillion tokens each over the last 12 months (was 330 in 2026 Q1), with 2,000 customers consuming over 100 billion tokens each; Booking Holdings has expanded a multi-year commitment with Google Cloud; it’s still early days for AI-native and AI-enabled workloads on the cloud, and it’s also early days even for overall workloads that have shifted to the cloud 

Next, Cloud. Our continued momentum is driven by our integrated AI portfolio consisting of chips, models, data, security, and agent platforms, all designed to work together. Gemini continues to be a key driver of growth and is deeply integrated across all of our cloud products, including Gemini Enterprise, data analytics, cybersecurity, and Google Workspace. We are seeing strong, diversified demand across products, customers, geographies, and industries. Our product differentiation is driving expansion in three ways. We are winning new customers, more than doubling our acquisition velocity year-over-year. We are deepening our relationships with existing customers who are expanding their usage and exceeding their commitments by more than 50%, also, an acceleration over last quarter. We are driving growth with partners, with transactions on Google Cloud Marketplace growing over seven times year-over-year.

One of the strongest parts of our growth comes from the rapid adoption of our Gemini Enterprise platform. It’s differentiated with easy-to-use tools to build agents and automate processes, connectivity to enterprise systems, cost management, and governance tools. In Q2, Agent Development Kit, our framework for building and deploying enterprise AI agents, reached nearly 70 million total downloads. As I said, nearly 90% of Fortune 100 are using Gemini Enterprise. We have customers like PepsiCo for AI and analytics solutions, Intel to streamline core processes, HSBC for wealth management, Bell Canada for customer engagement, Macy’s for commerce experiences, and SIGNAL IDUNA for knowledge management. More broadly, Gemini is transforming how millions of businesses use AI to build custom agents, automate processes, improve cybersecurity, manage customer relationships, streamline data analytics, collaborate effectively, and more.

All of this momentum is driving growth in our paid token usage. Nearly 500 Cloud customers have each processed more than 1 trillion tokens in the last year, and usage is so much deeper than that. Over the last 12 months, more than 2,000 enterprises consumed over 100 billion tokens…

…Booking Holdings expanded a multi-year cloud commitment and are partnering closely

to advance our AI-powered ad formats. It’s also deploying Google’s AI technology to enable new

customer experiences like agentic dining reservations on OpenTable, helping restaurants get

discovered and booked right when it matters most…

…We used to talk about cloud itself, very small percentage of overall workloads and enterprises have shifted to cloud. Think about what percentage of workloads are really AI native and AI enabled. It again feels very early. 

Alphabet’s management is seeing strong interest for Google Cloud’s AI-powered security offerings; 90% of Fortune 100 companies are Google Cloud Security users; 90% of Wiz customers are using AI-powered security features; in 2026 Q2, the number of AI workloads protected by Google Cloud’s security platform saw a 45% sequential increase; Alphabet has a new security offering called Google AI Threat Defense

We’re also seeing strong interest in our AI-powered security platform, which is differentiated because it integrates threat intelligence, cyber response prioritization with Wiz, AI-automated scanning, code remediation, and monitoring. Today, 90% of Fortune 100 are Google Cloud Security users, nearly 90% of Wiz customers are using AI-powered security features, and we have seen a more than 45% quarter-over-quarter increase in the number of AI workloads scanned and protected by our security platform. Our security tools are being used to protect critical infrastructure, including financial services organizations such as Morgan Stanley, telecommunication providers such as Telus, healthcare organizations such as Texas Children’s Hospital, software companies such as Atlassian, and several government agencies. With our new Google AI Threat Defense, we are really excited to bring our new cyber model and CodeMender to help our customers defend against AI threats.

Alphabet’s management believes Google Cloud offers the industry’s widest range of AI accelerators, including Alphabet’s latest TPU 8t and 8i, and NVIDIA’s latest Vera Rubin GPU platform; Google Cloud’s Virgo Network connects a million AI accelerators across multiple data center sites into a unified supercomputer; Google Cloud has native support for AI accelerator programming languages, which enables workload portability across GPUs and TPUs; Google Cloud’s agent-optimised Axion CPU has 30% better performance per dollar than other CPUs; Google Cloud’s AI infrastructure is seeing strong demand from a wide variety of companies, such as AI labs, technology companies, financial services providers, pharmaceuticals, and robotics companies

We offer the industry’s broadest range of accelerators from Google and NVIDIA, including the new NVIDIA Vera Rubin platform and TPU 8t and 8i, which delivers strong price performance…

…First, our Virgo Network, which is designed to meet the needs of modern large-scale AI workloads. It allows customers to connect a million AI accelerators across multiple data center sites into a unified supercomputer. Second, our software stack has native support for JAX, PyTorch, vLLM, and SGLang, enabling workload portability across GPUs and TPUs. Third, our new agent-optimized Google Axion CPU provides 30% better performance per dollar compared to peer offerings. 

We are seeing strong growth and demand for our AI infrastructure offerings from leading labs such as Ineffable Intelligence, next-generation AI builders, including Kakao, financial services like Deutsche Börse Group, pharmaceutical companies such as Pfizer and Roche, and robotics and spatial intelligence companies such as World Labs. 

Waymo has introduced a new vehicle, Oasis, to riders; Oasis is the 1st vehicle powered by the 6th generation Waymo Driver

Waymo introduced its newest vehicle, Oasis, to public riders. This is the first vehicle powered by the sixth-generation Waymo Driver and will welcome more riders in the coming months. 

Alphabet’s management continues to accelerate the deployment of Gemini across the company’s entire advertising infrastructure; Gemini improves Search Ads’ ability to find relevant ads for longer searches that were previously difficult to monetize; in 2026 Q2, Gemini drove a 20% improvement in showing highly relevant ads in Shopping ads; 500,000 advertisers have adopted AI Max; users of AI Max enjoy 15% more conversions or value on Search at similar ROAs; AAA Auto Club Enterprises used AI Max to drive a 17% improvement in conversion volume and an 11% decrease in cost per lead; half of Alphabet’s SMB (small, medium business) advertising customers use its generative AI tools for creative development

We continue to accelerate the deployment of Gemini across our entire ads infrastructure to boost performance in three areas mentioned before: ads quality, advertiser tools, and AI user experiences.

First, ads quality. At Google Marketing Live, we showcased how Gemini improves query understanding, allowing us to find relevant ads for longer searches previously difficult to monetize. The core engine of our Search Ads relies on a dual prediction, delivering immediate utility for the user while maximizing measurable value for the advertiser. Gemini completely supercharges this capability. We use Gemini’s advanced reasoning to decode the nuances of longer, more detailed queries. With Shopping ads, for instance, we drove a 20% improvement in showing highly relevant ads, helping shoppers immediately find the best match. 

Second, advertiser tools. Take AI Max. It’s out of beta, and 500,000 advertisers have already adopted it. Those who adopt our AI-powered campaigns like AI Max or Performance Max, see an average of 15% more conversions or value on Search at a similar ROAs. AAA Auto Club Enterprises used AI Max to personalize creative assets and capture growth from increasingly detailed insurance searches. This led to a 17% improvement in conversion volume and an 11% decrease in cost per lead. We see strong adoption of our generative AI creative tools, which makes creative development easier, especially for SMBs. In fact, over half of our SMB customers globally use AI to create or optimize their creatives…

Merchants are rapidly adopting Alphabet’s UCP (Universal Commerce Protocol) for agentic commerce

In collaboration with the retail industry, we established the open-source Universal Commerce Protocol, UCP, as the new standard for agentic commerce. Merchants are rapidly adopting UCP, with Target and Steve Madden now live, while new members have joined the UCP Shopping and Food Tech Councils to help steer its vision.  

Google Cloud had 82% revenue growth in 2026 Q2 (was 63% in 2026 Q1) driven by growth in GCP; GCP grew at a higher rate than Google Cloud’s overall growth; Google Cloud’s growth was driven by core GCP, AI solutions, and AI infrastructure; Google Cloud started recognising revenue for the first time in 2026 Q2 from selling TPUs; Google Cloud’s revenue growth still accelerated materially even after excluding TPU sales; Google Cloud operating margin was 35.6% (was 32.9% in 2026 Q1 and 20.7% in 2025 Q2); Google Cloud backlog grew 11% sequentially to $514 billion in 2026 Q2 (was $462 billion in 2026 Q1); most of Google Cloud’s backlog are GCP contracts and just over 50% of the backlog is expected to be recognised as revenue in the next 2 years; when external TPU sales agreements are signed, they are reflected in Google Cloud’s backlog; Google Cloud’s $514 billion in 2026 Q2 primarily consists of GCP contracts; external TPU sales impact Alphabet’s operating cash flow because the company needs to build ahead of selling the systems; the use of TPUs in Google Cloud brings better margins

Cloud revenues were up 82% to $24.8 billion, driven primarily by GCP, which grew faster than cloud overall. Core GCP, AI solutions, and AI infrastructure were all important drivers of growth. We also began to recognize revenues from TPU system sales, which we delivered to customer data centers for the first time in Q2. Cloud revenue growth accelerated meaningfully even after excluding the impact of TPU system sales. Cloud operating income was $8.8 billion, more than tripling year-over-year, and operating margin increased from 20.7% in the second quarter last year to 35.6%. Google Cloud’s backlog increased by more than $50 billion sequentially, reaching $514 billion in the second quarter. The increase was driven by strong demand for our enterprise AI offerings. The majority of the backlog is related to typical GCP contracts from a broad mix of customers, and we expect to recognize just over 50% of the total backlog as revenue over the next 24 months…

…In terms of revenue recognition and how we look at the TPU system sales. The way to think about it is the following. When we sign the agreements that I’ve mentioned in the prepared remarks, they would be then reflected in the cloud backlog. The vast majority of the $514 billion of cloud backlog is the GCP agreements, but the TPU system sales are reflected in that backlog. We start building inventory to be able to sell those systems. You see that impact on the cash from operations because we built ahead, obviously, as we’re building that business and ramping up. Once we start delivering the sales, generally that’s when we start recognizing revenue…

…On the TPU margins, we don’t break out margins for any specific products or infrastructure component. Certainly, there are benefits from designing and manufacturing our own chips.   

Alphabet’s management is seeing significant demand within Google Cloud; management continues to expect to recognise most of the revenue of external TPU shipments in 2027; management expects to use 3rd party centers in 2026 Q3 as a bridge while building internal compute capacity, as Google Cloud remains constrained by supply, and the use of 3rd-party data centers will result in near-term margin pressure; management has raised capex guidance for 2026 to $195 billion to $205 billion (was previously $180 billion to $190 billion; 2025’s capex was $91.4 billion, which was itself up 65% from $55.4 billion in 2024, and 2024’s capex was up 69% from 2023); the higher capex guidance is because of an acceleration in the delivery of capacity to meet growing demand (i.e. Google Cloud is pulling-forward capex); management continues to expect 2027’s capex to be much higher than 2026’s; management expects Alphabet’s free cash flow to remain under pressure because of AI-related capex; management continues to see attractive returns on Alphabet’s capex; the supply constraint of Google Cloud comes from strong demand from both external customers and internal use cases; the use of 3rd-party data centers is to support large customers at a high near-term cost, but there’s still high ROI (return on investment) over the entire life-time of the deals with these customers; when Alphabet’s input costs for capex goes up, management is able to raise prices for its services; Alphabet’s level of capex in 2027 is driven by strong demand indicators that management is seeing, and the dynamics in the market look even better today compared to a year ago

In Google Cloud, we are seeing significant demand for products and services, which we expect to drive strong growth. As I mentioned earlier, we started delivering TPU system to customer data centers in the second quarter. We continue to expect to recognize a relatively small portion of the revenues from our existing TPU system sales agreements this year, ramping as we exit 2026. We anticipate the vast majority of the revenues from these agreements will be realized in 2027. Given the supply-constrained environment, we plan to expand the use of third-party capacity in Q3 as a bridging strategy while we build out more internal capacity. This strategy allows us to keep growing our customer base and capture greater overall value. However, it will create modest margin pressure in the near term as we utilize this capacity…

…We are updating our full-year 2026 CapEx guidance range to $195 billion-$205 billion, up from our previous estimate of $180 billion-$190 billion. The increase in the range is primarily due to an acceleration in the delivery of capacity to meet growing demand. As we previously shared, we continue to expect our CapEx to increase significantly in 2027 and will provide more details at a later date…

…We expect the free cash flow will remain under pressure, driven by our investments in technical infrastructure, which enables us to capitalize on the AI opportunity and continue to drive attractive returns. Q2 represented another strong quarter…

…It feels like we are in very early innings of what feels like secular shift across multiple areas. In our core information businesses, just the possibilities when I see what all you can do with the absolute frontier capabilities, there’s still a lot of work ahead to translate all that into experiences for our consumer users. You can think about end-to-end agentic experiences to really meaningfully do a lot more for them. All of that looks like extraordinary opportunities with extraordinary returns for executing well on those opportunities… 

…We are seeing very strong demand, both from external cloud customers as well as across the business…

…I think on the bridge deal, the main thing I would say is, look, on the margin, there are very, very large customers of ours on Cloud who we are trying to support them through this extraordinary moment. The incremental opportunities they are bringing to us, while a short-term cost over a few months may be very high, in the lifetime of the deal, as we bring more capacity on, is highly ROI positive…

…To the extent that our input cost is going up to us, we reflect that in our ability to price our solutions and see returns there. All of that is factored into how we are planning…

…Our compute capacity investments in 2027, to the first question I answered, I think we are seeing strong demand indicators, including long-term deals, the existing deals which we have, which are renewing with exceptional demand on a moving forward basis. We are using all that to plan and invest accordingly. I think if anything, the dynamics look healthier than where we were about a year ago, that’s what gives us the confidence to undertake those investments.  

Alphabet’s management is aware that the company does not have frontier capabilities in coding and agentic coding and the teams are working on it; 3.6 Flash is already showing improvement in coding compared to 3.5 Flash; management is very confident that Alphabet will be at the frontier again and they are being very ambitious with Gemini 4; Gemini 4 will be a much larger base model compared to the previous generation; management expects Alphabet to pick up the pace to almost a monthly-cadence when it comes to model releases; management wants Gemini 4 to compete at the frontier level of where the frontier will be when Gemini 4 is released

There are areas where we’ve acknowledged we need to improve. Coding and agentic coding is an example of that, and the teams are very focused on it…

…In terms of agentic coding, we are iterating, and you will see us make continued iterations. 3.6 Flash, for example, compared to 3.5 Flash, jumped over 10 points in DeepSuite as a benchmark, and it is more token efficient doing so. We are using it internally. We are testing it with many customers in coding. We see the progress just in six weeks from the prior version to the new version, and you will see more continued iterations on that as well. 

In terms of the frontier, we are both very committed and very confident of being at the frontier. For the next generation of frontier, you’re going to need much larger base models. We are now training Gemini 4, and we’re being very ambitious with it…

…We will need Gemini 4 as a larger base model to compete at that frontier level, so we are focused on executing on that well…

…On the speed of model releases, I do think you will see us continue to pick up pace…

…Picking up pace and releasing models almost at a monthly cadence is part of our roadmap as we are building Gemini 4 as well…

…We want to compete at the frontier level of where the frontier will be when Gemini 4 comes out, we are applying a lot of our compute and effort in that direction.

Alphabet’s management’s priority with TPUs is to allocate to internal teams for frontier model development; to balance demand for TPUs, Alphabet is using both TPUs and GPUs to serve its models

In terms of allocating our TPUs, look, our first priority is making sure we are allocating what we need to compete at the frontier in terms of AGI development…

…Given the extraordinary demand to balance the external demand, even for most cloud customers, we are using both TPUs and GPUs mainly for serving our models. Think Vertex AI, Gemini Enterprise, the momentum we see. We’re using it or agentic workloads, et cetera. We are using it for those purposes…

Amazon (NASDAQ: AMZN)

AWS grew 36.7% year-on-year in 2026 Q2 (was 28% in 2026 Q1) and is now growing at its fastest pace in 18 quarters; AWS added $4.6 billion in revenue sequentially, 80% more than the largest increase; AWS’s backlog is $496 billion in 2026 Q2 (was $364 billion in 2026 Q1), up triple-digits year-on-year; AWS’s run rate has reached $169 billion (was $150 billion in 2026 Q1); AWS’s AI business revenue run rate is $25 billion in 2026 Q2 (was $15 billion in 2026 Q1), up triple digits year-on-year; management thinks customers are choosing AWS for AI for 4 reasons, namely, (1) AWS’s broad capabilities, (2) customers want their AI inference to be at where their other applications and data reside, and this happens to be in AWS, and (3) AWS has the strongest security and operational performance; AWS is seeing growth in both AI and non-AI workloads, and growth in one is driving growth in the other; AI workloads drive non-AI workloads because (1) post-training reinforcement learning and agent tool use are mostly run on CPUs, and AWS’s Graviton CPU has 30%-40% better price performance than competitors, and (2) AI workloads need databases, and this is an AWS strength too; management sees enterprises as being very early in using inference at scale; management now believes AWS can be a trillion-dollar annual revenue business in time, up from the previous view of a “few hundred billion dollar revenue business”; management is seeing customers who want to benefit from AI accelerate their migration to the cloud; management is seeing a strong correlation in customers’ AI spend and core growth in AWS; AWS is on track to doubling its 2025 power capacity by 2027 

Revenue growth of 36.7% year-over-year, accelerating for the fifth straight quarter, our fastest growth in 18 quarters back when AWS was less than half its current revenue size. We added over $4.6 billion in revenue quarter-over-quarter, about 80% more than our largest increase ever. Our backlog stands at $496 billion, growing triple digits year-over-year. AWS is now a $169 billion annualized revenue run rate business, which, for perspective, would place it 24th on the Fortune 500 list if it was a standalone company…

…Our AI revenue run rate climbed significantly quarter-over-quarter, and is now also over $25 billion, growing triple-digit percentages year-over-year.

Customers choose AWS because we offer the broadest capabilities. They want their AI inference to reside near their other applications and data, and more of it resides in AWS than anywhere else. Because AWS has the strongest security and operational performance, we’re seeing strong growth across both AI and non-AI, what we call core, and growth in one is driving growth in the other. Growth in AI drives core because post-training reinforcement learning and agent tool use is mostly done on CPUs versus AI accelerators. This is an advantage for AWS, as our Graviton chip is the strongest CPU chip, offering up to 30%-40% better price performance than other options. You need a place to store this AI data and to run vector databases, which are also emblematic of a meaningful edge for AWS because we have the broadest and most capable functionality by a fair bit in these core infrastructure areas. We feel similarly about the AI stack, top to bottom. We have a unique offering that customers are excited about…

…Remember, enterprises are still very early in using inference at scale in their current production applications. We long believed AWS could become a few hundred billion-dollar revenue business and now believe it’ll be at least double that, and very possibly be a trillion-dollar annual revenue business for us in time, with very appealing accompanying free cash flow and return on invested capital…

…Increasingly, customers seeking the full benefits of AI are accelerating their transition to the cloud. We see a strong linkage between AI spend and core growth. As customers invest in AI, we see a corresponding increase in core consumption…

…We’re on pace with the capacity build that we talked about a few quarters ago, where we said we expect to have double the power capacity by the end of 2027 that we had in 2025, and we continue to be on that track.

AWS’s chips business is now at a $25 billion annual revenue run rate in 2026 Q2 (was $20 billion in 2026 Q1), up triple-digits year-on-year; management thinks AWS’s Tranium AI chip and Graviton CPU both have leading price-performance; OpenAI and Anthropic, the 2 highest-profile AI labs, have multi-year, multi-gigawatt commitments to Trainium; a growing number of AI startups, and larger technology companies, are also adopting Trainium; Graviton is used by 98% of AWS’s top 1,000 EC2 customers; Graviton revenue commitments are up 3x sequentially in 2026 Q2; Graviton 5 is growing nearly 2x faster than Graviton 4 did; AWS continues to have a deep partnership with NVIDIA, as management knows that customers want choice; management sees incredible demand for Trainium and some customers who are interested in obtaining Trainium chips outside of AWS; management thinks there’s a real chance AWS will start selling Trainium chips to 3rd party data centers

Our chips business now has an annual revenue run rate of over $25 billion, growing triple-digit percentages year-over-year…

…We are unusually well-positioned for this AI inflection, given our leading price-performance chips in both AI with Trainium and CPU with Graviton. In addition to the two leading AI labs in the world, Anthropic and OpenAI, making multi-year, multi-gigawatt commitments to Trainium, an increasing number of AI startups are also adopting Trainium, including unicorns like Neurorobotics and Odyssey, joining startups like Twelve Labs, Descartes Labs, Poolside AI, Karakuri, Metagenomi, NetoAI, and Splash Music, and larger companies like Uber and Pinterest all adopting Trainium.Graviton is used by 98% of our top 1,000 EC2 customers. The revenue commitments have increased nearly three times quarter-over-quarter, and Graviton5 is growing nearly 2x faster as Graviton4 did.

We also continue to have a deep partnership with Nvidia, and we’ll continue making AWS the best place to run Nvidia chips, as we have customers who will run on Nvidia for as long as we can foresee, and we believe strongly that customers want choice…

…We just have an incredible amount of demand for Trainium. There are a lot of customers who are very excited about using it in the form that we’re providing right now. We do have an increasing number of customers who are interested in us providing the Trainium chips to them, separate from our cloud, and we’re actively having those conversations and exploring, and I expect there’s a real chance we’ll do that in the future.

Amazon’s management continues to think that technical companies will build their own foundation models, and AWS’s SageMaker AI service helps them do that; Bedrock, AWS’s fully-managed service for companies to build upon frontier models, provides high-performance and cost-effective inference; Bedrock provides the best selection of leading models at superior performance and with governance and security; Bedrock continues to grow rapidly; building agents at production scale is hard, so Amazon Bedrock Agents provides the building blocks for organisations to build and manage agents; management thinks most companies will use turnkey agentic services; AWS’s coding agent Kiro is 50% more cost-effective than competitors, and usage has tripled sequentially in 2026 Q2; Amazon Q is an AI work companion that lets users manage leading SaaS tools and take action for users; Amazon Q was recently made more capable with autonomous agents and more integrations with SaaS tools; Amazon Q’s customers include large multinational companies; AWS has agentic services such as Amazon Connect for call centers and AWS Transform for software migration; management recently released AWS Continuum, an agentic service leveraging frontier models for cybersecurity; management expects AWS Continuum to grow quickly; all 5 major airlines, and many leading banks and healthcare companies, are users of Amazon Connect; Amazon Connect is growing very quickly; 

As we’ve been saying for 18 months now, technically competent companies are going to build their own foundation models. Not the really big frontier models, but smaller models that leverage their proprietary data. There is no easier service for this than our SageMaker AI service. Customers also need a high-performance, cost-effective inference service, and that’s what Amazon Bedrock provides. Bedrock not only provides the best selection of leading models at superior performance and with the governance and security controls that companies need, it’s also continuing to grow incredibly quickly…

…After you’ve built an agent, you have a lot of muck to worry about. A production agent needs somewhere secure to run, memory so it holds context, an identity so it can act on a user’s behalf, tools and data to connect to, and a way to watch what it’s doing once real traffic hits. Stitching all that together reliably is hard, and it’s stalled many production deployments. It’s why we’ve built Amazon Bedrock Agents. It provides building blocks as managed infrastructure, and our teams keep iterating, recently adding features like policies which give companies deterministic controls over what agents can do, payments so agents can execute transactions autonomously, web search to ground agents’ knowledge without having to leave AWS, and a new harness that further speeds up how fast customers can put this all together, including creating the agent with Strands Agents. While companies will construct their own purpose-built agents from the ground up, most will also use turnkey agentic services…

…Our own spec-driven Kiro, which is up to 50% more cost-effective than others and tripled in usage quarter-over-quarter…

…Amazon Q, an intelligent AI work companion that helps you manage, search, and automate your digital workload across email, calendar, local or cloud files, and custom workflows. Unlike other offerings in this space, Q also lets you manage across leading SaaS tools like Slack, Salesforce, Jira, Teams, and ServiceNow. Q enforces a company’s existing access controls so each person sees only what they’re cleared to see. Then it takes action: scheduling meetings, drafting and sending email, updating a CRM record, building a dashboard, and more. In Q2, we made Q even more capable, adding autonomous agents that customers set up in plain language to run continuously in the background and carry out multi-step tasks, a personalized activity feed that pulls email, messages, calendars, and tasks into one prioritized view, and 16 new integrations, including Adobe, Moody’s, and Snowflake. Q has momentum, with 3M, Allianz, AstraZeneca, Autodesk, BMW, Exxon, FINRA, Hyundai, Intuit, Mondelēz International, Moody’s, the NBA, the NFL, Sun Life, and Southwest Airlines all using it.

We also have services like Amazon Connect, our call center service, and AWS Transform, which automates software migration growing quickly…

…We recently released AWS Continuum, which discovers, prioritizes, validates, and remediates code vulnerabilities. It starts by ingesting the backlog of vulnerabilities a team already has and then leverages the new frontier models to run comprehensive scans. Continuum uses agents in each company’s own business context to prioritize what matters, reasoning through questions like, “Is the affected component deployed? Is it reachable? Is it in a production path? What’s the impact if it’s exploited?” Then it validates vulnerabilities in a sandbox so teams aren’t chasing false positives. Finally, it recommends the fix. It is hard to talk with enterprises about AI right now without their mentioning security. We expect Continuum to grow quickly…

…Amazon Connect, which is our call center service, which is used by all five major leading airline providers, as well as many of the leading banks and healthcare companies, continues to grow very quickly.

Amazon’s management has clear line of sight to strong financial returns on the company’s originally-planned capex of $200 billion, or higher, for 2026; Amazon’s capex has 2 components, the data centers, and servers; spending on data centers is done 2 years before servers are slotted in to start monetisation; once a data center has servers in place, AWS immediately starts generating significant revenue; a data center can be monetised for 30-plus years without needing startup capital again; servers are typically purchased months before they are put into service; management will only buy servers when they see strong demand signals; servers take slightly less than 3 years to breakeven and have 5-6 years of useful life; most of AWS’s AI capacity is contracted for at least 5 years, so AWS earns significant free cash flow on servers in the 2-3 years after breakeven; AWS has a strong track record of pulling forward breakeven periods for servers; AWS typically gets 5-6 generations of server economics from data centers, with subsequent generations after the 1st having better overall economics because the upfront data center spending is not needed; when there’s demand for many data centers, Amazon has to spend ahead of time before the data centers can come online and be monetised, and management sees very compelling revenue, free cash flow, and ROIC (return on invested capital) a few years after the data centers are being monetised; management has gone through the same monetisation cycle during the 1st era of cloud computing; cloud computing’s demand-build was more gradual than AI; management is seeing margins and returns for Amazon’s AI build tracking slightly ahead of the cloud computing build at the same point of evolution; management has raised Amazon’s capex for 2026 to $220 billion (capex was $128 billion in 2025, and $83 billion in 2024 because of higher cost of memory chips; even with the higher capex guidance, management sees AWS as being supply-constrained in 2026 and 2027, and possibly 2028, with 2028 demand being striking; the lion’s share of AWS’s capacity in 2027 is already reserved, and quite a bit of capacity for 2028 is also already reserved

Earlier this year, we said we plan to invest approximately $200 billion in cash CapEx in 2026, the majority of which to support AI and AWS. At this level of spend and higher, we have clear line of sight to strong financial returns…

…There are two major parts of the investment, the data centers and the servers and networking equipment that go into them. These have different capital cycles. Data center capital is spent starting two years before we can put servers into them to start monetizing. Once a data center opens with servers plugged in, we start generating significant revenue right away and then get to monetize these data centers for 30-plus years without having to spend that startup capital again. Servers and networking equipment operate on a shorter cycle. We typically purchase these a few months before putting them into service, so we have strong visibility into customer demand before we trigger the spend. If the demand isn’t there, we won’t spend the capital.

For servers and networking equipment, on average, it takes a little less than three years to break even on that investment. The servers currently have a useful life of at least five to six years, and most of our AI capacity these days is being contracted for at least five-year terms. That means that we’re driving significant free cash flow on the servers and networking equipment in the two to three years after we break even. It’s also worth noting that AWS has a strong track record of pulling forward break evens on server equipment where we’ve already made meaningful progress and finding ways to extend the useful life of this equipment without sacrificing customer experience. For our data centers, which have 30-plus-year useful lives, we should get at least five to six generations of server economics, like I explained earlier, with subsequent generations after the first having even better overall economics because we don’t have to repeat that upfront data center investment I mentioned earlier.

This means in the short term, when demand is necessitating so many data centers being built simultaneously in advance of when we can start monetizing them, we’ll spend a lot of CapEx and encounter free cash flow headwinds until these data centers come online, can be monetized, and we get a few years into these servers being utilized. As we get a few years out and the revenue growth outpaces the incremental CapEx growth, which will happen at some point, the resulting revenue, free cash flow, and return on invested capital is very compelling. We’ve done this before in the first era of cloud computing, just over a longer time horizon, where demand built more gradually than it has in AI. We see the margins and returns in AI tracking what we saw with Core at the same point of evolution, actually a little ahead.

We now believe we will spend approximately $220 billion in cash CapEx in 2026. The higher cost of memory pushing this number up from our prior estimate of about $200 billion. Even at that amount, we will still not have enough capacity to meet all the demand we have in 2026, and I believe this dynamic will also be true in 2027, too. In fact, the demand we already have for 2028 is striking…

…We have so much demand right now. Apart from what we’ve talked about in 2026, the lion’s share of capacity in 2027, we’re adding a lot of capacity, as I mentioned just a few minutes ago, is largely reserved, and we have quite a bit of capacity that’s already been reserved for 2028. 

Over 350 million Amazon customers have used Alexa for Shopping, Amazon’s agentic shopping assistant, in the last 12 months; in 2026 Q2, active users of Alexa for Shopping was up nearly 100% year-on-year, and interactions was up 5x; management expanded Amazon Lens to 10 additional countries and it’s now available in 21 countries; in the US, customers who use Alexa for Shopping spend 40% more per order than those who don’t; customers who’ve tried Alexa+ are signing up for Prime at 25% higher rates

Customers love Alexa for Shopping, our agentic AI shopping assistant. It offers personalized recommendations, product comparisons, price history, and the ability to automate shopping through features like price alerts and auto-buy. Over 350 million customers have used it in the last 12 months, and engagement accelerated in Q2, with active users nearly doubling and interactions up over 5x year-over-year. We also expanded Amazon Lens, which lets customers take a photo of anything they see and instantly find the same or similar items on Amazon, to 10 additional countries, and it’s now available in 21 countries around the world…

…We find that everywhere Alexa goes, it drives momentum for the business. For example, in the U.S., customers who use Alexa for Shopping spend an average of over 40% more per order than those who don’t. Customers who’ve tried Alexa+ are signing up for Prime at nearly 25% higher rates.

Ads Agent is one of Amazon’s AI-powered advertising tools and it lowers campaign setup time from hours to minutes; advertisers using Ads Agent see 8% lower cost per impression and 6% lower cost per acquisition; management has expanded Ads Agent to 11 new countries in 2026 so far

We make it easy to create, launch, and optimize full-funnel campaigns using AI-powered tools, including Ads Agent, which turns hours of setup and targeting into minutes. Advertisers using Ads Agent targeting see 8% lower cost per impression and 6% lower cost per acquisition, and we’ve expanded it to 11 new countries this year.

Amazon’s management thinks that AWS can be wildly successful even if Amazon does not have its own frontier model, because there is not going be just one model to rule the world; management still wants Amazon to pursue building frontier models because it gives Amazon more control over costs and model-features; management thinks there will be at least 6 frontier models that are equally good over the next few years, and Amazon’s model will be among the mix

AWS and Amazon can have a wildly successful business without its own frontier model. A lot of that is because there is not going to be one model to rule the world. You already see that right now. You see it. It’s not just Anthropic, or it’s not just OpenAI. You see increasingly more and more companies being interested in the open models as well…

…All that said, we are pursuing our own frontier model, and we’re doing it for a few reasons. First of which is it just gives us additional control over cost. Cost for our own consumer applications, also we’re trying to drive costs down for customers. Having a player like ourselves that’s always focused on trying to take the price performance and the cost down for customers all the time, we think will help keep the models more cost effective for customers. I think also it allows us to have more control over prioritization on what models focus on. We have, both from our own external customers as well as our internal customers inside the company, certain priorities that matter that we want the models trained especially well for, then it gives us some control on speed. My view of it is that within the next few years, you’re going to have at least a half dozen models that are comparably good to each other.

In the 2025 Q4 earnings call, Amazon’s management said market demand for AI compute looked like a barbell with AI labs on one end spending a lot on compute for just a handful of applications, and with enterprises on the other end using AI for productivity purposes; now, the adoption curve for AI still looks like a barbell, with the middle being enterprise production workloads that are mostly not using inference pervasively; management thinks the middle of the barbell will become the largest AI workloads, and that AI will change every customer experience and lead to the invention of brand-new experiences; management does not know if the trajectory of the middle of the barbell will be as steep as seen currently with the ends of the barbell

We see this adoption curve in AI right now is very barbellled. There is, on one end of the barbell, the AI labs are consuming gobs and gobs of compute, and there are a few runaway successful generative AI applications like Claude Code and ChatGPT. On the other end of the barbell are enterprises who are getting real value from AI in cost avoidance and productivity. These are things like automating customer service or business process automation or fraud or things like that. In the middle of the barbell is all of the current enterprise production workloads, some of which are using inference in a pervasive way, but most of which aren’t. That is going to change very significantly over time. In my opinion, that will be the largest absolute segment, the existing production workloads in the enterprise and new businesses and workloads that startups build too. I think we’re still in the relative early stages of how much demand there’s going to be for AI. I think it’s going to change every customer experience that we know. I think that it will invent all sorts of new ones that we never imagined. I don’t know if the trajectory of that middle part of the barbell will be the same wildly steep trajectory that we’ve seen with the current barbell AI labs piece.

It seems that most of AWS’s contracts do not have built-in protections for cost inflation

[Question] Could you talk about how your pricing strategy at AWS incorporates future cost inflation? Do your longer-term contracts allow for stable return profiles despite cost inflation?

[Answer] What I would say is that most of the deals that you sign, there’s a certain amount of your demand that is on demand, where there aren’t contracts. A large amount of it tends to be deals and agreements that you’ve signed. The deals that you sign, those will be the prices and those will be the agreements that we have over the duration of that contract. New agreements that you sign, you always take into account what your costs are and how you ultimately build a price that you agree to with your customers. I think it’s no secret right now to any company in the world that there are inflated prices right now on some of the components like memory and hard drives and SSDs.

Apple (NASDAQ: AAPL)

Apple’s management recently unveiled the new Siri AI, and they are thrilled with the response from early users; Apple’s work on Apple Intelligence are done in a way that’s personal and private, with AI models that are running on-device and on servers using private cloud compute; management thinks Apple’s differentiating factor with AI is its massive unified memory bandwidth, industry-leading power-efficient performance, and deep on-device intelligence; management thinks Apple’s products are the best hardware for users to experience AI; Apple recently unveiled new AI-powered accessibility features; one AI-powered accessibility feature is for power wheelchair users to control drive systems with just their eyes, using the Apple Vision Pro; Siri AI had a public beta a few weeks ago, and the feedback has been great; Siri AI is private, based on a user’s personal context, and integrated into iOS; management is unsure what the compute costs will be for the new Siri AI, but they do see upgrade possibilities on iCloud+ for heavy users; management is still unclear if the introduction of Siri AI will result in a step-change in Apple’s cost structure; Siri AI’s initial roll out will not include China and Europe for regulatory reasons, but management is working with the relevant authorities to solve the problems

This year’s WWDC was a wonderful showcase of our latest innovations. We were tremendously excited to unveil the all-new Siri AI, a completely reimagined version of Siri that is profoundly capable, deeply personal, and integrated seamlessly across our platforms. We’ve been absolutely thrilled by the response from people who’ve been using Siri AI in the developer and public betas. The reviews from early users have been phenomenal, it’s been so wonderful to hear from people who are excited about the capabilities we’ve built. It underscores our philosophy that building AI that is private and based on personal context can change how users find information and get things done with our products in a way that truly enriches their lives…

…We’re excited about the work we’re doing on the next generation of Apple Intelligence, including Siri AI and the AI features we’re developing across our platforms. These experiences are intuitive and useful, while also deeply integrated in a way that’s personal and private with the latest models running on-device and on servers using private cloud compute…

…What sets Apple apart is the unique combination of massive unified memory bandwidth, industry-leading power-efficient performance, and deep on-device intelligence, all built around the customer experience from the ground up. The result is that Apple has created the world’s best hardware to experience AI, whether using Apple Intelligence, including Siri AI, or third-party offerings…

… In honor of Global Accessibility Awareness Day, we unveiled new features to help users get more out of the products they use every day. New intelligent capabilities are coming to VoiceOver, Magnifier, Voice Control, and Accessibility Reader to make them more useful and intuitive. We’re also using on-device speech recognition to generate subtitles for video content without captions. Apple Vision Pro is adding a feature for power wheelchair users to control drive systems using just their eyes…

…[Question] Just on iOS 27 and Apple Intelligence, went into public beta earlier this month. Could you talk about learnings from the public beta? Will the new Siri AI be a demand driver for iPhones this holiday?

[Answer] We released it to the public for a public beta a few weeks ago, the continued feedback is really, really great. I think it’s a very big idea to have AI that’s private, that’s based on your personal context, and that’s integrated across the operating system… In terms of what it means for compute cost, it’s obviously early going for us. I don’t want to say that we have a complete plan for that. We do believe there will be people that want to use it a lot. We will have some kind of upgrade possibilities on iCloud+ where people can buy up the stack on iCloud+. We’ll see how the pickup for that is…

…[Question] Is it right to think that the capital intensity of Apple will change in the future because of Siri AI?

[Answer] We use some third-party cloud, and we do our own data centers. There will be a mix. Generally speaking, as you know, we have been growing our OpEx and spending more in AI in general and quite a bit more. There are other locations on the P&L other than OpEx, like COGS etc., that also have AI expenditures. We’ll see what Siri AI does from the cost side of it. There’s also the ability when people use it a lot for them to move up on an iCloud Plan as well. What the balance of that is a bit uncertain at the moment…

…[Question] When you announced Siri AI, you also did mention, along with the rollout, that probably we won’t have the initial rollout in China and Europe. Just wanted to get your updated thoughts on that front.

[Answer] You look at the EU, we’re working closely with the commission. Obviously, our complete desire is to launch everything everywhere at the same time. That’s always the philosophy that we have. We have not been able to do that in the European Union, but we’re working closely with them to try to get to something that would allow us to offer Siri AI there. It is offered or will be offered for the Mac there, because the Mac is not covered by the same regulations as the iPhone and the iPad. Net-net, we’re working with them and hope to reach some sort of solution. You look at China, last week we received approval to ship sort of the original features of Apple Intelligence, things like cleanup and so forth. We’re working now through the rollout of those, and there will be more work required down the road for Siri AI. We’re at the front end of that.

Apple’s management thinks the Mac is an AI powerhouse, with excellent on-device inference and creation capabilities; management is seeing customers use the Mac Mini for agentic AI, and deploy clusters of Mac Studios to run frontier models locally; more companies are choosing Macs for their on-device AI advantages, including Disney and Credit Agricole; Credit Agricole is using on-device AI on MacBook Pros to reduce manual processing time of regulatory workflows by 80%

Mac delivered its best June quarter yet with $10.4 billion in revenue, growing an impressive 29% from a year ago despite significant supply constraints. This revenue growth was driven by the incredible strength of our latest lineup with MacBook Pro and the all-new MacBook Neo. According to IDC, we gained share globally. We also set a June quarter revenue record in developed markets and an all-time record in emerging markets with particular strength in Greater China, where we had an all-time revenue record. In addition, we achieved all-time records for upgraders and customers new to Mac.

With the power of Apple silicon, the Mac lineup delivers outstanding power-efficient performance, massive memory bandwidth, and next-level AI capabilities. Mac continues to be the ultimate AI powerhouse, excelling at high throughput, on-device inference, and creation across a broad range of AI workloads. We’re seeing customers increasingly put those capabilities to work, from using Mac Mini as a powerful platform for agentic AI to deploying clusters of Mac Studio systems to run frontier class models locally…

…More companies are choosing Mac for on-device AI advantages, including lower costs, better performance, and enhanced privacy and security. At Disney, creative teams are increasingly turning to Mac for on-device AI workflows that reduce overall cloud token costs and keep their IP secure. Crédit Agricole, France’s leading retail bank, is using on-device AI on MacBook Pro to streamline regulatory workflows, reducing manual processing time by over 80%.

Users of AirPods are using live translation, powered by Apple Intelligence

Meanwhile, we continue raising the bar across our AirPods lineup, whether it’s the immersive listening experience of AirPods Pro 3 or the premium listening experience and exceptional active noise cancellation of AirPods Max 2. With live translation powered by Apple Intelligence, people are crossing language barriers and connecting like never before.

Apple’s management will reinvest tariff refunds into the US; Apple recently announced a new agreement with Broadcom to design custom silicon and wireless connectivity technologies; management expects the new agreement with Broadcom to exceed $30 billion, and it is part of Apple’s $600 billion manufacturing commitment to the US, and the largest commitment to-date; Apple will soon open the Apple Advanced Manufacturing Center in Houston; the Apple Advanced Manufacturing Center is in a facility where Apple is currently assembling AI servers and will soon manufacture Mac Minis; the Apple Advanced Manufacturing Center will be imparting the processes Apple uses to make its products to strengthen the USA’s entire advanced manufacturing ecosystem; Apple will be sourcing 100 million components from TSMC’s Arizona fab in 2026 and management is really pleased about the fab; Apple’s use of the TSMC Arizona fab is part of Apple’s $600 billion commitment

 Last year, we made a $600 billion commitment to the U.S. over four years, and now, as we said before, we plan to reinvest the tariff refunds we’ve received into the U.S. We’re pleased with the progress we’ve already made advancing the American supply chain. Earlier this month, Apple announced a new agreement with Broadcom to design and produce custom silicon components and cutting-edge wireless connectivity technologies. The new multi-year agreement with Broadcom, which is part of Apple’s American Manufacturing Program, is expected to exceed $30 billion. This marks our largest-ever American manufacturing program commitment. It’s also an important step forward in our work to build an end-to-end silicon supply chain here in the U.S.

We’re excited for the upcoming opening of the Apple Advanced Manufacturing Center in Houston. The center is located in a facility where we currently assemble advanced AI servers. Later this year, we’ll make Mac Mini there, too. The center will teach students, supplier employees, and business of all sizes the same innovative processes we use to make our products. The goal is to empower American manufacturers to take their work to the next level and strengthen the entire advanced manufacturing ecosystem…

…In Arizona, we do source over 100 million components this year out of Arizona, it is part of our $600 billion commitment to the U.S., and we could not be more pleased with how that fab has ramped and is producing for us.

Apple’s management thinks it will be a great idea if there were more memory chip suppliers; management had to reluctantly raise the prices of Apple products to deal with a 100-year flood on memory pricing; it’s still unclear to management if Apple’s price hikes will impact demand

In terms of the sources of supply, primarily the DRAM market has three suppliers. Obviously if there were more suppliers, that would be good, and it would help us on the supply side and perhaps the pricing side. It’s unclear on the pricing side, it could help on the supply side…

…We reluctantly raised prices, I would say. We did it because we’re in what I would characterize as a 100-year flood on the memory pricing, with exponential increases in memory prices…

…Obviously, we’ve now had to increase prices on iPad and Mac — and the price elasticity there, it’s just too early to come to a definitive conclusion of what happens there, because it takes a little while for the channels to adjust since there’s channel inventory, and it takes a while for the consumer to respond. And so we’ll understand that more in the weeks ahead.

ASML (NASDAQ: ASML)

ASML’s management is seeing strong end-market demand that is motivating its customers to raise their capex; management is keen to support its customers’ demand; management is seeing both Logic and DRAM customers entering long-term agreements with their customers and having unprecedented visibility on future demand; management is seeing Logic customers add a lot of capacity on existing advanced nodes because of AI-related demand; management is seeing Logic customers aggressively ramp the 2nm node and even start ramping the 1.4nm node; the dynamics in the Logic segment are driving both an increase in litho intensity and more demand for litho; management sees a clear need for more supply of memory chips; management is seeing Memory customers accelerate capacity plans; management is seeing advanced Memory nodes calling for higher lithographic intensity, for both EUV (including low-NA EUV) and DUV (immersion belongs to DUV); management is seeing a perfect storm for ASML on DRAM (memory chip) in 2026 and beyond

The end market demand, this has motivated our customers to increase their CapEx but also accelerate all their plans. This really creates a need for more systems basically starting this year. We are doing the same on our side, extending output, extending basically our teams so that we can support them moving forward…

…What we also see for both Logic and DRAM is that our customers are getting long-term agreements with their own customers, which really invites them to commit for the long term. Because they have what I would call a quite unprecedented visibility on what will happen to the market…

…Quite a bit happening with Logic. So first, if we look at the existing advanced nodes 5nm, 4nm, 3nm, we see that our customers are trying to add a lot of capacity there now. This is because there is a huge demand on those technologies coming from AI. At the same time, the 2nm ramp is done as aggressively as possible. We see customers adding capacity, accelerating their plans and even start basically to look at the 1.4nm ramp…

…These dynamics in the logic segments are driving both an increase in litho intensity and greater demand for advanced lithography…

…If we look at the price of Memory today, either for DDR or for HBM, there is a clear need for more supply. This is translating into, again, acceleration of capacity plans from our customers. So this is happening with all customers. On top of that, as we discussed previously, the latest nodes are calling for more litho, for higher litho intensity, both on EUV but also advanced immersion…

…DRAM lithography intensity is rising as customers migrate to advanced nodes. This includes both EUV and deep UV immersion with EUV Low-NA growth driven by the increased replacement of multi-patterning with more cost-effective single-exposed EUV…

…HBM will require more wafers. So there’s a volume effect again. So that’s one element. The second element is, of course, the number of EUV and immersion layers, which has increased basically on the nodes that are ramping very, very strongly right now. So the 1c node, for example, which is going to be an enormous node, or even 1b are using more EUV layers. So this is really this combination, which creates a bit the perfect storm for ASML on DRAM this year and most probably the next few years to come.

ASML’s management is having very constructive discussions with customers for long-term business; ASML has already received nearly all the EUV orders needed for 2027, despite the company adding 30% more EUV capacity for 2027 compared to 2026; ASML has already received large orders for EUV for 2028, and this has driven ASML to increase EUV capacity in 2028 by another 30%; the capacity additions for EUV for 2027 and 2028 are for low-NA EUV systems; ASML can achieve the capacity additions by optimising existing cleanroom space, and has space to increase capacity by even more if needed; the 30% capacity increase in EUV for 2027 refers to tools, while the actual wafer capacity will be increased by 45%; the cleanroom optimisation does not require ASML to sacrifice any of its high-NA EUV supply; management thinks it’s possible that ASML’s anticipated capacity additions will increase further; the current planned capacity addition for 2028 is an act of preemption by ASML’s management, but they are based off strong customer signals; management sees plenty of operating leverage to come down the road for ASML

We are having very constructive discussions with our customers on the long term. Their own visibility to their business allows them to share with us also, longer than, I would say, usual visibility on their business. We are talking of course about next year, but also beyond that. Now this has also translated practically into very strong order bookings through the first half of 2026.

If we look into more detail starting with 2027, there we are pretty much already close to receive all the EUV orders we need for 2027. This is with us adding about 30% capacity for EUV in 2027 versus 2026. When we look at 2028, we have received already a large number of orders from our customers for EUV. This has also invited us very strongly to investigate another 30% increase in our EUV capacity for 2028. Now, of course, when EUV grows DUV grows as well. Immersion is going to be important. Also for 2027 and for 2028 we are going to look into a 30% increase of our capacity for both years…

… For 2027, we are now close to being fully covered with orders for Low-NA EUV, and we are planning to increase our Low-NA EUV capacity by around 30%. Looking ahead to 2028, we have already received a significant number of Low-NA EUV orders. Strong demand forecasts from our customers have led us to investigate a further 30% capacity increase for that year…

…[Question] The 30% increase in ’28, which implies 110 tools. And I mean, Christophe, you mentioned investigating. The word is precise. Do you need a new cleanroom for that?

[Answer] The number we are mentioning, we can achieve basically by optimizing the existing cleanroom space in the right way. So this is also why we can create basically that improvement in the short term…

…The balance between demand and supply as we see it today gets us to the 30%, right? So that’s the way we do it. If customers are going to come to ASML and say, “Hey, ASML, we need considerably more.” Then just as we’ve been doing it in the past couple of months, we need to look ourselves in the eye, we need to look at all the supply chain and just see what can further be done…

…When it comes to EUV in particular, right, the tool mix that we’re going to ship next year will be Es and Fs, while this year, it’s a combination of Ds and Es. And if you recognize the difference in output, then in essence what you’re looking at is not 30% improvement of wafer capacity that we’re adding, but approximately 45%…

…The optimization I was referring to before is really basically across all products. Now I think everyone understands that, of course, a lot more is being done today on Low-NA and immersion, for example, than High-NA. But we are not sacrificing, I would say, any of our High-NA supply by doing the rest of the optimization…

…[Question] The capacity increase to 85 and 110 units, you are meeting the demand, not undershipping. Is that correct?

[Answer] I think that as you have noticed in the last few months, I don’t think we have reached yet a stable state on what the demand will be for ’27, certainly not for ’28. So we keep on revising basically with our customers what that demand is. And again, the whole goal of our supply is to follow that demand. So I would not say that we are done with this discussion…

…We’re investigating the 110 scenario for EUV in Low-NA by 2028. Of course, we don’t have orders for 110 EUV Low-NA at this stage. So we’re not waiting. We’re preempting…

…The demand signals that we’re getting from customers also when it comes to ’28 are sufficiently strong for us to seriously investigate this 110 number and the related number on immersion that we signaled to you…

…In the past, we increased the headcount of R&D quite substantially. I would say that today, we believe that with the team that we have today, we can really entertain a very aggressive roadmap going forward. So all in all, I think you will continue to see us manage both R&D and SG&A quite nicely. And as a result of that, the operating leverage that you imply, I think the operating leverage will indeed become better in the quarters and the years to come.

Intel is now using ASML’s high-NA EUV systems in production for its most advanced nodes; management expects to soon enter discussions with all of ASML’s customers on how/when high-NA EUV systems will be inserted into their high-volume manufacturing flows; the maturity of the high-NA EUV systems is maturing to the level required for high-volume manufacturing; management believes that the high-NA EUV system, at single exposure, will bring cost benefits to customers; management thinks the high-NA EUV system’s cost will have more advantages over low-NA EUV once the high-NA EUV system matures; management thinks the partnership with Intel is the clearest example of the progress high-NA EUV is making in terms of its maturity; Intel was the first to get the high-NA EUV systems, so it became the first to implement in production; management thinks both logic and DRAM are good candidates for high-NA EUV systems; there’s very little fungibility in the optical tools used in low-NA EUV and high-NA EUV systems 

As you may have read it, Intel is basically now using High NA in production on their most advanced products. So it means that some of the products you buy today from Intel have been created with an High NA machine. So this is, of course, a very important milestone. This is the proof of the maturity of the tool. We talked a lot about that in the last quarter. We are seeing that happening with all customers and therefore expect to enter that discussion with all our customers on how exactly and when exactly the tool will be inserted in high volume manufacturing…

…Intel Foundry is using ASML High-NA EUV technology on the Intel 18A process node to produce a subset of its Intel Core Ultra Series 3 processors…

… We are continuing to work very closely with our customers to prove the value of High-NA technology for their process technology road maps. In parallel, the maturity of the platform is improving towards the level required for insertion into high-volume manufacturing…

…Every new generation of lithography system ASML ever brought to market was with a strong intention to reduce the cost of patterning. So when you look at High-NA single expose, the design of the tool, the performance of the tool will be such that it provides a cost benefit to our customer…

…We are still basically working on bringing the High-NA platform to the level of maturity of Low-NA. And when you achieve that, this is practically the time where the cost of High-NA is going basically to provide an advantage versus the existing technology…

…I think the key again for High-NA to be cost-effective, to beat the cost of Low-NA plus immersion multi-patterning is to bring High-NA to the right maturity…

…We are very happy with the press release this morning about Intel, because this is, I would say, maybe the strongest sign so far that we’re getting there…

…We talked about Intel today. I think you know that Intel was first to get the technology. So they are first to implement it in production…

…The opportunity for DRAM is significant also because the volume is also significant. But there’s no real change there. I think we still see both logic and DRAM being a good candidate for High-NA. The reason for that is both DRAM and advanced logic will be shifting more and more towards multi-patterning Low-NA over time. So that applies to both…

…In terms of the fungibility of equipment for High-NA and Low-NA, particularly when it comes to ZEISS, because that’s the way, C.J., I interpret your question, that really isn’t there. It’s totally different tools that you need that ZEISS needs to produce a High-NA optic versus a Low-NA optic. So it’s not that there is fungibility that you can use High-NA tools to get more Low-NA output.

Intel (NASDAQ: INTC)

Intel’s management sees demand outpacing supply for the company (despite growing its supply) as well as the semiconductor industry; the supply shortage is expected to last for some time; management sees a rapid and sustained build-out of compute infrastructure for the semiconductor industry; Intel’s customers are signalling a strong and sustainable spending environemnt 

Strong demand for our products continue to outpace our growing supply…

…Industry is facing one of the most severe supply constraints in its history across leading-edge logic silicon wafers, memory and substrates. These shortages will persist for the foreseeable future…

…Intel is uniquely positioned to benefit from the overwhelming demand for compute as the entire industry continue a rapid and sustained build-out of compute infrastructure…

…Customers continue to signal a strong and sustainable spending environment driven by the unprecedented demand for AI compute. Industry-wide supply constraints across wafers, memory and substrates remain the dominant challenge our customers are facing to support the AI infrastructure build-out.

Intel’s management sees the transition to agentic AI systems driving CPU density higher in AI data centers; management is seeing customers place growing importance to Intel’s x86 CPUs in AI data centers; Intel’s Xeon 6 server CPU product continues to be one of the company’s fastest ramping products ever; Intel’s design services business (where Intel helps customers build custom AI chips) saw revenue grow 3x year-on-year in 2026 Q2, driven with the help of the x86 CPU franchise; management’s outlook for server CPU demand has increased yet again and management expects strong growth for the industry for 2026-2028; management thinks the CPU-to-GPU ratio in AI systems is now at parity, and could even skew towards CPUs in the future; management sees significant growth in the CPU market

As AI expands from training to inference and increasingly to agentic and multi-agent systems, general purpose server CPU density continue to increase, and our core server CPU franchise is growing faster than ever…

…Demand accelerate across cloud and enterprise as customers increasingly recognize the critical role that CPUs in general and x86 CPUs, in particular, play in the AI infrastructure…

…Xeon 6 continue to be one of the fastest ramping products in Intel history, reflecting improving execution and strong customer demand…

…We continue to make steady progress in our newly announced design services business with revenue growing nearly 3x year-over-year. We see tremendous opportunities to leverage our strong x86-based general purpose computing franchise to build more purpose-built computing products for the AI era…

…Our outlook for server CPU demand has improved again since our last earnings report and we’re forecasting strong double-digit unit growth for the industry this year and next, with momentum extending into 2028…

…Lip-Bu has talked in the past about the ratio of CPU to GPU going up. And we now believe we’re almost in parity at this point and could eventually even skew more to CPUs on a unit basis…

…From all the inputs we’re getting from our customers in terms of the level of spend and also the long-term agreements we put in place and the visibility we’ve gotten, we feel like the growth is going to be significant. 

Intel’s management is more confident than ever of the company’s foundry roadmap for leading-edge chips; in 2026 Q2, Intel’s leading-edge 18A node exceeded internal volume and yield expectations, and saw output increase meaningfully; Intel is now ramping multiple new products on the 18A node while supporting demand for existing products; the ramp up of 18A for Intel’s internal products provides validation as the company engages external customers; management has started risk-production for the 18A-P node, which has performance and power improvements over 18A; Intel’s 14A node is outpacing the 18A node’s development; customer engagements for 14A has increasing momentum and management is increasingly confident that the node will be a highly competitive node in performance, power, density, cost, and schedule; Intel remains on track for 14A risk production for internal products in 2027 H2; management has decided to fully commit to a high volume ramp for 14A in 2028; the 18A node is now in volume production across commercial and consumer products in the PC client segment; the 18A node’s output was 25% over target in 2026 Q1, and up 50% sequentially; nearly all of Intel Foundry’s revenue in 2026 Q2 was from internal demand; Intel Foundry has lowered the cost of Intel’s Panther Lake SKU by 50% year-to-date in 2026 Q2, with further reductions expected for the rest of 2026 and 2027; the 18A node’s quarter-to-date yield for 2026 Q3 are ahead of targets set earlier this year 

My confidence in our foundry process road map has grown significantly since joining over a year ago. I am more confident than ever of the strategic in significant and unique value proposition of Intel Foundry. During Q2, our factories across Intel 7, Intel 3 and Intel 18A exceeded internal volume targets, driven by improving yields, better cycle times and increasing wafer starts. 18A output increased meaningfully in the quarter. Yields continue to track ahead of expectations. We are now ramping multiple new products on 18A, while supporting growing demand for our lead products, including Panther Lake and Wildcat Lake. I keep raising the bar on the internal targets, and the team continues to meet the challenge. The successful volume ramp of 18A for our internal products provide important validations as our Intel Foundry engaged with external customers. We also began risk production of 18A-P, providing additional performance and power advantages, while maintaining IP and design compatibility with Intel 18A, positioning 18A-P as a competitive node for external customers…

…I’m encouraged by our progress on Intel 14A. Defect density and transistor performance are all outpacing 18A development. PDK 0.5 is now complete, and PDK 0.9 is on track for October. We continue to build out and validate the IP portfolio for 14A as we position the 14A family for broad-based adoptions across a wide range of customers. I’m pleased to see the increasing momentum on customer engagements for Intel 14A, and I’m increasingly confident that the 14A will be a highly competitive process offering across key vectors of performance, power, density, cost and schedule. With encouraging external customer progress and increased demand for our internal products, we remain on track for 14A risk production for our internal products in second half of 2027, and we make the decision in Q2 to fully committed to high-volume ramp in 2028…

…In our core PC client segment, Intel 18A is now in volume production across multiple commercial and consumer products. Our factory output continues to increase sequentially every month. The successful high-volume ramp of 18A for our internal products provides important validation as Intel Foundry engaged with external customers…

…Our Client Group has now brought 18A to full scale with 400-plus designs for Series 3 across consumer and commercial…

…Turning to Intel Foundry, revenue of $5.8 billion was up 6% sequentially on higher fab volumes driven by strong growth in Intel 18A with output approximately 25% above target and up more than 50% quarter-over-quarter. External foundry revenue was $293 million in the quarter…

…Intel Foundry has driven down the cost of our primary Panther Lake SKU by roughly 50% year-to-date and is on track for an additional 20% this year with further meaningful reductions planned in 2027…

…Our wafer output across our major nodes exceeded expectations from 90 days ago, and Q3 quarter-to-date 18A yields are trending ahead of targets set in March.

Customer interest in Intel’s advanced packaging technology, EMIB-T (Embedded Multi-die Interconnect Bridge with Through-Silicon Vias) remains very high; EIMB-T has a growing backlog, and yields are hitting targets; management wants to ramp EMIB-T into high volume production in 2027

On advanced packaging, customer interest for EMIB-T continue to be very high. Technology is compelling, providing capabilities for advanced AI silicon solutions, which are not possible with today’s mainstream offerings. We continue to have a growing EMIB-T backlog, yield and reliability are hitting targets, and we focused on ramping the technology into high volume and high quality to support customer ramps in 2027.

Intel’s management sees a growing opportunity for AI with edge devices; management thinks the market for edge and physical AI will match that of enterprise AI; Intel had 130 Series 3 design wins for edge AI in 2026 Q2

We recently renamed our PC business to our Client Computing and Physical AI Group or CCPG. We did this to recognize the growing opportunity for AI at the edge…

…We expect enterprise adoption of AI to be a long-term tailwind for CCPG, but our AI-driven market prospects don’t stop there as the edge and physical AI opportunity is likely to at least match the client TAM over time. CCPG showcased this growing opportunity with 130 Series 3 design wins for edge AI applications, including brain and control deployments for robotics.

Intel has a multi-year collaboration with SambaNova, an AI inference chip startup

We also extended our heterogeneous AI strategy through multiyear collaboration with SambaNova. We are pleased with their growing momentum as we work with them to drive performance and power improvements with disaggregated inference.

Intel’s management thinks the company is the only one that can build the whole of compute solutions for AI, from traditional CPUs and GPUs, to custom AI chips

We are the only company that can design, manufacturing, build the entire range of computing solutions from general purpose, traditional CPUs and GPUs to more purpose-built ASICs and CPUs optimized for agentic AI, as we increasingly move from compute dominant by system on chip towards system in package.

Intel’s vPro mangeability software had 1,500% growth in the last 4 quarters, demonstrating that manageability and enhanced security are critical for agentic deployments

On the commercial side, activations for our market-leading vPro manageability software have surged 1,500% over the last 4 quarters, underpinning that manageability and enhanced security are critical must-haves in the agentic workplace.

Intel’s management now expects capital expenditure to be more than $20 billion in 2026, which is up from the previous expectation of around $18 billion; management is aggressively locking in tool purchase orders from vendors; management expects Intel’s capital expenditure in 2027 to be much higher than 2026’s level; Intel’s total capital expenditure on tools and space in the U.S. is approaching $100 billion for 2021-2026; management is committed to have capital expenditure match customer demand; Intel’s capital expenditure is fairly broad-based and includes advanced packaging, although more is allocated for front-end manufacturing than advanced packaging; the higher capital expenditure for 2026 is a signal of management’s confidence in Intel’s customers, and the higher visibility management now has on the long-term demand outlook; management sees significant returns from Intel’s capital expenditure; Intel’s external customers became excited when they started to see the 14A node’s PDK (process design kit yield, and this gave management comfort to invest more in capital expenditure; most of Intel’s capital expenditure in 2026 will be for tools for Intel’s leading edge nodes; management thinks Intel has sufficient liquidity on the balance sheet for now to fund its capital expenditure, while prepayments from customers are also helpful, although the company may need to tap the capital markets if the business becomes really successful and more capital expenditure is needed to fulfil demand; management is willing to commit capital expenditure only if there are customer commitments

Due to strong customer demand signals, we’re raising our outlook for 2026 and now expect our CapEx to be more than $20 billion, which is up significantly versus our expectations entering the year. We’re also aggressively locking in tool purchase orders from our vendors, accelerating our clean room build-outs and actively securing supply of substrates and memory. As a result, we’re forecasting 2027 capital expenditures to be significantly above the 2026 levels with the vast majority spent across our U.S. network. In fact, as we look back from 2021 through 2026, our total capital spending in tools and space in the U.S. is approaching $100 billion, significantly higher than any other semiconductor company over that time frame. We remain committed to tightly matching our expenditures with customer demand and remain financially disciplined as we capture the growth ahead…

…The CapEx is fairly broad-based. It’s going to include advanced packaging. . As Lip-Bu talked about, we’re pretty excited about our prospects on EMIB-T. And so we will be investing in that. That said, the cost of a fab for the front end is much more expensive than a packaging facility. So it will be skewed towards the front end…

…This increased investment is a signal of our confidence in customers across all of our business units. We feel very confident, particularly in places where we’ve gotten long-term agreements that we now have the signal to be able to kind of forecast out what the outlook looks like for the next few years in terms of demand, and we’re putting forth the capacity in anticipation of that across all of our businesses…

…That’s why you see the CapEx going up next year. But over time, they generate significant return. And particularly now as we migrate towards a model where we keep these processes on longer, the returns are quite significant…

…The engagement with the customer, the feedback have been very positive, tremendous demand for our own products and also external foundry customer engagement that give me the confidence the moment they’re starting to see the 0.9 PDK the yield, they’re starting to get excited about what kind of product they want to run that and how much capacity we can provide them. So those are very positive signs that they are really serious about going forward. And that’s why, as I mentioned earlier, I don’t put CapEx unless I see the yield performance, the IP is ready to serve the customer and also customer engagement, the level of engagement I see…

…Most of the CapEx dollars are going to at this point is tooling. We’ll increase tooling in ’26 by 40% relative to ’25. So we’re investing a significant amount in tooling. And it’s where you might expect, it’s Intel 3, it’s 18A, it will be 18A-P…

…[Question] As you are planning these investments for the back half of the year and into next year, how are you thinking about the balance sheet?

[Answer] We feel like we’re in a really good place from a balance sheet perspective. We have over $30 billion of cash. We have a $10 billion revolver. So we’ve got $40 billion of liquidity… We have seen, by the way, our customers willing to invest with us. And we’ve had prepays from customers that we’ve been — that has enabled us to unlock capacity that’s helped us. That said, if we’re super successful, which we’re driving to, we may need to tap the capital markets to drive some more investment…

…We’re just going to be very careful around making bets ahead of customer commitments. I think that’s the most significant change with Lip-Bu is until we really know that we’ve got the customers, we don’t want to put a significant amount of capital.

Intel’s management thinks the ASIC (application specific integrated circuit) for AI business has a market opportunity of more than $100 billion; management expects the annualised revenue run rate for Intel’s ASIC for AI business to increase rapidly

[Question] On the ASIC business. I guess based on what was disclosed last quarter, it’s about a $1.2 billion run rate business now growing well for the company. How do we think about the diversity of that business? And just you’ve announced Fortinet. I’m just curious of how you’re thinking about the growth profile of that business?

[Answer] This is a massive opportunity. I think potentially it’s over $100 billion TAM market…

…Growth rates of ASICs. I mean I would say that today, we’re probably running at about a $2 billion run rate or at least approaching a $2 billion run rate for that business. We think in the not-too-distant future, we’ll be at a $4 billion run rate for that business.

Intel’s management sees memory chips having a severe supply shortage; Intel is collaborating with all 3 major memory chip companies; management sees memory chips as a bottleneck in AI currently, and so Intel is working on memory, with the recent hire of SK Hynix’s CEO (SK Hynix is one of the 3 major memory chip companies)

Memory becomes a big supply constraint challenge. And we’re collaborating with the 3 big memory vendors. That’s very important to serve our customers as our #1 priority. The next thing, as you recall, Intel has a rich history in the memory. And recently, we hired Seok-Hee Lee to join us. He used to be the CEO of SK Hynix. And clearly, memory, it becomes the bottleneck, a lot of AI infrastructure and pain point for customers. And we’re also looking at how other areas that we can integrate compute and memory and also how the stacking and then how can we use the memory more — utilization more efficiently. So I think there’s a lot of areas we are working on. Stay tuned, and we will work on that. 

Mastercard (NYSE: MA)

Mastercard’s management is seeing higher demand for the company’s security solutions; Mastercard is able to identify threats before they materialise through Recorded Future’s capabilities (Recorded Future was acquired by Mastercard in 2024 Q4 and it provides AI-powered solutions for real-time visibility into potential threats related to fraud); Mastercard Threat Intelligence was launched in 2026 and in its first 3 quarters, it has identified more than 7 million card-testing transactions in 192 countries and prevented $172 million in fraud; Mastercard Merchant Trust Services is a new suite of AI-powered capabilities to help identify fraudulent merchants

We’re seeing increased demand for Mastercard’s robust and unique security solutions as clients navigate the ever-expanding threat landscape. Today, we have differentiated capabilities that span cybersecurity, identity, and fraud…

…We’re identifying threats before they materialize. How? We’re bringing Recorded Future’s market-leading intelligence capabilities to our clients globally and seeing strong engagement across sectors. This quarter, we partnered with Wipro, a global information technology and consulting company, to further scale our capabilities. Building up on our acquisition of Recorded Future, we also launched Mastercard Threat Intelligence specifically for payment fraud. In its first three quarters, Threat Intelligence has identified more than 7 million card testing transactions across 192 countries. Stopping that activity prevented an estimated $172 million in fraud linked to malicious domains. That’s real value to us, our customers, and of course, our cardholders…

…Mastercard Merchant Trust Services is a new suite of AI-powered capabilities to help identify fraudulent merchants. Keeping the scammers from setting up shop will reduce fraud, cut out disputes, and provide greater security. The goal, real transactions for real purchases from real merchants.

Mastercard’s management believes agentic commerce is the next evolution in payments and is a significant opportunity for the company as it leads to incremental transactions and services; Mastercard Agent Pay helps power secure and trusted agentic transactions across Mastercard’s global acceptance network; management expects cards to prevail in an agentic world; early engagement with Mastercard Agent Pay is encouraging; management sees agentic commerce giving rise to a new class of payments in machine-to-machine (M2M) payments; management believes M2M payments expands Mastercard’s addressable market; management recently announced Mastercard Agent Pay for Machines to allow AI agents to purchase low-value digital services at machine speed; Mastercard is currently the only network enabling M2M payments; there were already 30 industry leaders participating in Mastercard Agent Pay for Machines at its launch; management sees agentic transactions in consumer-oriented use cases happening through existing card networks, and these transactions also bring the opportunity for services; Mastercard Agent Pay has a capability called Verifiable Intent, developed together with Google, that allows consumers to challenge a transaction; management sees agentic transactions in B2B use cases happening through existing card networks; management sees agentic transactions in M2M as possibly requiring different types of payment infrastructure, such as stablecoins

Agentic commerce is the next evolution in payments, where the importance of security, transparency, and control only increase. Agentic commerce creates a significant opportunity for Mastercard. It leads to incremental transactions and even more opportunity for our services. Through Mastercard Agent Pay, we’re helping power secure and trusted agentic transactions across our global acceptance network using tokenization, zero liability protections, and unique dispute resolution capabilities. These capabilities are just a few of the reasons why we expect cards will prevail in an agentic world, both in consumer and commercial use cases. It’s early days, but engagement across the globe is energizing.

The rise of agentic commerce also brings about an entirely new class of payment use cases, machine-to-machine payments. This is an expansion of our addressable market and one that we are at the forefront. We recently announced Mastercard Agent Pay for Machines, which enables AI agents to purchase low-value digital services such as APIs, compute, data, content at machine speed. With on-chain permissioning and off-chain settlement, Mastercard is the only network enabling machine-to-machine payments. An ecosystem is rallying behind us. At launch, we had more than 30 industry leaders participating, including Adyen, Ant International, BVNK, Checkout.com, Cloudflare, Coinbase, and OKX. We are a first mover in this space and one with credibility and trust to deliver…

…There’s the consumer-oriented use cases in agentic commerce where our keyword search turns changes, and we may use agents for that. That could be an LLM, that could be a first-party agent by a large retailer. There are transactions that are now then delegated to agents, and that can happen very well through the existing card networks. What you need for that is, that’s what merchants always need. They need reach, they need predictable user experiences. That’s the same, that’s true for consumers. We really believe that cards will prevail in that world. This is a tremendous opportunity for us also on the services side through tokenization, inside tokens and so forth…

…We have additional capabilities that we put into the Agent Pay protocol from us, and one of them is Verifiable Intent, which allows you to basically challenge a transaction, say, “I never wanted to buy this,” and then the chargeback process can kick back in. This was innovated together with Google…

…If you look on the B2B side, you can see there’s a range of agentic commerce transactions that can happen where you have an agent that does purchasing for a company. That can very well happen on the card ecosystem, very similar to what I just said on the consumer side. Amounts, speeds, purposes, they will need the protections, they need the global reach, all of that can apply, and we believe that’s a continued opportunity for us, particularly on the services side again…

…Machine-to-machine payments. That is low-ticket, micro-ticket transaction that happens at very high velocity. For that, we can see a world emerging where different kind of underlying infrastructure is required. For that, we’ve put out our protocol, which is an evolution of Agent Pay, which is Agent Pay for Machines. This is the only network protocol that’s out there today to facilitate that. Now, the underlying infrastructure for that, we’ve mapped it out. You can start to see that there is a transaction that is recognized from one agent to another, these machines talking to each other, but the settlement happens through different kinds of rails. That could involve stablecoins, but it also could involve different types of settlements. We’re actually quite open to that. What it needs is the immediacy of these agents to recognize that transaction, and that is what Agent Pay for Machines actually does.

Meta Platforms (NASDAQ: META)

Meta’s management sees the company’s AI investments accelerating all parts of the core business; the AI investments are improving users’ experiences, performance for advertisers, and the speed of shipping new experiences; management is optimistic about the integration of LLMs (large language models) into Meta’s recommendation systems for Instagram and Facebook because the models the company to show more relevant and engaging content; management thinks the new Muse Image and Muse Video models will lead to a huge expansion of new and personalised content to show people; Meta is using LLMs in its advertising systems to (1) improve the prediction and ranking of ads that are shown, and (2) drive significant increases in advertising relevance and conversions on both Instagram and Facebook; management sees clear signs of Meta’s AI investments paying off because Meta’s advertising business is reporting faster year-over-year revenue growth than any other competitor; 9 million small businesses are now using at least one of Meta’s Gen AI advertising creative tools (was 8 million in 2026 Q1); management is rolling out end-to-end creative solutions for advertisers and Muse Image will be important for this as it can produce better ad variations; Muse Image is getting great feedback so far; management has launched Meta One, a subscription that provides more tools and AI features; management will explore different tiers and pricing options for Meta One as demand grows; recent new experiences Meta has launched include Instagram Instants, Forum, and Seller; management will use recommendation systems to scale new experiences; adoption of Image Generation more than doubled among advertisers in 2026 Q2

We are now at a point where our investments in AI are accelerating every major part of our core business. They’re improving the experience for people using our apps, driving better performance for advertisers, and helping our teams build new experiences and ship faster…

…In Instagram and Facebook, I am very optimistic about our work to integrate large language models into our recommendation systems. LLMs add a first principles understanding of what the content is about and why it is compelling, as well as a deeper understanding of what people are interested in and what their goals are when they are using our apps. This means that we can show more relevant and engaging content that better reflects people’s goals and interests. Our new Muse Image and Muse Video models will also dramatically expand the universe of content that people can discover across our platforms. There are already two large sets of content to draw from. First, from your friends and the people you follow, and second, from creators that you don’t follow. Now there’s going to be a whole new and nearly infinite universe of personalized content. This is going to make our services a lot more useful and engaging for people.

For ads, we are using LLMs to improve how our systems predict and rank the ads that we show. We’ve expanded the context that we can take into account around a person’s organic and ads activity to determine an ad’s relevance, driving significant increases in relevance and conversions on both Facebook and Instagram. On a dollar basis, our ads business is reporting faster year-over-year revenue growth than any other company’s reported ad business. These AI investments are paying off.

We are also seeing a lot of demand for our new AI-powered creative tools. 9 million small businesses on our platforms are now using at least one of our AI ad creative tools, and we’re rolling out new end-to-end creative solutions that help advertisers translate performance data into their creative decisions. Muse Image is going to supercharge this. The model can analyze images, improve its own work, and produce better ad variations based on advertiser input. We’re getting great feedback on this so far.

We also just launched Meta One, a new subscription offering that provides more tools and AI features across our apps. As demand grows, we’re going to offer a variety of different tiers and pricing options there as well…

…Earlier this year, we shipped Instagram Instants. We also just launched Forum, a standalone groups app, and Seller, a standalone marketplace app. I expect it to become a lot easier to ship new apps, we are planning to build out more ideas and use our recommendation systems to scale them to the people who will find them interesting, as we’ve done with Threads…

…Image Generation, which now lets advertisers produce more creatives at scale from existing content, including a new ability to create images from video assets, saw adoption more than double this quarter. 

Meta’s management is building new personal agents that will be the foundation for new products and revenue streams in the future; Muse Spark will be a base for Meta’s personal agents; management thinks personal agents will need to work right out of the box, compared to enterprise agents where software engineers are more willing to make them work; management thinks WhatsApp will become an important surface for individuals to interact with multiple agents; WhatsApp is already the leading surface where people engage with Meta AI; management wants to deliver a private and secure AI experience for individuals; Meta recently launched incognito mode on WhatsApp and the Meta AI app for users to have private conversations with their assistant; management believes personal agents will be a massive market a few years down the road

We are developing new personal agents that will be the foundation for our next wave of products and revenue lines in the months and years ahead…

…One reason that we are so focused on making Muse Spark great at agentic capabilities is that we think that there’s a very big opportunity to ship a few types of agents that are aligned with our mission and business. The first is personal agents. Soon, we will have agents that can work 24/7 on your behalf to help you achieve your goals and improve your life, your health, your relationships, your finances, whatever you want…

…The first domain that agents have really taken off in is coding. Engineers are more technical and willing to spend time making those agents work. To build great personal agents, this needs to be a great consumer product that just works out of the box and is easy enough for billions of people to adopt and use…

…As we move towards a future where we’re all interacting with multiple agents, I think that WhatsApp and our other messaging surfaces are going to become increasingly important. WhatsApp is already the leading surface where people engage with Meta AI. As we build out a platform for more agents across our messaging apps, we’re going to innovate on how to deliver a private and secure AI experience. We launched incognito mode this quarter on WhatsApp and the Meta AI app, allowing people to have private conversations with their assistant that even Meta can’t see. We’re planning to make strong privacy and security a fundamental part of the agents that we’re building as well…

…We think that consumer personal agents is going to end up being an extremely important and massive market. I think that it’s extremely unlikely if you look out five years from now, for example, that whatever period of time you want, that you don’t have billions of people with a personal agent that understands your goals and that is just working on your behalf 24/7 to achieve your goals in whatever the domain is that you care about, whether it’s helping you with your health or your hobbies or your personal finances or your productivity in running your home better or improving and enhancing your relationships, helping with your career.

Meta’s management sees a large opportunity to sell to businesses, including business agents and compute; Muse Spark 1.1 is an agentic coding model that is available through Meta’s new public API (application programming interface); management is ramping up distribution of Muse Spark and is making it easier for enterprises to adopt the model; management recently made Meta Business Agents available globally on WhatsApp and Messenger, and will soon roll it out on Instagram; 1 million businesses are already using Meta Business Agents; Meta Business Agents can learn from daily conversations with customers and bring insights to businesses; management’s goal is to build a business-in-a-box service for an individual to start and run a business on Meta’s platforms; management intends to monetise Business Agents through subscriptions and volume-based pricing, and eventually results-based pricing; management thinks results-based pricing will allow Meta to run efficient auctions over the company’s compute capacity; management recently introduced Meta Business Agent Platform for enterprises to build and manage agents at scale on WhatsApp; Brazilian car rental company Movida used a business agent on WhatsApp and saw a 44% increase in daily bookings compared to a year ago, with 85% of conversations in WhatsApp resolved entirely by the AI agent without human assistance

We see a large enterprise opportunity to sell to businesses, including APIs, Business Agents, potentially selling Compute directly and other services that we’re building for large customers…

…Muse Spark 1.1 is a strong agentic encoding model that is very efficient and excels at computer use, tool use, and multimodal understanding. It’s available through our new public API. We are ramping up distribution through partner channels and more coding agents over the coming weeks. We are also building out features to make it easier for enterprises to adopt Muse Spark…

…We made Meta business agents available globally this quarter on WhatsApp and Messenger. There are already more than 1 million businesses using them to talk to their customers or complete sales every week. We’re rolling business agents out on Instagram now, too. One interesting thing about having an agent talk to your customers every day is that it learns over time and can bring all of those insights back to you. We’re building more agentic capabilities to summarize all these conversations, digest what happened overnight, and surface what customers are asking for. Soon it’ll go further, including suggesting ways to grow your business, giving you competitive intelligence and real-time insights into what’s working and what’s not. Over time, we’d like to build this into a business-in-a-box service that can help you start and run a whole business using Meta’s platforms. In terms of how we will monetize these, we have a mix of subscriptions, volume-based pricing. I expect that we’re going to continue to evolve more of these products to be like our ad systems, where businesses only pay us when we achieve results for them. Over time, that will let us run an efficient auction over our compute, similar to how we do that for advertisers today…

…Earlier this month, we also introduced the Meta Business Agent Platform, which gives enterprises the infrastructure to build, customize, and deploy their business agent at scale on WhatsApp. The platform provides larger businesses with enterprise-grade controls, guardrails, and measurement built in so they can define rules and offer personalized experiences, starting within the messaging apps that their customers already use. Movida, one of Brazil’s largest rental car companies with nearly 400 locations, deployed a business agent on WhatsApp to handle the entire booking flow, from vehicle selection and pricing to payment, in a single conversation. Returning customers could complete a reservation in as few as three messages. In a one-month period, Movida reported a 44% increase in daily bookings through WhatsApp when compared to the same period in the prior year, that 85% of conversations in the channel were resolved entirely by the AI agent without human assistance.

Meta’s management thinks the trajectory of Meta Superintelligence Labs (MSL) is strong; MSL recently shipped the Muse Spark 1.1 and Muse Image models; Meta AI has seen a 60% increase in daily users since Muse Spark was integrated with it; Muse Spark 1.1 is an agentic coding model that is available through Meta’s new public API (application programming interface); management still believes in open sourced models, and expects MSL to release open source models soon; management thinks that closed models can be more jagged in intelligence than open sourced models; some experiences on Meta’s apps are not possible to be built by others because Meta has always taken a full-stack approach to its technology stack and this is why management wants control over the models Meta is using, and hence why the company is building its own models; management thinks current open source models are not as good as frontier models, so they do not want to use them; management thinks companies who invest in building models will be rewarded well over time

It’s been a little more than a year since we launched Meta Superintelligence Labs, our trajectory is strong. In the last month, we shipped Muse Spark 1.1 and Muse Image. Since we rebuilt Meta AI and integrated Muse Spark, we have seen a 60% increase in the number of people interacting with the assistant each day, that continues to grow quickly week-over-week…

…On open source, I think we have always felt like open source was an important part of the ecosystem, it’s good for the world…

…In ramping up Meta Superintelligence Labs, in some ways, actually counterintuitively, it takes some more work to do open source models because if you’re doing something as a closed system that you’re only building for your own use cases, it can be a little more jagged. Whereas if you release it as open and it’s going to be used for a lot of things, you want to make it more well-rounded…

…We expect that we will get back to releasing some open source models at some point soon…

…The question is, do we think that because there are some open-weight models that we can just rely on those? Right now, the open-source models are not as strong as the frontier models, so no is the basic answer…

…Some qualitative experiences are just not even possible for others to build because we go all the way down the stack. It just seems to me pretty clear that having kind of sovereignty over building your own models is going to be an important part of that stack going forward, which is why it is important for Meta, but is also why other people care about open source and why open source matters overall…

…I get that this is a big investment and it’s a big bet. We see the technology working. We’re happy with the trajectory of the lab. I’m excited about the products that are coming we believe that this is going to be a big thing. I get that this is sort of a big bet across the industry. My personal bet is that the people who invest in this are going to be rewarded and feel very good over time.

Meta’s management continues to see glasses as the ideal form factor for hardware that people use to interact with AI because they can be worn throughout the day and can provide insights while people remain present in the moment; Meta recently released its own line of Meta glasses in collaboration with EssilorLuxottica, and these glasses come with Muse Spark out of the box; early sales for the Meta glasses have exceeded management’s expectations

As we get closer to personal superintelligence, we are also going to need hardware that allows you to seamlessly interact with it. Glasses are the ideal form factor since they can be with you throughout the day and they can assist you without pulling you away from the moment. Our glasses remain one of the fastest-growing consumer electronics of all time. We continue adding to the lineup. We just released our own line of Meta glasses in collaboration with EssilorLuxottica, including a style that we designed with Kylie Jenner. They’re the first glasses to ship with Muse Spark out of the box so that they can understand what you’re seeing and give even more helpful answers. Early sales have been strong, exceeding our expectations.

Meta’s management sees the company as the only one in the world building AI with the main goal of putting it in people’s hands and distributing it widely

We’re the only major company building AI with the primary goal of putting superintelligence directly into people’s hands. Rather than centralizing superintelligence, we are focused on distributing it widely and giving everyone the ability to direct it towards what matters to them. That’s the way that society has always made progress. I think that these are the right values for building a positive AI future, and if we help build this, then I think that we will continue to build a very strong business as well.

Improvements to Meta’s Feed and Reels recommendations drove a double-digit year-on-year increase in global time spent on Instagram; improvements to rankings drove a 9% year-on-year increase in video time spent on Facebook, and over 10% in US and Canada

On Instagram, global time spent this quarter grew double digits year-over-year this quarter, largely driven by improvements to our Feed and Reels recommendations. On Facebook, video time spent increased 9% globally year-over-year and over 10% within the U.S. and Canada, where it was driven by ranking improvements.

LLMs improve Meta’s ranking and recommendation systems by (1) making existing systems smarter by understanding content and generating better training data, (2) helping with engineering development by evaluating content quality, detecting trends, and testing ranking changes, and (3) increasing the level of personalisation in recommendations; on engineering development, Meta recently had every public Reels and Feed post on Instagram being automatically analysed by an LLM, and management is working to include more surfaces on Facebook; on engineering development, management recently began using the Muse family of models to understand content and early results are positive; on increasing the level of personalisation, management recently shipped the largest single-release ranking improvement to date on Reels, which drove a 15 basis point increase in sessions on Instagram, and this ranking improvement is now being brought to Feed; on increasing the level of personalisation, Meta’s largest ranking models can now identify high-quality new Reels at creation, and over half of all recommended content on Instagram Feed is now less than one day old, more than double from 2025 Q2; on increasing the level of personalisation, users can now use natural language prompts to tune their recommendations on Instagram and Facebook, and early users have an 80% retention rate; Meta is building its next-generation recommendation systems, and the effort includes building foundation models for simultaneously recommending organic content and advertising, and developing LLM-native recommender systems; on building next-generation recommendation systems, in 2026 H1, Meta had continuously pre-trained a large-scale model with recommendations data, and observed healthy scaling laws; managment sees room to continue improving Meta’s recommendation systems into 2027

We are finding that LLMs are increasingly capable of delivering ranking and recommendations gains.

First, they make our existing systems smarter by understanding what the content is actually about and generating better training data.

Second, LLM-powered agents are also helping with engineering development by evaluating content quality, detecting trends, and testing ranking changes. Earlier this year, we reached a milestone of every public Reels and Feed post on Instagram being automatically processed through an LLM and analyzed across dimensions from topics to tone, and we’re working towards including more surfaces on Facebook as well. These signals can then be passed to downstream applications across ranking, recommendations, and content policy enforcement, which is a key building block toward greater personalization. This quarter, we also began using our Muse family of models to conduct content understanding across signals like video topic classification and summarization, and we’ve seen positive early results.

Finally, our recommendations are also becoming more personalized, surfacing more fresh content while giving people more direct control over what they see. On Reels, we shipped our largest single-release ranking improvement to date, combining faster inference with a new architecture that draws on deeper user history to improve predictions. This drove a 15 basis point increase in sessions on Instagram, with particular strength in reshares and time spent, which are both strong indicators of better content to user matching. We are now bringing this to Feed, where early results look comparable. We are also getting new content to people more quickly. Investments we’ve made in more real-time infrastructure and modeling improvements on new videos are allowing our largest ranking models to now identify high-quality new Reels at creation. On Instagram Feed, over half of all recommended content is now less than one day old, more than double from a year ago. We’re also giving people more direct control of the content they see. Today, Instagram users can visit the Your Algo page, which lets users write natural language prompts to tune their recommendations. Similarly, on Facebook, we launched Shape Your Feed. Early results show over 80% retention among users who engage with it.

Looking forward, we’re executing on our longer-term efforts to develop the next generation of our recommendation systems. This includes building foundation models that are designed to power organic content and ads recommendations simultaneously, as well as developing LLM native recommender systems. We hit our first research milestone this half by continuously pre-training a large-scale model with recommendations data, and observing healthy scaling laws in the process. We’re encouraged by this milestone and expect continued progress in the second half of the year…

…We certainly see further headroom to continue improving recommendations over the rest of the year and into 2027. We expect that will help us drive additional gains on both engagement on Facebook and Instagram.

Meta continues to enhance its systems to show advertising to users at the optimal time and location; management recently introduced Meta Generative Recommender, which uses LLMs to reason about ad content and user preferences simultaneously; early pilots of the Meta Generative Recommender architecture drove a 1% increase in app event conversions on Instagram; in 2026 Q2, management improved Meta’s user understanding models, which when combined with GEM (Generative Ads Model), drove an 8.3% increase in ad clicks and a 15.7% uplift in conversions on Facebook; Advantage+, Meta’s suite of AI-powered advertising automation tools, is now at a $75 billion revenue run rate (was $60 billion in 2025 Q3); advertisers who use multiple tools within Advantage+ see compounding gains; Indian apparel company Underneat adopted Advantage+ and saw a 13% incremental lift in purchases and a 16% increase in add-to-cart conversions 

The first part of this work is optimizing the level of ads within organic engagement. Here, we continue to enhance our systems to show ads at the optimal time and location…

…Within our ad systems, we’re delivering performance gains as we deploy more complex and predictive models. This quarter, we introduced Meta Generative Recommender, a paradigm shift in how our ad system works. Rather than scoring every possible ad individually, we are now using LLMs to reason about ad content and user preferences together and predict the best ad for each person. This makes our ad matching more intelligent and more precise, which compounds performance gains for advertisers. We deployed the first generative model into our ads retrieval system and saw notable improvements in ads performance. Early pilots using LLMs to better understand user preferences drove a 1% increase in app event conversions on Instagram. In Q2, we also advanced our user understanding models to analyze ads and organic activity, simultaneously improve both user experience and advertiser performance. Combined with our GEM model for ads ranking and sequence learning, these advancements generated an 8.3% increase in ad clicks and a 15.7% uplift in conversions on Facebook…

…Our AI-powered Advantage+ end-to-end solutions continue to grow, reaching over $75 billion in annual revenue run rate this quarter. We’re working to deepen adoption as advertisers who leverage multiple tools see compounding performance gains…

…Underneat, an online apparel brand in India, had been setting up each campaign manually across Facebook and Instagram. After adopting Advantage+ sales campaigns layered with Advantage+ audience placements and budget optimization, they saw a 13% incremental lift in purchases and a 16% increase in add-to-cart conversions.

Meta’s management continues to invest aggressively in AI infrastructure to meet growing AI usage; management recently announced a new strategic venture with BlackRock for a 1 GW data center; management expects a significant part of Meta’s compute to go towards training models, growing the core business, and building AI agents and new products; Meta is getting a lot of offers for its compute at a significant premium over what it paid; management’s approach to building compute capacity has 3 key elements, (1) the industry has historically under-built for AI adoption, (2) management has high confidence in Meta’s ability to utilize the capacity effectively, and (3) industry capacity will remain tight for some time; management’s current plans are to maximise capacity for 2026 and 2027, and to have flexibility to grow compute in 2028 and beyond if needed; the flexibility comes from the long-lived nature of the compute assets and Meta’s efforts in developing custom AI chips; management thinks that Meta’s models, consumer experiences, and enterprise offerings will be the best and highest ROI use of its compute infrastructure; a substantial amount of Meta’s compute goes towards training models; management believes there’s a much higher margin on selling intelligence rather than on selling compute directly; management believes that near-term compute capacity is more valuable than long-term capacity; Meta is compute-constrained today, so much so that there are many ROI-positive areas that management would put compute toward if there was any capacity

As AI usage in our products and businesses continues to ramp, we continue to invest aggressively in infrastructure to meet the demand. Yesterday, as part of our Meta Compute effort, we announced a new strategic venture with BlackRock to develop a new one gigawatt data center in El Paso, Texas. Overall, we expect that a significant portion of our compute is going to go towards training our models, growing our core business, and delivering personal agents and new products. We also expect to grow a large business serving large customers as well…

…We’re getting a lot of offers for compute at a significant premium over what we paid for it, and we have more coding and productivity tools on our roadmap as well…

…Our approach to building capacity is strongly influenced by several key elements. First, the broad environment for building infrastructure is dynamic and uncertain in both near-term and longer-term time horizons. The industry has under-built historically for the wave of AI adoption, making existing capacity, including our own, extremely valuable. Longer-term, the supply chains need to be built out to support the capacity that we anticipate we and others will need for AI-powered experiences. Second, we have high confidence in our ability to utilize capacity to scale and build on top of our existing experiences, as well as continue to invest in foundational models that will create substantial new opportunities. Consequently, our current plans are geared towards maximizing 2026 and 2027 capacity. When we have had incremental capacity in the past, it has proven extremely valuable in scaling experiences like Reels, and we are confident that this will be true in this timeframe as well. Longer-term, it’s harder to predict the exact usage scaling curves…

…We believe that being on the frontier will unlock new markets and opportunities for which we may need additional compute. Therefore, our longer-term capacity strategy aims to give us the flexibility to continue growing compute in 2028 and beyond by laying down data center and network foundations to accommodate future server decisions. The long-lived nature of these assets inherently provides the flexibility that will make it possible to adjust our investment to the pace of AI adoption. In addition, we have been making strategic investments in areas like our internal custom silicon effort, which will provide long-term strategic flexibility and supply chain leverage…

…Finally, we believe that overall industry capacity is going to remain tight for the foreseeable future. As we’ve said earlier, we strongly believe that the models, consumer experiences, and enterprise offerings that we are building will be the best and highest ROI use of our infrastructure. Those enterprise offerings have the potential to take multiple forms, as Mark mentioned. Agentic tools, our API, or monetizing compute directly, given outsized market demand…

…A substantial amount of the compute goes towards training models to be a leading lab, and I think that’s an important investment…

…We believe that there will continue to be a significantly higher margin on selling intelligence rather than selling compute directly…

…Generally, we believe near-term capacity is more valuable than long-term capacity…

…We are today and expect to be in the sort of foreseeable future demand-constrained. That really includes our core business too, where we still have numerous ROI-positive places that we would put compute toward if we had it.

Meta’s management believes that the company’s distribution advantage will give the company the opportunity to serve AI products to many users, even if the company’s models are not on the frontier

We believe that our distribution advantages will give us the opportunity to serve AI products that are valuable for everyone, both our 3.6 billion users and millions of businesses. This should be true regardless of whether our models are on the frontier.

Microsoft (NASDAQ: MSFT)

Microsoft’s management added 31 new data centers across 5 continents in 2026 Q2 (FY2026 Q4), or 1 gigawatt of compute capacity; Microsoft is bringing capacity online faster than ever; Microsoft has reduced dock-to-live times for new GPUs by nearly 50% over the last 12 months; Microsoft is on track to double its overall compute footprint in 2 years; 

We added 31 new data centers across five continents this quarter, bringing the total to 88 this year as we expand our footprint in response to accelerating demand. We’re also bringing capacity online faster than ever. Over the last fiscal year, we’ve reduced dock-to-live times for new GPUs in our largest regions by nearly 50%. All up, we added another gigawatt of capacity this quarter and remain on track to roughly double our overall capacity in just two years.

Microsoft’s management is optimising silicon across systems and software to get more from the company’s compute infrastructure; Microsoft has increased the throughput for Copilot workloads by 4x since the start of FY2026; Microsoft’s AI infrastructure utilises chips from NVIDIA, AMD, and itself (Maia); Microsoft’s Maia 200 chip continues to scale, has 30% better tokens per dollar compared to other leading AI chips, and now supports both OpenAI and MAI (Microsoft AI) models; Microsoft will be among the first cloud providers to provide the latest generation GPUs from AMD (Helios) and NVIDIA (Vera Rubin); management sees CPUs as being just as important as GPUs when running agentic workloads; Microsoft’s Cobalt CPUs are powering 1st-party and 3rd-party workloads; management expects to have Microsoft’s own Cobalt 200 racks in over 25 data centers around the world by the end of July

We’re also getting more from the infrastructure we already have by optimizing across silicon systems and software. For example, we increased the throughput for Copilot workloads 4x since the start of the year…

…We also continue to modernize our fleet with our own silicon innovation alongside the latest from NVIDIA and AMD. Maia 200 continues to scale. It delivers 30% better performance per dollar than the latest generation hardware in our fleet and is now supporting both OpenAI and MAI models. And we will be among the first cloud providers to deploy next generation rack-scale AI infrastructure based on AMD Helios and NVIDIA Vera Rubin.

When it comes to running agents, CPUs are just as important as GPUs. Our Cobalt VMs are powering both our own first-party workloads as well as workloads for customers, including Adobe, Arm, Elastic, OpenAI, Sprinklr, and TomTom. By the end of this month, we expect to have our Cobalt 200 racks in over 25 data centers around the world as we rapidly expand capacity.

Microsoft’s management thinks customers want the right model for the right task; management thinks Microsoft offers the broadest selection of models, more than 11,000, among the cloud hyperscalers; the number of customers building with models from different providers has increased by 5x since the start of FY2026; Levi Strauss is using Foundry to bring together models from OpenAI and Anthropic, and 1,000 domain-specific agents; management is building Foundry to be the complete app and agent stack, which gives agents access to the IQ layer, tools, durable state and memory, sandboxes, rubrics, and evals; there are more than 100,00 Foundry customers and revenue doubled year-on-year in 2026 Q2 (FY2026 Q4); Telefonica is using foundry for its first wave of agents tackling mission-critical network operations; the number of Foundry customers at an annualised run rate of 1 trillion tokens was up 4x year-on-year in 2026 Q2 (FY2026 Q4)

Every customer wants the right model for each task based on quality, latency, cost, and compliance. We offer the broadest model catalog in the cloud with over 11,000 models, including the latest from OpenAI, Anthropic, Mistral, xAI, as well as our own MAI family. Since the start of the year, we have seen 5x increase in the number of customers building with models from multiple providers. Levi Strauss & Co., for example, is using models from OpenAI and Anthropic on Foundry as it brings more than 1,000 domain-specific agents into a unified enterprise AI platform…

…Beyond model choice, data, and context, we are building Foundry as the complete app and agent stack. It gives agents access to the IQ layer, the tools they use, along with durable state and memory, secure sandboxes, rubrics and evals, and even their own self-improvement loops. We now have 100,000 Foundry customers and revenue more than doubled year- over- year. Telefónica, for example, adopted Foundry as the foundation of its corporate agentic platform with its first wave of agents tackling mission-critical network operations. All up, the number of Foundry customers at 1 trillion tokens annualized run rate increased 4x year- over- year.

Microsoft recently announced more than a dozen new models across modalities, and all offer cost-efficient inference; Microsoft’s models are co-designed with its own AI chips; MAI (Microsoft AI) models have 40% better performance per watt when they are run on Maia 200 chips; management is building a new model system where the harness, context, memory, and action space are separate from the models, meaning every model is substitutable; Microsoft is using the new model system in its products; developers using MAI-Code-1-Flash on GitHub Copilot are getting higher code acceptance rates and 10% lower median token usage while still getting frontier capabilities; MAI-Code-1-Flash is delivering comparable quality to GPT-5.6 in Excel for common tasks with much lower costs; MAI-Cyber-1-Flash achieves better security performance than much larger models but at half the cost; MAI-Voice-2-Flash has led to an 89% reduction of GPU costs in Dynamics 365; MAI-Image-2.5 has led to an 84% reduction in GPU costs in PowerPoint; any company can use Microsoft’s new model system in Foundry; management thinks that it will be increasingly clear that organisations want AI system providers who will help them with outcomes and knowledge creation, and not simply extract knowledge, and this direction of travel has led to the architectural design of separating the harness from the model, so that the model layer is swappable; management thinks that the recent Hugging Face breach by an unreleased OpenAI model is a good example of why enterprises cannot depend on just one model; MAI-Cyber-1-Flash can achieve better security performance than much larger models but at half the cost because 90% of the tasks are done by MAI-Cyber-1-Flash whereas the remaining 10% is done by the frontier model; the work done by Microsoft on model diversification also helps with margin improvement

We are also accelerating our own model development. We announced more than a dozen new models across image, voice, transcription, coding, security, including our first reasoning model, MAI Thinking-1, all with cost-efficient inference at the core for the enterprise use cases. We are co-designing these models with our silicon. We are seeing 40% better performance per watt when running MAI models on Maia 200. More importantly, we are building a new model system where the harness, context, memory, and action space are separate from any one model family, thereby moving the frontier on the cost-to-outcome curve. It’s not just about cost. It also has the added benefit of business continuity and resilience because every model is substitutable.

This is the system we are using in our products with great results. For example, millions of developers have used MAI-Code-1-Flash on GitHub Copilot, achieving higher code acceptance rates and 10% lower median token usage while still having access to frontier capabilities from OpenAI and Anthropic. In Excel, MAI-Code-1-Flash is delivering comparable quality to GPT-5.6 for the most common tasks while operating at significantly lower costs. In security, MAI-Cyber-1-Flash achieves better performance than much larger Mythos model, but at half the cost when combined with our multi-agent security harness. More broadly, across our model implementations, we are seeing significant efficiency gains, including 89% reduction of GPU costs in Dynamics 365 with MAI-Voice-2-Flash and up to 84% reduced GPU costs in PowerPoint with MAI-Image-2.5. This system is available to any company to use as part of Foundry…

…Every firm is going to evaluate who are the providers who are helping them with their outcomes and their knowledge creation. I think that that is now fairly clear, and it’s going to become clearer by the day. This is not going to be about, come in and take all my knowledge and benefit yourself, whereas I am not getting anything out of it. Given that direction of travel, we are very clear about the architectural design of the platform, which is you’ve got to keep your harness separate from the model. The harness will ensure that your memory, your context, all of that is external. That means any given model at any given time is swappable…

…If you look even at the Hugging Face incident, the biggest thing that we should take away from that is you can’t depend on any one model. You will maybe need multiple models to even remediate some challenges that get caused by one model. That’s the way to think about it, which is you can’t be subject to the refusal of the one model…

…Essentially you can have Mythos level performance, with 50% less cost because of this MAI-Cyber-1-Flash. The reason is because 90% of the tasks are done by the MAI-Cyber-1-Flash model, and 10% of the tasks, you still go to the frontier. This is sort of that mixing of the right model for the right task in what is essentially a pipeline job, is a super important characteristic…

…The work, frankly, on model diversification also is a margin improvement opportunity. Being able to serve the best possible outcome with a more efficient, or both efficient in terms of token usage and efficient in terms of cost structure, are also margin levers.

Microsoft’s management sees the data estate shifting from primarily supporting human users to supporting agents; Microsoft’s customers are adopting the company’s AI-optimised databases; PostgreSQL revenue was up 55% in 2026 Q2 (FY2026 Q4); the number of PostgreSQL customers who also use Foundry was up 80% in 2026 Q2 (FY2026 Q4); customers are increasingly choosing PostgreSQL as the database for their AI workloads; Microsoft has Horizon DB, a fully-managed PostgreSQL service that has 3x the throughput of self-managed deployments;  Fabric customers grew 60% year-on-year in 2026 Q2 (FY2026 Q4) to 40,000 (was 35,000 in 2026 Q1); 17,000 customers now use both Fabric and Foundry, up 60% year-on-year (was 15,000 in 2026 Q1); management recently introduced the agent-fore Rayfin SDK (software development kit) for building apps in Fabric; tens of thousands of customers, including 90% of the Fortune 500, are grounding their agents; management recently introduced Web IQ, which allows agents access to the web; Web IQ is already being used by popular AI assistants, including ChatGPT; Agent 365 is a control plane for managing agents’ governance, identity, and security; tens of thousands of companies are already using Agent 365 to manage 40 million agents in 2026 Q2 (FY2026 Q4) (was “tens of millions of agents” in 2026 Q1)

The data estate is evolving from primarily supporting apps used by people to supporting agents. Customers are rapidly adopting our AI-optimized databases like Cosmos DB and PostgreSQL to give agents fast, secure access to real-time data and context they need for memory and retrieval.PostgreSQL revenue was up 55%, accelerating for the third consecutive quarter. Also, the number of PostgreSQL customers also using Foundry increased 80% as customers increasingly choose it as the database for AI workloads. We are going further with Horizon DB, our new fully managed PostgreSQL service on Azure, which delivers three times the throughput of self-managed deployments…

…We now have over 40,000 paid Fabric customers, up more than 60% year-over-year, and over 17,000 customers now use Foundry and Fabric, up 60% year-over-year as enterprises connect agents to real-time operational, analytical, and unstructured data in Fabric…

…Tens of thousands of customers, including nearly 90% of the Fortune 500, are already grounding their agents in enterprise context with Foundry, Fabric, and Work IQ. This quarter, we introduced Web IQ, which gives agents access to real-world intelligence from across the web. It is already being used by the most popular AI assistants, including ChatGPT…

…With Agent 365, we offer a control plane that extends companies’ existing governance, identity, security, and management frameworks to agents they build. Just two months in, Agent 365 now has nearly 40 million agents registered across tens of thousands of companies.

There are over 30 million Microsoft 365 Copilot seats in 2026 Q2 (FY2026 Q4) (was “over 20 million” in 2026 Q1), with net seat adds up more than 100% sequentially; management made Copilot generally available in June 2026 to help customers complete multi-step tasks; management recently introduced Autopilots, which are autonomous long-running agents powered by OpenClaw; in 2026 Q3 (FY2027 Q1), management will bring all of Microsoft’s Copilot experiences together in one super app that spans both consumer and commercial experiences; management has been improving Copilot and recent customer feedback has been good; Copilot’s user satisfaction scores have doubled over the last 3 quarters; Copilot’s latency was reduced by 25% in 2026 Q2 (FY2026 Q4); usage intensity of Copilot is at a record, with conversations per user up nearly 100% year-on-year in 2026 Q2 (FY2026 Q4), and average weekly engagement on par with Outlook and Teams; the time taken for Copilot to move from deployment to high usage has fallen from months to just days over the last 12 months; the number of Copilot customers with more than 50,000 seats was up 7x year-on-year in 2026 Q2 (FY2026 Q4); the number of Copilot customers deploying Copilot to the majority of their knowledge workers was up 75% sequentially in 2026 Q2 (FY2026 Q4); Copilot has saved NHS England employees 43 minutes per year, and NHS England is rolling out Copilot to 505,000 employees; KPMG is rolling out Copilot to 276,000 employees; HSBC has committed to 200,000 Copilot seats; many customers have bought 60,000 or more Copilot seats; management recently added usage-based billing to Copilot Cowork; the super app will have Chat, CoWork, Autopilot, and Code in it; management thinks the addition of usage-based billing to Copilot Cowork has increased Microsoft’s addressable market

We now have over 30 million paid Microsoft 365 Copilot seats with net seat adds more than doubling quarter-over-quarter…

…Last month, we made Copilot generally available, helping customers complete multi-step tasks grounded in their work data while meeting enterprise security and compliance requirements. This quarter, we also introduced Autopilots, autonomous long-running agents with full enterprise compliance, including always-on personal agent powered by OpenClaw.

This quarter, we are bringing these Copilot experiences together, including code in one super app spanning both consumer and commercial experiences. This is a major step forward and I look forward to sharing more soon. More broadly, we have steadily been improving the quality and performance of Copilot and have been delighted by the recent customer feedback. Over the last three quarters, user satisfaction scores have doubled and are now at an all-time high. This quarter alone, we cut latency by 25%. These quality improvements, together with continued product innovation, are driving record usage intensity. The number of conversations per user nearly doubled year-over-year. Average weekly engagement is on par with Outlook and Teams. The time from deployment to what we think of as high usage, meaning monthly active usage, about 80% across a customer’s user base, has fallen from months to just days over the past year. The number of customers with more than 50,000 seats increased over 7x year-over-year, and the number of enterprise customers deploying Copilot to the majority of their information workers grew nearly 75% quarter-over-quarter, a signal of how central Copilot has become to their operations.

NHS England, for example, is rolling out Copilot to 505,000 clinicians and staff, the largest healthcare deployment of its kind after a trial showed it saved employees an average of 43 minutes per day. KPMG is expanding its deployment across its global workforce of more than 276,000 professional, HSBC committed to 200,000 seats to accelerate its workforce transformation. AstraZeneca, Boeing, Infosys, Coke Inc., Procter & Gamble, Stellantis, Tata Consultancy Services, University of Pittsburgh Medical Center, Wells Fargo, and Wipro each purchased 60,000 or more…

…Earlier this month, we added usage-based billing to Copilot Cowork with thousands of customers already paying for and actively using it…

…We now have Chat, CoWork, Autopilot, Code all coming to essentially what is going to become this flagship super app that various roles can use it…

…If I think about historically Office compared to what Microsoft Copilot is much more narrower. This is the first time where you really have an enterprise-wide tool, which has a both per seat and usage-based pricing. The TAM is much more expansive.

Microsoft’s management has added consumption plans to its business model; customer service within Microsoft Dynamics 365’s MCPs (model context protocols) have seen the highest usage-based credit consumption, up 4x sequentially in 2026 Q2 (FY2026 Q4); management recently introduced usage-based billing to GitHub Copilot; after the introduction of usage-based billing, GitHub Copilot had significant consumption revenue and continued to have business and enterprise seat growth

In addition to this, we are also evolving our business model beyond per seat to per seat plus consumption, further expanding our TAM and delivering more customer value. Earlier this month, we added usage-based billing to Copilot Cowork with thousands of customers already paying for and actively using it…

…In Biz Apps, we have been reinventing Microsoft Dynamics 365 for an agent-first world. We are exposing over 650,000 MCP actions across sales, finance, supply chain, HR, and customer service so that agents can now access business context and take action using the same data models, rules, permissions, security guardrails, and audit trails as any application user. We are also moving from seats to seats plus consumption model. Customer service is at the forefront of this transformation with usage-based credit consumption in this category up 4x quarter-over-quarter with customers like Northern Trust using our tools to drive proactive intelligence…

…GitHub Copilot now has 50 million users. This quarter, we introduced usage-based billing and have continued to see business and enterprise seat growth and also significant consumption revenue after the new model went into effect. Copilot revenue accelerated over 60% quarter-over-quarter.

Microsoft’s management recently launched Microsoft Frontier Company for customers to build AI systems that learn and improve with usage, through customers’ workflows, domain knowledge, and judgment; Microsoft will embed 6,000 industry and engineering experts with customers as part of Microsoft Frontier Company; Microsoft has completed over 330 of such projects with 164 customers; Microsoft’s teams worked with Novo Nordisk to build an agent that analyses clinical data under strict compliance requirements; Microsoft worked with LSEG to embed AI into LSEG Workspace

There is a tremendous opportunity to turn customers’ workflows, domain knowledge, and accumulated judgment into AI systems that learn and improve with every usage. To help customers capture that opportunity this month, we launched Microsoft Frontier Company, the largest outcome-driven engineering organization in the industry. We will embed 6,000 industry and engineering experts with customers to co-design, co-innovate, and continuously improve AI systems at scale. We’ve been testing this model over the past year, completing over 330 projects across 164 customers, including many of the world’s leading companies across industries. For example, our FD teams worked with Novo Nordisk to build an agent that helps analyze clinical data while meeting its strict compliance requirements. We partnered with LSEG to embed AI into LSEG Workspace, helping finance professionals ask complex questions and quickly find answers across structured and unstructured financial content.

Microsoft’s management sees a big opportunity for Windows to become a place for unmetered intelligence by combining on-device compute with enterprise-grade security

We see significant opportunity for Windows to become the offload for unmetered intelligence, combining powerful on-device compute with enterprise-grade security.

1/3 of Microsoft’s cloud and AI-related capex in 2026 Q2 (FY2026 Q4) are for long-lived assets that will support long-term monetisation, while the other 2/3 are for CPUs and GPUs; management is able to easily slowdown Microsoft’s capex, both for the short-lived and long-lived assets, depending on the demand environment; in thinking about Microsoft’s capex, management is studying history, such as the railroad boom-and-bust described in the book 1873

Capital expenditures were $41 billion, including the impact from higher component pricing as noted in our guide. Roughly two-thirds of our CapEx was for short-lived assets, primarily CPUs and GPUs, as customers increasingly build solutions that leverage both AI and non-AI infrastructure. The remaining spend was for long-lived assets. This quarter, total finance leases were $5.6 billion and were primarily for large data center sites, and cash paid for PP&E was $35.8 billion…

…You’ve seen our CapEx really pivot toward what I would call and do call short-lived assets, which really, right, that’s CPUs and GPUs that have relatively shorter lead times. If the demand environment changes, you just slow down what is, in fact, the largest component, right, and the driver of COGS. The investment into land and data center builds is actually quite flexible, right? It’s a smaller percentage of the overall cost structure, and timing can be changed on much of that, especially on the builds. You can stagger the timing of the build-out of, as I was saying, some of the GPUs and CPUs that you plan to put in. When you think about being able to manage through that, hyperscalers have been doing that for quite a long time in terms of having the flexibility and the understanding of manage those changes in demand…

…All of us are reading this, “1873” is the book to be read. So in my mind, I think you’ve got to get the product shape right. That’s sort of a lot of what we are focused on. You have to get the portfolio right. Amy talked about what we are doing, whether it’s in Copilot or the super app, bringing all the form factors or all the way to Azure and the agent first sort of primitives in Azure. You kind of have to really get that portfolio to all come together. The mix of customers is super important. You have to recognize the breadth, the geo mix, the segment mix, the workload mix, and you got to really think about all of those when you’re even building capacity. You’ve got to run an efficient railroad.

Azure grew revenue by 43% in 2026 Q2 (FY2025 Q4) (was 40% in 2026 Q1); Azure’s revenue growth was better than expected because of efficiency gains in the compute fleet and the earlier delivery of new capacity; Azure continues to be constrained by capacity; the new capacity was quickly monetised; Azure’s revenue in 2026 Q2 also benefitted from stronger than expected GitHub Copilot consumption following the June business model change to usage-based pricing; Azure’s compute infrastructure is fungible when it comes to running any model family; Azure has a very diverse book of business; Azure has been able to adjust pricing to counter a rise in hardware costs 

In Azure and other cloud services, revenue grew 43% against a prior year that included accelerating growth. Customer demand continues to exceed available capacity. Revenue growth was ahead of expectations, driven by efficiency gains across our CPU and GPU fleet, as well as process improvements to enable earlier delivery of new capacity. That additional in-quarter capacity for Azure was quickly monetized. Results also benefited from stronger than expected GitHub Copilot consumption following the June business model change to align pricing with usage and value…

…Given that we continue to see growing demand, no matter what model is chosen or what model family or whether it’s run a model of your own, the Azure platform’s quite efficient at delivering that. Think about that infrastructure as being pretty fungible…

…The other thing is, that’s important, Mark, is you just have an incredibly diverse book of business, by geo, by segment, by industry…

…[Question] How do you manage through the hardware price increases that we’re seeing, the component prices?

[Answer] We’re adding this capacity, to your point, but a lot of this obviously is also being sold in newer contracts, and we’re able to have the pricing reflect it, but keep value.

From FY2027 onwards, management will extend the useful life of Microsoft’s data centers and offices from 15 years to 25 years; the extension of the useful life will only have minimal benefit to FY2027’s operating income; the extension of the useful life will have greater impact to capital expenditures, as the extension will shift Microsoft’s future data center leases from finance leases to operating leases (finance leases are included in capital expenditures, while operating leases are not); management now expects 2026’s total capital expenditure to be $175 billion (was previously $190 billion) because of the shift from finance leases to operating leases

Effective at the start of FY 2027, we are extending the estimated useful life of our data centers and office buildings from 15- 25 years, reflecting our operating history and expected use of these assets. The impact of this update is reflected in today’s guidance. This change affects only the timing of future depreciation and is expected to have a minimal benefit to FY 2027 operating income. The greater impact is on capital expenditures, as more of our future data center leases will shift from finance leases to operating leases as a result of this update. Finance leases are included in capital expenditures, while operating leases are not.

Outside of this useful life impact, our calendar year 2026 CapEx investment expectations remain unchanged. However, the shift from finance to operating leases adjusts our expectation to approximately $175 billion…

…At the company level, with strong commercial momentum, we continue to expect another fiscal year of double-digit revenue and operating income growth. Operating expenses should grow in the mid to high single digits, reflecting continued investment in R&D compute capacity, talent, and data. We expect FY 2027 capital expenditures will grow year-over-year, given demand signals across our portfolio. Even as we invest to meet growing demand, full fiscal year operating margins should be down less than a point. In addition, we expect to remain free cash flow positive in FY 2027.

Netflix (NASDAQ: NFLX)

Netflix’s management is using LLMs (large language models) to improve title discovery and understanding of member preferences; management is using AI to improve Netflix’s search function; in 2026 H1, GenAI (generative AI) workflows were used in 300 of Netflix’s titles, with the workflows mostly in post-production; the use of GenAI helps Netflix to deliver higher quality output faster and at lower cost, and also produce sequences that otherwise would have to be skipped; it’s still early days for Netflix’s acquisition of InterPositive (a startup focused on providing AI tools for filmmakers) but Netflix is already seeing the impact of genAI on its productions; Netflix has other genAI tools outside of InterPositive; management believes that the presence of AI tools does not change the fact that it takes great artists to create great content; Netflix is using AI tools for set references, pre-vis, VFX, sequence prep, and shot planning; management is seeing AI use cases in content production scaling faster and faster; a recent Netflix documentary, American Experiment, features 17 minutes of AI-enhanced footage that were produced twice as fast and at half the cost compared to non-AI options

We are leveraging LLMs to improve title discovery and to better understand member preferences. We’re also enhancing search for our members with new voice search functionality and AI-powered natural language search…

…In 2026, GenAI workflows have been used in roughly 300 of our titles, with the largest concentration of work in post-production. We are increasingly leveraging these tools to deliver higher quality output more quickly and at a lower cost than traditional methods. In some cases, productions would have had to leave out key shots and sequences in the absence of GenAI technology. For example, Glory (India), Brasil 70: A Saga do Tri (Brazil), and The American Experiment (US) utilized GenAI tools to create highly complex sequences (e.g., enhanced crowds, historical battle sequences, and worldbuilding establishing shots)…

…It’s early days for InterPositive, but we’re broadly seeing that gen AI is starting to have an impact across hundreds of our productions. So important to note that we have other gen AI tools in addition to InterPositive…

…On the content side, we believe it takes great artists to make something great, and AI is not changing that. AI will give creators better tools to bring their visions to life. Movies are being made by people who make movies. AI provides them with better tools to make them even better…

…So today, our talent leverages tools for things like set references and pre-vis and VFX and sequence prep and shot planning, which all makes the production itself so much more smooth and efficient and fast. And that’s just the beginning. We’re seeing it across the entire production life cycle and AI — those use cases are scaling faster and faster. So our documentary series we just released called American Experiment. That series features 17 minutes of AI-enhanced footage. It enabled us to expand the scope of the series in ways that just wouldn’t have been feasible before. Those 17 minutes, they were produced twice as fast and at half the cost of previous options.

Netflix’s management has expanded Netflix’s AI-powered advertising tools across its full advertising lifecycle

In Q2, we expanded our AI-powered tools across the full advertising lifecycle, from planning and creative production to campaign management, optimization, and reporting. 

Taiwan Semiconductor Manufacturing Company (NYSE: TSM)

TSMC’s management is seeing very strong demand for TSMC’s leading edge nodes; TSMC’s capital expenditure is always in anticipation of growth in future years; management has raised capex guidance for 2026 to US$60 billion to US$64 billion partly because of agentic AI (previous guidance is for US$52 billion to US$56 billion; capex growth at the high end of the new guidance would be 56% from 2025’s capex of US$41 billion); management does not foresee any bottlenecks to TSMC’s capacity expansion plans; most of TSMC’s capex for 2026 will be for advanced process technologies; management now expects TSMC to grow revenue by above 40% in USD terms in 2026 (previous guidance was for growth to above 30%); TSMC’s capex in the last 3 years was ~US$100 billion, and the next 3 years is now expected to be much, much higher (previous guidance was for it to be “much higher”), although management does not have a specific 3-year capex outlook to share for 2026-2028; management now thinks the AI accelerators business will have an even stronger CAGR for 2024-2029 than the high-end of the mid-to-high-50% CAGR communicated in the 2026 Q1 earnings call; it’s very likely for TSMC’s capex guidance for 2026 to continue increasing; management thinks the next few years will look really good for TSMC; management now thinks the CAGRs for 2nm, 3nm, and 5nm nodes will be even higher than the very strong CAGRs mentioned during a recent symposium; TSMC’s capex guidance for 2026 was raised partly because customers are willing to cooperate with TSMC (hinting at pre-payments?), and because of price inflation of tools

Demand for our leading-edge technologies is very strong…

…At TSMC, a higher level of capital expenditures is always correlated to higher growth opportunities in the following years…

…Given the continued strong structural demand from our customers, including the newly emerging Agentic AI market, we have decided to raise our full year 2026 capital budget to be between USD 60 billion and USD 64 billion as we continue to invest heavily to support our customers’ growth. We always collaborate closely with the tool suppliers well in advance to prepare the capacity, whether it is a strong up cycle or down cycle, just like our customers collaborate with us well in advance to plan our capacity. Thus, we do not foresee any bottlenecks to our capacity expansion plans.

About 70% to 80% of the 2026 capital budget will be allocated for advanced process technologies. About 10% will be spent for specialty technologies and about 10% to 20% will be spent for advanced packaging, testing, mask-making and others…

…Supported by our robust technology differentiation and the broad customer base, we now expect our full year 2026 revenue growth to be slightly above 40% year-over-year in U.S. dollar terms…

…We do not have a number to share with you. But as you know, we invest CapEx this year for the future business opportunity. And as long as there are business opportunities, we will not hesitate to invest. As you can hear from our prepared remarks that we — our conviction in the megatrend, AI megatrend multiyear is very strong, and we are stepping up the CapEx, including increasing this year’s CapEx. Last time, we said our CapEx in the next 3 years will be significantly higher than the CapEx in the past 3 years. Now is the — the CapEx in the next 3 years will be even more significantly higher than the past 3 years…

…If you read our message that we continue to invest more. We increased the CapEx with a good reason. So if you’re asking about the AI’s CAGR, let me give you not a number, but it’s stronger and stronger and stronger. So we don’t give you the number today because it continue to increase. So we don’t know how to answer this question, but stronger than what we said before…

…This year, we say we increased the CapEx from $52 billion to $56 billion, now $60 billion to $64 billion. And you bet, that will continue to increase…

…Because of the revenue corresponding to our investment, right, because we know we forecast our demand, and then we make an assessment, and then we do the CapEx. Next few years is going to be a very good business for TSMC. That’s all I can say…

…[Question] I noticed that during your symposium that you actually mentioned about 2-nanometer family capacity growth will be growing at around 70% CAGR from ’26 to ’28 and N3 plus and N5 to grow by 25% CAGR from ’22 to ’27. So I was just wondering, are those numbers still right assumptions today?

[Answer] We showed the chart. Okay. Now it’s bigger. That’s what I say…

…[Question] From year-to-date, so TSMC raised the CapEx guidance by almost USD 10 billion. So can you give me some color where is the upside from? How you guys see the difference from 6 months ago?

[Answer] The most important reason is because of the demand continued to increase, and we feel the pressure from the customer to drive TSMC, not drive actually, to cooperate with TSMC for the capacity increase. That’s one of the major reasons. The second reason is inflation. Now we buy the tools with inflation price.

TSMC’s management sees very robust AI-related demand; management continues to see very strong signals and positive outlooks from TSMC’s customers’ customers, who are the cloud service providers; management’s conviction in the AI megatrend remains very high; management thinks demand for chips, driven by AI, will be incredibly strong to at least 2030, but is unsure if there will be a dip in-between; management thinks the collective trend of AI is so robust it is creating a new industry altogether

AI-related demand continues to be extremely robust. The AI megatrend continues to drive the need for more and more computation, which supports the robust demand for leading-edge silicon. Our customers and customers’ customers, who are mainly the cloud service providers, continue to provide us with a very strong signal and positive outlook. Thus, our conviction in the multi-year AI megatrend remains very high…

…I believe from this day on all the way to probably 2029, 2030, the demand is very strong. Whether in between there’s a dip or not, I’m not very sure. But the trend is so robust that I believe we are witnessing a kind of a new industry. I would like to say the new industry called AI industry, which is so common in our daily life because it’s going to affect our automotive, affect the humanoids, robot, and also impact to all the industry. So by the amount of money we put in, I mean, including all the CSPs, this alone is a very important new industry to the world. And so the demand will be there. And the fundamental thing is semiconductor chips, and most of them in TSMC.

TSMC’s management sees the emergence of agentic AI leading to a resurgence of the importance of CPUs (central processing units) in AI data centers; management thinks the resurgence of CPUs is positive for TSMC because nearly all the companies behind the major CPU architectures are customers of TSMC

The emergence of Agentic AI is leading to a resurgence in the role of CPUs in AI data centers, which drives more silicon demand in addition to AI accelerators. We believe this is positive for TSMC as no matter what CPU approach is taken, whether it’s x86, ARM-based, or RISC-V architecture, they are almost all TSMC’s customers. We are already collaborating closely with our CPU customers and working to support them with the most advanced technologies and necessary capacity, so they can capture the Agentic AI market opportunities.

TSMC’s management collaborates closely with its customers, and customers’ customers, when planning the company’s capacity; management is aware that while the CEOs of TSMC’s customers’ customers are telling their own truths, the combined picture is not the truth, and management is adjusting for that in thinking about true end-demand so that TSMC does not end up with over-capacity

To address the structural increase in overall long-term semiconductor market demand profile, TSMC collaborates closely with our customers and our customers’ customers to plan our capacity. Given the fundamental complexity of leading-edge technologies and the design-in and lead time involved, we also have a very good idea of their multiyear product road map and production plans. This is important because it takes more than 5 years to develop the technology and product, prepare the capacity, and ramp it up to high-volume production. Internally, TSMC employs a disciplined capacity planning system to assess the market demand from both a top-down and bottom-up approach. This is a continuous and ongoing process. Based on our assessment, we are stepping up our CapEx investment to increase our capacity, to support our customers’ future growth…

…Now remember that I believe every customer tell me the truth, everyone. You put all the truths together, it’s not the truth. So we have to make some of the judgment. You know what I mean, since you are laughing. Because all the customers are very aggressive, right? That’s the CEO’s job. CEO got to be aggressive. So they give me the number of their demand, and I believe they try their best to tell me the truth. So I put all together, all the truths together is not a truth. Mark down that word. So yes, we do a very careful judgment. May not be correct, may not be correct, but we did carefully and because this is a big money, right? 

TSMC’s management has announced an additional US$100 billion investment in Arizona for fabs for nodes of 2nm-and-below, in partnership with TSMC’s US customers and the US government; management believes the latest Arizona investment will strengthen the US’s semiconductor supply chain; the additional US$100 billion investment in Arizona will see TSMC build at least 4 additional fabs, with both front-end and back-end fabs; TSMC is also receiving government support in the USA, similar to Intel, but TSMC does not announce the support; there’s no firm timeline for the additional US$100 billion in investment, as it depends on the market situation and customers’ demand, but management wants to move fast

With a strong collaboration and support from our leading U.S. customers and the U.S. federal state and city governments, we would like to announce an additional USD 100 billion investment in Arizona. This is to build several more semiconductor logical wafer fab for 2-nanometer and below technologies as well as advanced packaging fabs to support the strong multiyear demand from our leading U.S. customers. We believe this investment will help to further foster the development of the U.S. semiconductor ecosystem, strengthen the supply chain, and support an increasing number of high-tech, high-paying jobs in the United States…

…We announced additional $100 billion investment in Arizona. How many fabs? Many. So actually, let me say that, say probably, additional 4 more fabs will be built.

[Question] And that’s combining front- and back-end?

[Answer] Yes…

…The other one in the U.S., they got a very strong U.S. government support. We also got the government support, by the way, although we don’t announce it…

…[Question] Do you have any schedule or time frame to share about this additional $100 billion?

[Answer] Most of the time, it depends on the market situation and our customers’ demand. So if you ask me to give you a firm schedule, no, we don’t have it today. But we do have a plan. And we speed it up. We try to speed it up as fast as possible.

TSMC’s management is building new fabs in Taiwan; management is continuing to convert 5nm tools for 3nm capacity in Taiwan; management is focusing on flexible capacity support among the N7, N5, and N3 nodes; the upcoming A14 node has 10-15 speed improvement at the same power compared to N2, or 25-30 power improvement at the same speed, and a nearly 20% chip density gain; the A14 node is on track and progressing well; management has introduced the A13 and A12 extensions, which are both superior to A14; A13 and A12 are scheduled for volume production 2029; management believes the A14 family will be an even larger and long-lasting node than N2; management wants to move as fast as possible in Taiwan and other countries

We are building 13 leading-edge and advanced packaging fab in Taiwan over the next several years, and we will continue to further invest in Taiwan…

…We continue to convert 5-nanometer tools to support 3-nanometer capacity in Taiwan…

…We are also focusing on capacity optimization across node, which including flexible capacity support among N7, N5 and N3 nodes…

…Our A14 technology representing the second generation of nanosheet transistors and deliver another full node stride from N2 with performance and power benefit to address the incessant need for high-performance and energy-efficient computing. Compared with N2, A14 will provide 10 to 15 speed improvement at the same power or 25 to 30 power improvement at the same speed and close to 20% chip density gain. A14 technology development is on track and progressing well…

…We also introduced A13 and A12 as extension of the A14 family. A13 represents a further advancement of A14, achieving an over 6% die area saving through an innovative 97% optical shrink. Through continuous design technology co-optimization, A13 also drive further performance and power efficiency improvement. A13 design rule are backward compatible with A14 to ensure smooth IP migration. We also introduced A12, which will bring our innovative superpower rail technology to the A14 platform for superior performance, power, and area benefit. Both A13 and A12 are scheduled for volume production in 2029. We believe A14 and its derivative technologies will propel our A14 family to be an even larger and long-lasting node for TSMC than N2. Just like 2-nanometer technology is a larger and longer-lasting node than 3-nanometer, and here further extend our technology leadership position well into the future…

…We’re also moving the new fabs and the facilities in Taiwan as fast as possible. And the same thing, we try to bring up a new fab in the Japan as fast as possible. Because of the situation today is the demand and the supply, the gap is so big. So we are working very hard to narrow the gap.

TSMC’s management has demonstrated 90% device performance and 90% SRAM yield; management is seeing strong customer interest and engagement from smartphone and HPC AI applications

Internal product-like vehicle demonstrated close to 90% device performance and close to 90% 256 megabits SRAM yield. We are observing a strong level of customer interest and engagement from both smartphone and HPC AI applications and customer now tape-out activity is ongoing and ahead of schedule. Pre-production will start in 2027 and volume production is scheduled for 2028.

When dealing with foundry competition, TSMC’s management thinks 3 things are the most important, namely, technology, manufacturing, and customer trust; management is not that concerned about foundry competition because they think choosing a foundry partner requires deep work to understand the foundry’s technological road map and it’s not as simple as changing a brand of milk from a convenience store; management is welcoming competing advanced packaging technologies because it lessens TSMC’s load and provides more flexibility for customers; advanced packaging is a back-end business, and it’s where TSMC is less worried about competition

The most important thing as we continue to say is the technology, manufacturing, and customer trust. These 3 fundamental never change. For my 30-some years, 40 years career, it’s always the most important thing. And that’s always the TSMC’s secret recipe to win the business. So from my – from the competition point of view, choosing a technology, ramping it up is not buying a milk from 7-Eleven. Well, I’m using that — I’m quoting the sentence for my customer, anyway. It says that you’re choosing a kind of a technology partner, it is no shortcut. You need to understand the technology. You need to really utilize it using the test chip, and then something, and work together, and then prepare the capacity and ramp it up. That’s why I would say, it takes about 5 years. It’s not that today, you think this milk is better, you go to the next store, it’s a 7-Eleven. You don’t like it, you go to another store. No…

…[Question] My question is regarding the new advanced packaging technology. We noticed that especially the EMIB-T is gaining traction. So how will TSMC react this request?

[Answer] Our packaging capacity is so tight that now it’s limited by customers’ growth. So we welcome that additional flexibility in the market. And so that will help TSMC’s front-end wafer business growth, which is a majority part of TSMC’s business. The technology looks good, according to the newspaper. And we hope they will be successful and so that share some of the loading from TSMC. Today, we’re working very hard to shorten the gap between the demand and the capacity. And so as I said, we welcome to have this additional alternatives, and so the flexibility for my customer…

…[Question] If these technologies have some small problem, and then ask our company to support. So how our company accommodate it?

[Answer] Our #1 is to support our customers’ success. So whatever that we can hear about our customers’ business, we want to win…

…The front-end’s wafer business and the back-end’s business are 2 different things, right? If they are the same, then you can expect ASE become the front-end competitor also. It’s 2 different things. And I also say that since our capacity in the back-end is so in shortage mode, the gap is bigger. And so, I welcome that the competitor offers some of the flexibility to my customer so that their front-end wafer can be put into the package, and that help TSMC’s front-end wafer business. 

TSMC’s management is diligently checking to make sure its AI chips do not end up in its customers’ inventory and are actually being put into production

Are we sure that we deliver the chips to our customer, and they were not put into inventory? So we — actually, we are checking the AI data centers progress, the building, the location, the demand, the racks, we’re checking all that to make sure that TSMC chips will not be put in inventory.

TSMC’s management is not worried about customer concentration for AI-related demand; management sees a lot of new players in the AI industry

[Question] Could I ask about the risk that you see around customer concentration, as AI demand continues to significantly outgrow other end markets? I think your exposure to your top 5 customers is becoming meaningfully larger than at any point in your history.

[Answer] No, that’s not our concern. Besides what you say the customers are growing bigger and bigger, we are very happy about it. And some of the customers also growing very fast. So it’s not — Jim, it’s not what you said that the bigger customer is growing bigger and bigger. No. I mean that’s — there’s a lot of new player in the AI industry.

TSMC has started production for its COUPE (Compact Universal Photonic Engine) platform; COUPE is TSMC’s silicon photonics platform designed to send data using light instead of copper wire; management believes AI data centers will need to lower power consumption while increasing communication bandwidth, and these are important functions of the COUPE platform; management believes COUPE’s demand will grow in the next few years and it will become a fairly important technology for TSMC

[Question] When should we expect the COUPE platform to have a material contribution to your top line?

[Answer] We start the production right now, and it will be ramped up. As time goes by, I think the AI data center need to lower down the power consumption and increase the bandwidth of the communication channel. So I believe the COUPE will continue to increase the demand, and then will become a fairly important technology in the next few years.

All the silicon-roads for AI lead to TSMC

[Question] You mentioned about the Agentic AI and the CPU growth potential. But can you give us more update among that AI, different kind of chips between GPU, accelerators or CPU? What you see the growth potential and your visibility?

[Answer] I don’t think I can give you a very specific number. But let me share with you. All of them are in TSMC. And they’re also using the same kind of leading-edge technologies. We’re working with our customers to allocate the wafer, the supply to balance the CPU, GPU, XPUs ratio.

In mature nodes, there is a shortage for nodes that are related to AI, such as for power management ICs (integrated circuits); the demand for mature nodes from consumer products is not high

The mature node cover a lot of different segments. Only the one which related to AI is in shortage, which is the most important one, is the #1 is power management IC, because all the AI data centers need a lot of power management. And those are the mature node technology like 0.18 micron, 90-nanometer or something like that. Those are in shortage definitely. And also the sensor portion because of — you need a lot of sensor to detect the environmental information and put into the AI data center to analyze it. Other than that, other area, just like you pointed out, the consumer product is not in a high demand. And so other segment is not so strong demand. And as I pointed out in my statement, other area, no, it’s not so much of, say, in a lot of shortage, not at all.

Tesla (NASDAQ: TSLA)

Tesla’s management is seeing very high take rate of FSD in locations where it is approved; management thinks consumers are purchasing FSD primarily, instead of buying Tesla vehicles; management thinks Tesla will enjoy similar uptick in demand as it gets approval for FSD in different countries; 55% of Tesla’s vehicle deliveries in North America in 2026 Q2 had FSD subscriptions enabled; FSD now has 1.5 million paid customers globally (1.3 million in 2026 Q1), 55% of which are paid upfront and 45% are subscriptions; management has turned off upfront payments for FSD, so they expect FSD monetisation to come from subscriptions in the future; the Robotaxi fleet is currently running v15 (version 15) of FSD; v15 has 7 major parallel tracks of improvement over v14; Robotaxi’s v15 already has 40% of the tracks merged together; management thinks v15 will further improve the safety profile of the Robotaxi fleet; management thinks it makes sense to upgrade all Tesla vehicles to at least Hardware 4 

We’re seeing in locations that have FSD approved, we’re seeing a very high take rate of FSD. In fact, I think for a lot of people, they’re actually buying Tesla Full Self-Driving with a car attached, as opposed to a car with FSD. They’re coming into our stores in the U.S. and telling me they want the Full Self-Driving and with whatever car it comes with, essentially. Clearly this is a significant demand driver and as we get approval for FSD in different countries, I think we’ll see a similar uptick in demand…

…In Q2, we had, in North America, about 55% of our deliveries had FSD subscription at the time of delivery enabled. Overall, FSD attach rates continue to improve, reaching nearly 1.5 million paid customers globally, of which 55% is upfront purchases and the remaining 45% is subscriptions. We expect that the bulk of the growth in FSD monetization will come from subscriptions as we’ve removed the purchase option in most markets…

…The currently operating Robotaxi fleet is already running early versions of the V15 FSD software that we had referred to in the past. For V15, we had planned roughly about seven major improvement tracks, and they’re all happening in parallel. The early V15 builds that are running on Robotaxi have already 40% of those tracks merged together, and that’s what’s running in the fleet right now. As we continue to complete our work on V15, we will see that the car is going to be ridiculously safe and capable…

…I think it’s going to make sense to upgrade all cars that have less than Hardware four. Any cars that have cameras, basically, because otherwise it would be probably too many modifications. Anything that’s set up for cameras, it’ll be financially sensible at some point to upgrade them. I think we’d want to upgrade them to the next generation of AI board.

Tesla’s management thinks the company’s energy business will be crucial for scaling up AI data centers; power constraints are a major issue for the deployment of AI data centers; power can cycle dramatically in AI data centers, especially for training runs, so the data centers require fast-acting power electronics to smooth out the huge changes in power 

The energy business is also growing incredibly fast and, I think will be crucial for the scale-up of artificial intelligence data centers…

…We think power constraints are going to be, they already are a major issue for AI. Just turning on the AI computers, the AI compute demand is so high that even the hyperscalers are having trouble turning on their AI compute and finding the power, and then smoothing the power, especially for the training runs where the power cycles dramatically in a very short period of time. You can have, during a training run, the power consumption can drop by 70% for 100 milliseconds. You really need fast-acting, advanced power electronics to be able to smooth out the massive changes in power, especially during their training runs. That’s why SpaceX has bought so many Megapacks for the data centers. It’s actually mostly for smoothing out the power for the training runs.

Tesla’s management said during the 2026 Q1 call that the company will increase capital expenditure significantly, partly for AI-related investments; management is confident that the capex will yield massive returns; the higher capex in 2026 Q2 caused free cash flow to become negative; management continues to expect capex for 2026 to be more than $25 billion; management is tapping on the debt markets to help fund Tesla’s capex

This is a massive CapEx year. I’m confident that all the things that we’re investing in will yield incredible returns. Really, maybe the best CapEx returns that we’ve ever seen…

…Our free cash flow ended up being negative for the quarter. Most of the reason for it going negative is because CapEx more than doubled sequentially. We expect it to increase further in the second half of 2026. We continue to expect that CapEx for this year will be more than $25 billion. CapEx will grow for the next two or three years as we expand our Robotaxi fleet, expand our production capacity for Optimus, make investments for semiconductor fab, install solar manufacturing capacity, and AI compute infrastructure, in addition to all the other expansions we’ll do for other manufacturing for automotive. In addition to using our cash for such investments, we are being opportunistic in securing certain debt facilities that will give us the capacity to borrow up to $30 billion to help accelerate such investments.

Tesla’s management will soon start production for Optimus, the company’s autonomous humanoid robot; management continues to think Optimus will be the biggest product ever; no one has ever built an autonomous humanoid robot that can do tasks without any programming; there are many challenges in the electromechanical design of Optimus; management foresees very substantial challenges in scaling the production of Optimus, as the degree of difficulty is proportionate to the newness of a robotic part, and there is no existing supply chain for Optimus; Tesla has in-house a significant amount of Optimus’ production; Tesla has an Optimus production line in Fremont, California; management expects the initial part of the Optimus production S-curve to be long and flat; management thinks a lot of videos of robots on the internet are pre-programmed or remote controlled; management believes Optimus will be the first humanoid robot capable of doing generalised tasks; Optimus will come with hands that have the same dexterity as the human hand, and then eventually have superhuman dexterity; management is confident that the same AI technology for self-driving cars can be used for a digital version of Optimus; the development of Digital Optimus is in partnership with SpaceX’s (an Elon Musk company) Grok models; management thinks having Optimus take on the human form-factor gives the robot access to huge amount of learning-data as the form factor allows the robot to learn directly from humans; workers at Tesla factories are providing the data for Optimus to learn from; Optimus benefits from a reinforcement learning loop when a large number of robots practice their tasks; management has the same end-to-end AI approach for Optimus’s AI as they do for FSD; management sees Optimus 4 as having much higher vertical integration than Optimus 3; management sees 10 million units a year for Optimus 4, compared to 1 million units annually for Optimus 3

We will soon start production with Optimus…

…I think Optimus will be the biggest product ever…

…It’s one of the hardest things to solve, to make an autonomous humanoid robot that can do tasks that if you simply ask it to do something or show it a video, it can do the task without any programming. No one’s ever achieved this. There are many challenges in the electromechanical design of the robot to achieve sufficient dexterity, also to be very reliable and have long wear and tear. Meaning, it needs to be out in the field and not break down…

…The production scaling challenge is very substantial. This is going to be the hardest product to scale manufacturing that we’ve ever made at Tesla, because everything on the robot is new. The difficulty of scaling the production ramp is proportionate to the newness of the parts in the robot…

…With Optimus, there is no supply chain. We’ve had to build up a supply chain in its entirety, or in-house the production. We actually have in-housed a tremendous amount. The Optimus production line that we’re building out in Fremont, in place of what used to be the Model S, X production, it looks incredible…

…Optimus will follow the normal S-curve of a manufacturing ramp, but the initial portion of the S-curve will be quite flat and long because of the newness of the parts in the robot. 

You’ve probably seen lots of impressive demonstrations of robots on the internet, but those demonstrations you’re seeing are pre-programmed or remote controlled. There is no humanoid robot that is actually able to do generalized tasks. Optimus will be the first one that is capable of doing that, where it’s not just a demo, it’s genuinely useful in day-to-day life. And Optimus is designed to have full human dexterity. A hand that has the same level of dexterity, if not higher, than a human hand. The human hand is an incredible thing. The more you study the human hand, the more you realize how amazing hands are. It’s more than just opposable thumbs. The nuances of how human hands work are amazing. The closer you look, the more amazed you are. Optimus will have that capability. It will have human and then superhuman dexterity…

…We have Digital Optimus, which is basically driving a computer screen, I guess you could call it computer use or something like that, but driving a computer screen in the same way you drive a car. For the car, it’s pixels in or photons in and controls out. The same thing is true for Optimus and Digital Optimus. It’s photons in, controls out. We feel confident we can adapt the same Tesla AI technology that we developed for self-driving cars to have a self-driving computer screen or self-driving computer, essentially. The self-driving computer, where the very low cost, Tesla AI computer can handle all of the sort of real-time tasks. They’re doing real-time video control of the screen of the computer. It’s not like screenshots type of thing, it’s real-time video, at high frame rate. Digital Optimus will be important, obviously, for physical Optimus because physical Optimus needs to be able to operate computers. It can’t come up to a touch screen and not know what to do. It’s got to have a generalized touch screen and computer use capability.

That’s looking promising and this is in partnership with SpaceX. SpaceX’s Grok, sort of the big model that is the manager of Digital Optimus and tells Digital Optimus what to do, provides it with a series of tasks, then Digital Optimus goes and does those tasks…

…Optimus has been designed to not just match the appearance of human beings, but also the functionality and dexterity of humans. One of the main reasons for this approach is that having the human form factor and function allows us to learn from humans on how to perform a wide variety of tasks. This opens up the entire world to provide data for training Optimus. Just like FSD, we have access to a broad fleet of humans giving us data from all of the workers at our factory…

…When we have a large number of Optimus robots practicing their tasks in what we call the Optimus Academy. The data from the bots experiencing the task themselves will be invaluable and help us close any minor form factor gaps that may exist between the bot and the humans. This is also when the reinforcement learning loop kicks in, where the bot initially attempts some tasks, fails sometimes, learns from both the successes and failures of those tasks, and eventually learns to master those tasks at perhaps a superhuman level. Our AI strategy for Optimus is aligned the same. It’s the same end-to-end strategy that drives FSD, pixels in, controls out. Just like FSD, we expect it to work broadly. In FSD, you can get in the car, type in an address, and then hit start, and it just handles all of driving from park to park. The same thing is going to be true for Optimus as well. You’re going to just ask it to do anything, and then it should just perform the entire task on its own without you having to do anything along the way…

…For Optimus 4, which will be built in Austin, that will be a much more vertically integrated supply system for Optimus 4. That would aim to have an order of magnitude more production of Optimus 4 than Optimus 3. Sort of aspirationally 10 million units a year versus 1 million units a year of Optimus 3. With all the caveats there, which is insanely difficult to scale production.

Tesla’s management is trying to scale Robotaxi as fast as possible, but they are doing so while being very careful not to harm anyone; Robotaxi has been introduced in a number of cities in Florida, Texas, and the Bay Area; management expects Robotaxi’s miles driven to increase by more than 10% a week; Robotaxi is in 7 markets in the US currently; management expects to accelerate the ramp of Robotaxi throughout the year and expand into new US markets; Robotaxi has driven more than 380,000 miles unsupervised with zero incidents; Robotaxi did not require LIDAR, radars, HD maps and more to drive safely; management thinks Robotaxi’s current safety record is validation of the company’s entire AI approach to autonomy; Robotaxi’s number of unsupervised miles has been growing double-digits every week since the start of 2026; it has been taking management relatively less effort to launch Robotaxi in new cities and management expects the time to launch in a new city to trend towards zero; management expects the Robotaxi business to be fully vertically integrated; management thinks Robotaxi will not face any demand challenges; the Robotaxi fleet is still small (only in the dozens) but management wants to expand into new cities to test the generalisability of Robotaxi’s autonomy technology; vehicles in the Robotaxi fleet are driving 24/7, so Tesla can get a lot of miles from them

There are, I think, 30,000 – 40,000 automotive deaths per year in the U.S. alone, most of those do not generate any press or maybe, you never really read about almost any of those. If we injure even one person, it’ll be worldwide headline news, and regulators will immediately clamp down on our activities. We don’t want to injure anyone. We’re going as fast as humanly possible in scaling Robotaxi, but while trying to ensure that we do not harm anyone at all, and ideally do not even run over a pet. That’s really the constraint is we want to grow as fast as possible with Robotaxi without harm to anyone…

…We’ve opened up in a number of cities in Florida and in Texas, obviously in the Bay Area. We’ll continue to scale, I think, very rapidly with more than 10% a week in terms of miles driven…

…We continue to grow the Tesla Robotaxi fleet and have expanded to a total of seven markets in the U.S. We expect the ramp of the fleet to accelerate throughout the year, along with expansion into new U.S. markets…

…In terms of safety, the program has had an impeccable safety record. We have driven more than 380,000 miles of unsupervised Robotaxi, now across six cities in two different states. We have had zero notable incidents. Any reports have been of other actors impacting us when we were stationary…

…Historically, the so-called experts have always claimed that you need LiDARs, radars, HD maps, and the entire kitchen sink to drive safely. Here we show that such is not true. You can have safe, comfortable, and affordable autonomy with just cameras. This record should be a huge validation of Tesla’s entire AI approach…

…Since the beginning of this year, we have grown at double-digit growth rates to the number of unsupervised miles that the fleet drives every week…

…For expanding to new cities, it has been relatively less effort on our front. We expect that the time to launch to a new city will continue to trend towards zero, towards an end where we operate in entire states as a whole instead of going city by city…

…[Question] Would you ever consider third-party distribution partnerships, such as with rideshare providers, to increase utilization? Is the plan for now to keep Robotaxi fully vertically integrated?

[Answer] We expect to be vertically integrated with Robotaxi as we are in the rest of our business. I don’t think we’re going to have any demand challenges with Robotaxi. The economics will be so compelling that I think we will really have a lot more desire to use the service than I think demand will outstrip our ability to service the demand…

…[Question] On Robotaxi. If I look at the launches, you’ve been adding cities, the number of units is still, it looks like based on media reports, sort of in the dozens as opposed to hundreds. Why not just sort of scale up Austin or one or two cities before adding cities, what do you need to sort of get that higher volume numbers in a major city? What sort is the roadblock to start adding more vehicles on the ground?

[Answer] The reason we have been expanding across different cities instead of just doubling down on a single city, is that we want to make sure that our stack is a very general one. It is a general one. We just want to both prove to ourselves and to other folks that it is working across a lot of different cities without too much effort per city…

…In terms of miles versus vehicles, since these vehicles operating the Robotaxi fleet drive basically continuously as opposed to human drivers who use vehicles for maybe a couple of hours a day or something like that. These vehicles are in mostly continuous operation, which means that even for a few vehicles, you can get a lot of miles out of them, that’s why we are tracking the amount of unsupervised miles.

Tesla’s management expects to announce the location for the TeraFab soon; management believes TeraFab is a necessary project for Tesla because the company will be constrained by the supply of AI chips, otherwise, for the scaling of Optimus; management has placed semiconductor manufacturing orders for the development fab in Austin; the development fab will have lithography mask production, logic, memory, packaging, and chip testing all under one roof, which is unique among fabs, and gives it a rapid iterative cycle to test high-risk ideas

The Terafab, we expect to announce a location soon, and provide more details about our plans in that regard…

…I do think Terafab is going to be an amazing initiative and a necessary one, and one without which we will be constrained in our ability to scale Optimus production, because we simply won’t have enough AI chips. It’s crucial to solve that, and we’ll have to solve memory, logic, and packaging in order to scale Optimus…

…We’ve placed equipment orders for our development fab in Austin. That development fab, I think, is pretty cool because it’s intended to have lithography mask production, and then logic, memory and packaging and chip testing all under one roof. You can have a very fast iterative cycle, and try out new chip designs very quickly and see if they work. I don’t think such a building exists anywhere on Earth. This is really going to be super helpful as we try some exciting, adventurous, high-risk, high-payoff bets on AI chips.

Tesla’s management is building a Megapod design that combines an AI computer with an x86 CPU; the Megapod comes in a box that can be placed anywhere, and thus can allow Tesla to scale AI compute with disaggregated electricity production, such as its Supercharger network; Tesla’s Supercharger network has 7GW of power, and growing

We’re also building out a Megapod design that has Tesla AI computer with x86. It’s a pairing, an x86 computer with a Tesla AI computer in a box, and it’s got this Digital Optimus in a box, and in a Megapod, kind of like the Megapack packaging. Where we put just a large number of AI full plus x86 combos in a giant box, essentially. These boxes can be placed anywhere in the country or outside the country. This allows us to scale AI compute using disaggregated electricity production. Because there’s lots of places all around the world, including at our Superchargers. I think we’ve got something like seven gigawatts of power at our Superchargers and growing. We can place Megapods at many of these Superchargers and have distributed power for AI.

Tesla’s management thinks the Cybercab, which is the autonomous vehicle for the company’s Robotaxi fleet, is a phenomenal product; the manufacturing targets for Cybercab are now roughly aligned with the growth of Robotaxi’s unsupervised miles; v15 of FSD will also work on Cybercab; the Cybercab is a new form factor, so Tesla needs to accumulate driving data that is specific to the Cybercab before the company can put a lot of them on the road

Regarding the Cybercab, first of all, it’s a phenomenal product. Anyone who rides in it instantly falls in love with the experience. We have aligned our manufacturing targets to roughly match the projected growth rate of the unsupervised miles. The same V15 models that power the Model Y and other platforms will also work on Cybercabs…

…Because it is a new vehicle chassis, we need to accumulate driving data that is specific to the Cybercab before we can put a lot of them on the road.

Tesla’s management is seeing Tesla’s suppliers making investments to support Optimus; Samsung and Micron are providing good support to Tesla

Our suppliers have been great, they have made and are making tremendous investments in support of Optimus and Robotaxi and whatnot. Samsung and TSMC, in particular, are building fabs, TSMC in Arizona and Samsung in Texas, and putting in tens of billions to build AI compute for Optimus and Robotaxi. Panasonic has also invested many billions in increasing battery cell production…

…The Samsung fab, that’s going to be pretty significantly dedicated to future projects. That’s a massive investment, multi-billion dollar. We’re seeing the same level of investment going to memory and also new specific items like metal injection molded parts, flexible printed circuits, and all sorts of nonlinear technologies that are more based for the robot as opposed to the traditional vehicle supply chain we’ve had…

…I’d actually also like to thank Micron for giving us memory allocation. They’ve got to make some very tough decisions on memory allocation. We really appreciate Micron making room for Tesla in the years to come and giving us actually a very significant allocation on reasonable terms given the pretty insane pricing of memory these days.

There are increasing levels of collaboration between Tesla and SpaceX, and management appears to be hinting at some form of combination; Starlink, in particular, is important for Tesla because Starlink ensures that Robotaxis have constant internet connectivity

As you can tell from all the many collaborations on so many fronts with SpaceX, there’s more and more overlap, especially with Terafab, that’s really going to be a gigantic project. Obviously, we can’t talk about combining companies and that kind of thing on an earnings call. It’s got to be done with the appropriate process…

…We continue to benefit from our relationship with SpaceX, they’ve been a great partner, we have numerous beneficial transactions with them. Earlier this year, we deepened our relationship through an investment and a framework agreement. This will allow us to continue to work with them on projects that Elon mentioned, like Terafab and Digital Optimus…

…You’ve got Grok in the car. Grok helping drive Digital Optimus. You also got Starlink being integrated into the Cybercab, Starlink will be integrated into all our car vehicles, at least for markets that Starlink is active. Because for a Robotaxi situation, you need to have coverage everywhere. There are many places, even in Silicon Valley, where the cellular coverage is terrible or sometimes non-existent, which is surprising for Silicon Valley. I know when I drive to work, those first 10, 15 minutes, I can’t actually do any calls because the cellular connectivity is so bad. We can’t have Robotaxis getting stuck in these Bermuda Triangles of lack of cellular connectivity. Starlink with its ability to do connectivity anywhere is actually quite important, so we don’t have Robotaxis missing in action. Obviously if people are sitting in the car, they go, “Blip,” and want to do high productivity stuff or want entertainment. With Starlink, you can watch 4K live sports in the car and with very low cost per gigabyte of data, as that’s really not feasible via the cellular system.

There are a shortage of truck drivers in the USA and a fully autonomous Tesla Semi will be important to address the shortage; management’s focus at the moment is on solving autonomy for Tesla’s high-volume vehicles before doing to for the Semi, which is a low-volume vehicle; management thinks autonomy for the Semi will come in 2027, and in time for the scaling up towards volume-production of the Semi

There is a really serious shortage of truckers. There just aren’t enough people around who want to drive trucks, which are crucial for transport throughout America. An autonomous Semi is actually going to be very important to address the shortage of truck drivers. Obviously will be great for improving safety, and making it easier on truck drivers to have a self-driving Semi is going to dramatically improve safety and comfort for truck drivers that use the Tesla Semi. Since the total number of units of the Tesla Semi is still low, and will continue to be a very small percentage, even by the end of this year, of our total vehicle fleet. It makes sense for us to focus our self-driving efforts on our high-volume vehicles, Model 3 and Y, and solving self-driving for, and really getting to the point where it’s generalized, unsupervised self-driving for those vehicles, and Cybercab. We expect to get self-driving working on the Tesla Semi probably around the end of this year or early next year. I just don’t want it to be a distraction on the march of nines of safety for self-driving on three, Y, and Cybercab. It’s taking a bit of a backseat for the next six months or so for autonomous Semi. It will definitely be working next year and in time for the scale-up to high production of the Tesla Semi.

The Tesla chip team is making fast progress on the AI5 chip; management is very excited about the design of the AI6 chip as they think it’s going to be the best edge computing chip in the world

The Tesla chip team is really doing great work and making incredibly fast progress on AI5. I’m very excited about the design of the Tesla AI6 chip. I think it’s going to be the best edge computing chip in the world.

Visa (NASDAQ: V)

When generative AI first emerged, Visa’s management quickly deployed the technology to engineering, client service, and model orchestration; with the emergence of agentic AI, management has used the technology to build end-to-end pipelines with human oversight and autonomous capability, leading to a reduction in the size of product development teams from 10 or more people to 2 or 4; teams in Visa that use agentic AI have showed 80% more code commits and a 65% increase in the pace of feature development; Visa now has more than 150 AI-powered applications; in the last 12 months, Visa has shipped more than 300 major product releases with the help of AI; management has eliminated certain roles in Visa to fund AI-related investments; Visa has an AI financial assistant, which allows banks to provide white-lablled AI insights to cardholders; Visa has the Vulnerability Agentic Harness to allow users to use frontier models to fix problems; Visa has increased the velocity of its consulting projects with the help of AI; management thinks Visa still has enormous opportunity ahead to deploy AI within the company

AI is changing how work gets done at Visa. With the dawn of the generative AI era, we moved quickly to deploy AI across our enterprise to assist us in areas like engineering, client service, and model orchestration. As we enter the era of agentic AI, we are going beyond AI assistance and harnessing the power of AI to execute work and tasks with our supervision. We have progressed materially in product development and engineering, deploying new tools, plugins, agent skills, and persistent sessions to create an end-to-end pipeline with human oversight and autonomous capability. As a result of the unlocks we can realize with this new tooling, we are reforming our product development teams that used to be 10 or more into smaller and more nimble agentic squads of two to four. The results are meaningful for those teams that are using the agentic tool chain, with 80% more code commits and 80% plus improvement in requirement definition from 30 days to five days, which has translated to 65% plus faster feature development…

…We now have more than 150 AI-powered applications, and over the last 12 months, we have shipped more than 300 major product releases…

…Today, we announced that we are eliminating roles, with the majority being in our technology and product teams, to ensure that we are continuing to position Visa for future growth…

…The investments that we have in front of us are enormous. I think what we’ve shown over the last couple of years is that we can drive efficiencies, we can take the savings that we generate from those efficiencies, and we can invest those savings against the strategic levers that we laid out at our strategy at our Investor Day and deliver accelerated performance as a result….

…We continue to develop new products, including our AI financial assistant, enabling banks to white label our AI-powered financial insights from their data and Visa’s network data for their cardholders right in the bank’s own app and website…

…We built the Visa Vulnerability Agentic Harness, an orchestration layer that allows us to use models like Mythos to find and fix issues at AI speed. It is available now on GitHub to our clients, along with a technical blueprint, remediation, and validation agents…

…For our advisory and other portfolio, in addition to the strong marketing services revenue growth, we’ve also increased the velocity of our consulting projects through the help of AI. Just this past quarter alone, for over 700 clients across 100-plus countries and territories, we delivered 1,200 consulting projects, which is more than we delivered for all of 2019…

…I think we still have enormous opportunity ahead of us there, and we’re just going to continue to lean in to using these tools to drive efficiency and effectiveness, shipping products better, faster, and ultimately, better serving our clients in the ecosystem.

Visa’s new agentic-AI-driven way of working has made big impacts on its initiatives for stablecoins and agentic commerce; management believes that agentic commerce will expand Visa’s addressable market; Visa recently enabled new seller capabilities and infrastructure for agentic commerce; Visa is partnering with OpenAI to enable secure payments in agentic commerce; Visa is partnering with Meta Platforms to enable consumers to transact in Facebook and Instagram seamlessly and securely with Visa tokens; management thinks the world of commerce is still only in the very early stages of a major adoption curve for agentic payments; management thinks the adoption curve of agentic payments will follow the same patterns as previous technological cycles; management thinks that trust will be the ultimate accelerant for adoption of agentic payments; Visa’s products within agentic commerce have all been about building trust for consumers

As a result of this new way of working, we are able to design, build, and ship products at an increased velocity with continuous innovation and improvement. I want to touch on two areas where we are deploying our new way of working with great impact, stablecoin and agentic commerce…

…If stablecoins are reshaping the back end of commerce, we see AI as transforming the front end. We believe agentic commerce will expand our addressable market and drive future growth for Visa. This quarter, we continued to work across the ecosystem by enabling new seller capabilities such as our Agent Score and Agent Directory and building infrastructure such as our Token Assurance Framework to ensure agent-initiated transactions are transparent and trusted…

…We are excited to be partnering with OpenAI to enable secure Visa payments within agentic commerce. Through the partnership, Visa will provide its global network, credentialing capabilities, and security infrastructure to support agentic commerce experiences, helping consumers and businesses interact and transact with confidence. As part of our partnership with Meta, Visa is enabling new ways to pay across Facebook and Instagram, powered by Visa Intelligent Commerce, allowing consumers to transact seamlessly and securely with Visa tokens…

…We believe that AI and agentic commerce will expand our addressable market. We believe we’re in the very early stages of what’s going to be a major adoption curve in payments. I think to get a sense of how this progresses from here, it’s instructive to look at other kind of major cycles that we’ve been through, whether it was e-commerce or mobile commerce, tokenization, tap-to-pay. These innovations and these kind of major forces, they followed a similar pattern, right? You have an early period where Visa and other players are establishing standards, we’re announcing, launching, and shipping new products, as you mentioned. You migrate into the early adoption period of the curve, which ultimately then leads to growing consumer momentum and ultimately broad scale. All of those kind of ones that I mentioned previously have gone through that, and they’ve achieved that broad scale. We don’t think agentic commerce will be any different…

…I think the ultimate thing that’s going to accelerate that adoption is going to be trust. Trust that the payment is secure, trust that the agent is authorized, trust that the transaction reflects the consumer’s actual intent, and then the protections exist if something goes wrong. If you look at the products that we’ve announced over the last several quarters, they’re all directly intended to address that trust and ensure that our users have trust in using Visa credentials to make agentic commerce transactions. That’s where we’re investing.

---

Disclaimer: The Good Investors is the personal investing blog of two simple guys who are passionate about educating Singaporeans about stock market investing. By using this Site, you specifically agree that none of the information provided constitutes financial, investment, or other professional advice. It is only intended to provide education. Speak with a professional before making important decisions about your money, your professional life, or even your personal life. I have a vested interest in Alphabet, Amazon, Apple, ASML, Mastercard, Meta Platforms, Microsoft, Netflix, TSMC, and Visa. Holdings are subject to change at any time.

Mastercard (NYSE: MA) and Visa (NYSE: V) are two of the largest payments companies in the world. As a result, they have a great view on consumer spending that’s taking place. With both companies reporting their earnings results for the second quarter of 2026 earlier this week, the bottom line is that consumer spending remains strong in the USA and other parts of the world. Here’s what they are seeing.

*What’s shown in italics between the two horizontal lines below are quotes from Mastercard and Visa’s management teams that I picked up from their earnings conference calls.

---

From Mastercard

1. Mastercard’s management sees consumers and businesses being healthy and continuing to spend and they are supported by positive job growth, low unemployment, and purchasing power; management is monitoring geopolitical risks; management is seeing economies around the world adapting to changing conditions; the fundamentals of consumer and business spending and travel remain healthy

Consumers and businesses are healthy and continue to spend, supported by positive job growth, low unemployment, and real purchasing power in many major economies. At the same time, we continue to monitor geopolitical uncertainty and its potential economic impacts…

…Around the world, economies are adapting to changing conditions, with consumers and businesses continuing to demonstrate resilience…

…Overall, the underlying fundamentals of consumer and business spending and travel remain healthy.

2. Worldwide GDV (gross dollar volume) was up 8% year-on-year in 2026 Q2 in constant-currency basis; cross-border volume was up 12% globally in constant-currency, driven by both travel and non-travel cross-border spending (cross-border volume growth was 13% in 2026 Q1); switched transactions was up 9% year-on-year in 2026 Q2; card growth was 5% in 2026 Q2, with Mastercard ending the quarter with 3.7 billion cards in circulation (there were 3.7 billion cards in 2026 Q1, and year-on-year growth was 5% then); domestic assessments were up 10%, cross-border assessments were up 20% and transaction processing assessments were up 12%

I’ll speak to the growth rates of our key volume drivers for the second quarter on a local currency basis. Worldwide gross dollar volume, or GDV, increased by 8% year-over-year. In the U.S., GDV increased by 6%, with credit growth of 10% and debit growth of 1%. As a reminder, the Capital One debit portfolio migration was basically complete in Q1. Excluding the impacts from that migration, our U.S. debit GDV growth would have been 8%. Outside of the U.S., GDV increased 9%, with credit growth of 9% and debit growth of 10%. Cross-border volume increased 12% globally for the quarter, reflecting continued growth in both travel and non-travel related cross-border spending…

…Switched transactions grew 9% year-over-year in Q2…

…Card growth was 5%. Globally, there are 3.7 billion Mastercard and Maestro-branded cards issued…

…All growth rates are described on a currency-neutral basis unless otherwise noted. Looking quickly at each key metric. Domestic assessments were up 10%, while worldwide GDV grew 8%. The two PPT difference is primarily driven by pricing. Cross-border assessments increased 20%, while cross-border volumes increased 12%. The eight PPT difference is driven primarily by pricing in international markets and mix. Transaction processing assessments were up 12%, while switched transactions grew 9%. The three PPT difference is primarily due to favorable mix and pricing, partially offset by lower revenue from FX volatility, and other network assessments were $326 million this quarter.

3. In 2026 Q2, Mastercard’s operating metrics had good year-on-year growth and were stable sequentially; in July 2026 so far, Mastercard’s operating metrics continue to be strong with worldwide switched volume growth of 9% (6% in the USA, and 11% outside of the USA), switched transactions growth of 9%, and cross-border volume growth of 11%; card-not-present ex-travel’s sequential decline in growth rate was driven by timing; Mastercard’s US business had some benefit from the World Cup in 2026 Q2, but it was hard to quantify; Mastercard’s US business continues to have healthy consumer and business spending trends; management is seeing strong spending trends in both mass and affluent consumers in the US and around the world, although affluent consumers have higher growth in spending

Let me comment on the operating metric trends for Q2 and the first four weeks of July. Switched metrics were generally in line with Q1, and underlying spend remained stable. Of note, excluding Capital One debit, on a like-for-like basis, U.S. switched volume growth was 10%, or two PPT higher sequentially. This increase was driven by higher spend on fuel and overall strong consumer and business spending.

Moving to our cross-border metrics, our overall cross-border volume growth remained healthy at 12% in the second quarter. Cross-border card-not-present ex-travel remained strong at 20%, benefiting from increased card-not-present spend from Venezuela and the timing of large retail promotional events. While cross-border travel was down sequentially, relative to the April metrics we discussed on our last earnings call, we saw improved growth in the quarter due to lower impacts from the developments in the Middle East and timing of holidays.

As we look at the first four weeks of July, our metrics remain relatively stable and strong. Looking specifically at card-not-present ex travel, let’s focus on July compared to June. The sequential decline is primarily driven by timing, including the large retail promotional events that happened in June this year as compared to July last year, and by mix of days…

…The strong underlying consumer and business spending, which we’re seeing in the U.S. There has been a tailwind which has come on account of higher fuel prices, so let’s recognize that. You probably have some impact coming through from the World Cup as well, as it relates to the second quarter in particular. Hard to really quantify what that is just because we can’t really identify exactly what that is. I would tell you the underlying consumer and business spending trends continue to hold up well in the U.S. To your point, if I look at it’s broad-based. We see it across credit and debit. We see it across consumer and commercial…

…I would say they’re holding up well, both across mass and affluent. Certainly in the U.S., but across the world as well. We try and track the best we can as it relates to what we’re seeing in spending patterns based on the product codes that are out in the market, which serve the different categories of customers. When we look at that, we’re seeing generally strong trends, across both mass and affluent. What you do tend to see is higher growth in the affluent side of spending. That’s kind of not a new phenomenon. That’s been with us for some time now.

From Visa

1. US payments volume growth was good at 10% in 2026 Q2 (FY2026 Q3), a growth rate not seen since FY2019; there was good growth in both US credit and debit volumes; growth across consumer spend bands improved sequentially, with the highest spend band continuing to grow the fastest; both discretionary and non-discretionary spend remained strong; management did not see a deterioration in spend in the lower bands

U.S. payment volume grew 10% year-over-year, up about two points from Q2, a growth rate not seen since fiscal 2019, excluding the post-COVID recovery, with both card present and card not present growth accelerating strongly. U.S. payments volume growth was the result of several factors, including higher tax refunds, the cost of fuel, retail, including the timing of promotional shopping events, strong Visa Direct growth, and FIFA-related spend. U.S. credit rose 11% year-over-year, up more than a point from Q2. Debit accelerated by more than two points from Q2 to grow 9% year-over-year. Growth across consumer spend band saw incremental improvement from Q2, with the highest spend band continuing to grow the fastest. Across our volume, both discretionary and non-discretionary spend remained strong. We do not see signs of the lower spend consumer weakening in our volumes.

2. Visa’s cross-border volume growth remained strong in 2026 Q2 (FY2026 Q3) at 12%, up from 11% in 2026 Q1

Q3 total cross-border volume grew 12% year-over-year, up more than a point from Q2. Cross-border e-commerce volume was up 16%, three points above Q2, primarily driven by retail, including the timing of promotional shopping events. Travel-related cross-border volume was up 10%, consistent with Q2. While the conflict continued to be an offsetting factor, commercial and U.S. inbound continued to improve, and in June, the FIFA World Cup boosted inbound North America and Latin America volume.

3. Payments volume on Visa’s network continues to grow in July 2026, with US payments volume up 9%, cross-border volume up 14%, e-commerce volume up 18%, and processed transactions up 9%

Now, let’s look at drivers through July 21st, with volume growth in constant dollars. U.S. payments volume was up 9%, with both credit and debit up 9% year-over-year. A step down from June, primarily due to retail, including the timing of promotional shopping events, a lack of a day’s mix benefit that helped June, and the change in the cost of fuel. For cross-border volume excluding transactions within Europe, total volume grew 14% year-over-year, with e-commerce up 18% and travel up 12%. Processed transactions grew 9% year-over-year.

---

Disclaimer: The Good Investors is the personal investing blog of two simple guys who are passionate about educating Singaporeans about stock market investing. By using this Site, you specifically agree that none of the information provided constitutes financial, investment, or other professional advice. It is only intended to provide education. Speak with a professional before making important decisions about your money, your professional life, or even your personal life. I currently have a vested interest in Mastercard and Visa. Holdings are subject to change at any time.