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. Both are provided by OpenAI and are software products that use AI to generate art and writing, respectively (and often at astounding quality). 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 2025 – 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

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

Alphabet (NASDAQ: GOOG)

Alphabet’s 1st-party AI models, including Gemini, now process 7 billion tokens per minute from direct APIs; the Gemini App now has 650 million monthly active users; queries on Gemini App has 3x-ed from 2025 Q2; management sees Alphabet’s AI models as world-leading; 230 million videos have been created with Veo3; 13 million developers have built with Alphabet’s generative AI models; management will release Gemini 3 in 2025 Q4; the number of tokens per month processed by Alphabet has increased from 980 trillion in May 2025 to 1.3 quadrillion, up 20x from a year ago; Alphabet is applying Gemini internally and this has increased the productivity of its sales team by over 10%, leading to hundreds of millions in incremental revenue; Alphabet’s customer support division has used Gemini to manage over 40 million customer sessions year-to-date; management thinks the pace of frontier model development is still phenomenal 

Our first-party models, like Gemini, now process 7 billion tokens per minute via direct API used by our customers. The Gemini app now has over 650 million monthly active users, and queries increased by 3x from Q2…

…Our models are world-leading. GEMINI 2.5 Pro, Veo, Genie 3 under viral sensation Nano Banana are among the very best in class. Over 230 million videos have been generated with Veo 3, and more than 13 million developers have built with our generative models. We are looking forward to the release of Gemini 3 later this year…

…In July, we announced that we processed 980 trillion monthly tokens across all our surfaces. We are now processing over 1.3 quarterly and monthly tokens, more than 20x growth in a year, phenomenal…

…We’re also applying Gemini internally to help us serve customers with increased speed, intelligence and efficiency. Our sales teams use Gemini enriched with ads knowledge to streamline customer interactions. This increased productivity by over 10% led to hundreds of millions in incremental revenue and frees up sellers to engage with more customers at a deeper, more strategic level. In our customer support division, Gemini-powered solutions have managed over 40 million customer sessions so far this year and resolved hundreds of thousands of customer inquiries, and we’re just getting started…

…On the on the pace of frontier model research and development. Look, I think 2 things are both simultaneously true. I’m incredibly impressed by the pace at which the teams are executing and the pace at which we are improving these models. But it also is true at the same time that each of the prior model you’re trying to get better over is now getting more and more capable. So I think both the pace is increasing, but sometimes we are taking the time to put out a notably improved model, so I think — and that may take slightly longer. But I do think the underlying pace is phenomenal to see.

Google Cloud saw accelerating growth in 2025 Q3 with AI as a key driver; Google Cloud backlog grew 46% sequentially to $155 billion in 2025 Q3 (was $106 billion in 2025 Q2); Google Cloud is singing new customers faster, with 34% year-on-year increase in new GCP (Google Cloud Platform) customers in 2025 Q3; Google Cloud signed more deals over $1 billion in 2025 9M than in 2023 and 2024 combined; more than 70% of existing Google Cloud customers use Alphabet’s AI products; Google Cloud has 13 product lines that have annual run rate of more than $1 billion each; management thinks Google Cloud offers the widest array of chips, and 9 of the top 10 AI labs are on Google Cloud; revenue from products built on Alphabet’s generative AI models in 2025 Q3 was up more than 200% year-on-year; nearly 150 Google Cloud customers have each processed 1 trillion tokens in the last 12 months; WPP is using Alphabet’s AI models to improve efficiency by up to 70% when creating advertising campaigns; Swarovski is using Alphabet’s AI models to raise e-mail open rates by 17% and accelerate campaign localization by 10x; management recently launched Gemini Enterprise and it is seeing strong adoption of agents; the packaged enterprise agents by Gemini Enterprise have already exceed 2 million subscribers aross 700 companies

Cloud had another great quarter of accelerating growth with AI revenue as a key driver. Cloud backlog grew 46% quarter-over-quarter to $155 billion…

…Next, Google Cloud. Our complete enterprise AI product portfolio is accelerating growth in revenue, operating margins and backlog. In Q3, customer demand strengthened in 3 ways. One, we are signing new customers faster. The number of new GCP customers increased by nearly 34% year-over-year. Two, we are signing larger deals. We have signed more deals over $1 billion through Q3 this year than we did in the previous 2 years combined. Third, we are deepening our relationships. Over 70% of existing Google Cloud customers use our AI products, including Banco BV, Best Buy and FairPrice Group…

…Today, 13 product lines are each at an annual run rate over $1 billion…

…We have a decade of experience building AI accelerators and today, offer the widest array of chips. This leadership is winning customers like HCA Healthcare, LG AI Research and Macquarie Bank, and it’s why 9 of the top 10 AI labs choose Google Cloud…

…In Q3, revenue from products built on our generative AI models grew more than 200% year-over-year. Over the past 12 months, nearly 150 Google Cloud customers each processed approximately 1 trillion tokens with our models for a wide range of applications. For example, WPP is creating campaigns with up to 70% efficiency gains. Swarovski has increased e-mail open rates by 17% and accelerated campaign localization by 10x…

…Earlier this month, we launched Gemini Enterprise, the new front door for AI in the workplace, and we are seeing strong adoption for agents built on this platform. Our packaged enterprise agents in Gemini Enterprise are optimized for a variety of domains, are highly differentiated and offer significant out-of-box value to customers. We have already crossed 2 million subscribers across 700 companies.

Alphabet has a full-stack approach to AI, spanning infrastructure, research, products, and platform; management continues to see Alphabet’s AI infrastructure as a key differentiator; Alphabet is the only company scaling both NVIDIA’s GPUs as well as its own TPUs; Alphabet is now shipping the new A4X Max instances powered by NVIDIA GB300; the 7th-generation of Alphabet’s TPU will be available soon; management is seeing tremendous demand for TPUs; AI startup Anthropic recently announced that it would access up to 1 million TPUs

Our full stack approach spans AI infrastructure, world-class research including models and tooling, and our products and platforms that bring AI to people everywhere…

…Our extensive and reliable infrastructure, which powers all of Google’s products is the foundation of our stack and a key differentiator. We are scaling the most advanced chips in our data centers, including GPUs from our partner, NVIDIA, as well as our own purposeful TPUs. And we are the only company providing a wide range of both. As we announced yesterday at NVIDIA GTC, we are now shipping the new A4X Max instances powered by NVIDIA GB300 to our cloud customers. We are investing in TPU capacity to meet the tremendous demand we are seeing from customers and partners, and we are excited that Anthropic recently shared plans to access up to 1 million TPUs.

Alphabet’s management sees AI expanding Google Search; the growth in overall queries and commercial queries seen in 2025 Q2 accelerated in 2025 Q3, driven by AI Overviews and AI Mode; the acceleration of growth from AI Overviews in Google Search in 2025 Q3 was more pronounced with younger people; AI Mode has seen strong and consistent week-over-week growth in usage since launch in the USA and queries doubled sequentially; AI Mode has been rolled out globally in 40 languages; AI Mode now has 75 million daily active users; AI Mode is driving incremental total query growth for Google Search, including commercial queries; Google Search users can now shop conversationally in AI Mode; all US users of Google Search now have access to try-on capabilities for clothing items; management sees agentic experiences as additive to the way Google Search users seek information; management is working on agentic experiences across key verticals and they think it’s important that Alphabet also creates value for its partners when building these experiences; Alphabet has introduced agentic checkout and partnerships for agentic commerce; AI Overviews now has 2 billion users; AI Overviews continue to monetise at a similar rate as traditional Google Search, but management sees the opportunity for the monetisation to improve; Google Search’s paid clicks and CPCs were both up 7% year-on-year in 2025 Q3; management sees the opportunity in AI Mode to take queries that are not fully commercial and yet still serve attractive advertising offerings

AI is driving an expansionary moment for Search. As people learn what they can do with our new AI experiences, they’re increasingly coming back to Search more. Search and its AI experiences are built to highlight the web, sending billions of clicks to sites every day. During the Q2 call, we shared that overall queries and commercial queries continue to grow year-over-year. This growth rate increased in Q3, largely driven by our AI investments in Search, most notably AI Overviews and AI Mode…

…AI Overviews drive meaningful query growth. This effect was even stronger in Q3 as users continue to learn that Google can answer more of their questions, and it’s particularly encouraging to see the effect was more pronounced with younger people.

We’re also seeing that AI Mode is resonating well with users. In the U.S., we have seen strong and consistent week-over-week growth in usage since launch and queries doubled over the quarter. Over the last quarter, we rolled out AI Mode globally across 40 languages in record time. It now has over 75 million daily active users, and we shipped over 100 improvements to the product in Q3, an incredibly fast pace. Most importantly, AI Mode is already driving incremental total query growth for Search…

…Our investments in new AI experiences, such as AI Overviews and AI Mode, continued to drive growth in overall queries, including commercial queries, creating more opportunities for monetization. These AI experiences are enhancing how people connect with businesses and shop on Search. We recently added shopping capabilities in AI Mode, which now help people shop conversationally in Search, and we expanded try-on capabilities to more clothing items, now available to anyone in the U.S…

…This is all early, but we see agentic experiences really as additive to the way people seek information. It helps us answer people’s tough questions. It helps us — it helps people get stuff done, and it helps businesses in the process…

…We’re working on multiple agentic experiences across key verticals such as travel, commerce, shopping and so on, and we’re paying a lot of attention to creating a seamless user experience but also to the fact that we need to integrate different partner ecosystems in a way that it creates value for them…

…At I/O, we also introduced new agentic checkout, which will let shoppers use like agentic AI to buy products from merchant sites and so on. We have a partnership with PayPal to help merchants build agentic commerce experiences. We have a new open protocols for agent-to-agent transactions and so on and so on…

…AI Overviews is scaling up and working for our entire user base. We’re now scaled to over 2 billion users here, and we’re continuing to expand ads in AI Overviews in English to more countries, across desktop, mobile and so on. And as I’ve shared before, for AI Overviews, even at our current baseline of ads below and within the AI’s response, overall, we see the monetization at approximately the same rate…

…We’re excited about the opportunity of richer experiences in AI Mode and AI Overviews to basically open up then the opportunity for also much richer placements…

…As you will see in the 10-Q, paid clicks were up 7% year-on-year and CPCs were up 7% year-on-year…

…There is the question of whether queries actually increase with AI Mode, and Sundar actually talked about it and mentioned the opportunity that he sees here. So I think it’s important to separate those 2 things. And I personally also see this, what I just said in my last remarks, that I think, over time, there’s an opportunity to actually take, let’s say, queries that are not fully commercial but could have an adjacent commercial relationship to basically expand this into more attractive ads offerings without — while really creating a really interesting user experience at the same time.

Alphabet’s management recently rolled out AI features that help Youtube content creators streamline their entire content creation workflow; Youtube can now automatically products in content creators’ videos to make them more shoppable; Alphabet’s recommendation systems are driving watch time growth in Youtube; the use of Gemini in Youtube is driving improvement in content discovery; management is excited about the revenue growth powered by Demand Gen in Youtube’s direct response advertising business; Alphabet has improved Demand Gen’s performance on Youtube where it can now increase conversion value by more than 40%; Demand Gen is helping Youtube further monetise shopping-related categories; more advertisers are adopting interactive direct response ads on Youtube in the living room, with an annual revenue run rate exceeding $1 billion globally; management has introduced Veo 3 integration and speech to song for content creators in Youtube; Youtube Shorts has lower revenue-share than traditional Youtube

At our Made on YouTube event, we rolled out a number of AI-powered features that are helping create a supercharged creation and build their businesses. AI is now streamlining the entire content creation workflow from generated video tools and more efficient editing to AI-powered insights that help creators optimize their channels. We are also using AI to expand monetization, automatically identifying products to make their videos more shoppable…

…Our recommendation systems are driving robust watch time growth in our key monetization areas like Shorts and Living Room. As we leverage Gemini models, we’re seeing further discovery improvement…

…On direct response, we’re excited about the growth in revenue we’re seeing, especially from small and medium advertisers adopting Demand Gen. We also improved performance on Demand Gen with over 100 launches helping to increase conversion value by more than 40% for advertisers using target-based bidding on YouTube. The retail vertical continues to lead our growth on YouTube with Demand Gen helping us further monetize shopping-related categories.

Looking at the living room, our long-term bet, more advertisers are adopting interactive direct response ads, leading to an annual revenue run rate exceeding $1 billion globally for this format…

…We continue to invest in AI-powered features that are helping creators supercharge creation and build their businesses. with Veo 3 integration and speech to song, creators go from idea to iteration quicker, and new channel insights help them better understand performance…

…Shorts, which has a lower revenue share than in stream that helps to improve some of our gross margins.

Alphabet’s management intends to launch Waymo in London and Tokyo in 2026; Waymo has expanded operations in a number of US cities, and testing in New York City continues to scale; Waymo now has the option for enterprises to offer Waymo as a work-travel option; management launched Waymo teens accounts in Phoenix recently and usage is growing steadily; management thinks there’s a real opportunity to infuse Gemini into Waymo to improve the in-vehicle experience for users in 2026

Next year Waymo aims to open service in London, and they are working to bring service to Tokyo. They’ve also announced expansions to Dallas, Nashville, Denver and Seattle and secured permission to operate fully autonomously at San Jose and San Francisco Airports. Autonomous testing continues to scale in New York City. The new Waymo for Business allows enterprises to offer Waymo as a work travel option. And we launched Waymo teens accounts in Phoenix this summer. We are pleased to see usage steadily increase with positive feedback from teens and their parents alike…

…[Question] How far are we from an integration of Waymo into more of the core Gemini capabilities and the users on the platform taking your user data of where I’m going, what hotel I’m staying at, what airport I’m staying at and having integrated that into Waymo?

[Answer] Waymo clearly is scaling up, particularly in 2026. And I think the possibility, as you said, of Gemini, particularly with the multimodal experience as well as services like YouTube, I think there’s a real opportunity to make the in-car experience dramatically better. Definitely something we are excited about, and you’ll see newer experiences in 2026 for sure.

Alphabet’s management recently rolled out AI Max in Search for businesses and it can understand and predict consumer intent in Google Search;  AI Max is already used by hundreds of thousands of advertisers and is Alphabet’s fastest-growing AI-powered Search ads product; AI Max unlocked billions of net new queries in 2025 Q3; AI Max helps advertisers discover new customers at the exact moment they need their product or service; Kayak used AI Max in Search and grew conversion value by 12%

Businesses can now tap into our most powerful AI search experiences. Using our most advanced AI models, we can understand and predict intent like never before, unlocking entirely new commercial pathways to provide valuable new consumer connections and helping us monetize even more efficiently. Rolled out globally in September, AI Max in Search is already used by hundreds of thousands of advertisers, currently making it the fastest-growing AI-powered search ads product. In Q3 alone, AI Max unlocked billions of net new queries. By delivering the most relevant ad across surfaces and matching advertisers against additional queries they weren’t reaching before, AI Max helps advertisers discover new customers at the exact moment they need their product or service. Kayak, for example, look to grow conversions while staying within their ROAS goals. After turning on AI Max and Search, they grew their conversion value by 12% in early tests.

Alphabet’s management notes that GCP is seeing strong demand for enterprise AI infrastructure and enterprise AI solutions; management notes that GCP will be in a tight demand/supply situation going into 2026; management now expects capex of $91 billion to $93 billion in 2025 (up 66% from $55.4 billion in 2024 and 2024’s capex was up 69% from 2023), up from previous guidance of $85 billion; management expects capex to increase significantly in 2026; management expects the growth rate in depreciation to accelerate in 2025 Q4; when management makes capex decisions, they go through a rigorous process of assessing the return on the investment

In Cloud, demand for our products remains high as evidenced by the accelerating revenue growth and the $49 billion sequential increase in Cloud backlog in Q3. In GCP, we see strong demand for enterprise AI infrastructure, including TPUs and GPUs, enterprise AI solutions driven by demand for Gemini 2.5 and our other AI models, and core GCP infrastructure and other services such as cybersecurity and data analytics. As I’ve mentioned on previous earnings calls, while we have been working hard to increase capacity and have improved the pace of server deployments and data center construction, we still expect to remain in a tight demand-supply environment in Q4 and 2026.

Moving to investments. We’re continuing to invest aggressively due to the demand we’re experiencing from Cloud customers as well as the growth opportunities we see across the company. We now expect CapEx to be in the range of $91 billion to $93 billion in 2025, up from our previous estimate of $85 billion, keeping in mind that the timing of cash payments can cause variability in the reported CapEx number. Looking out to 2026, we expect a significant increase in CapEx, and we’ll provide more detail on our fourth quarter earnings call.

In terms of expenses, first, as I’ve mentioned on the previous earnings calls, the significant increase in our investments in technical infrastructure will continue to put pressure on the P&L in the form of higher depreciation expenses and related data center operations costs such as energy. In the third quarter, depreciation increased $1.6 billion year-over-year to $5.6 billion, reflecting a growth rate of 41%. Given the overall increase in CapEx investments, we expect the growth rate in depreciation to accelerate slightly in Q4. Second, we expect sales and marketing expenses to be more heavily weighted to the end of the year in part to support product launches and the holiday season…

…When we make a decision on investment in the long term, we go through a very rigorous process of assessing what the return could be and over what time frame we will see that return to give us the high level of confidence to then invest and make those investments for the long term.

Nearly half of all code in Alphabet is now generated by AI

The percent of code, now nearly half of all code generated by AI, that’s a way for us to leverage AI to drive further productivity across the business.

Amazon (NASDAQ: AMZN)

AWS grew 20.2% year-on-year in 2025 Q3, and is now growing at a pace last seen in 2022; AWS’s run rate has reached $132 billion (was $123 billion in 2025 Q2), and management thinks 20% growth off such a huge base is more impressive than what competitors have achieved (faster growth off a much smaller base); AWS’s backlog is $200 billion in 2025 Q3 (was $195 billion in 2025 Q2, up 25% year-on-year) and is higher now given unannounced new and large deals in October 2025; AWS has been a Gartner Magic Quadrant leader for 15 consecutive years; management sees AWS continuing to be the destination for most big enterprises and governments’ cloud migrations; AWS is where most companies’ data and workloads reside, and why most companies want to run AI in AWS; AWS operating income in 2025 Q3 was $11.4 billion, reflecting 34.6% operating margin (was 32.9% in 2025 Q2 and 38.1% in 2024 Q3); the AI–portion of AWS’s growth in 2025 Q3 come from both training and inference; a broad-base of AWS’s AI products also contributed to AWS’s AI-growth; cloud migrations by enterprises was also a strong contributor to AWS’s growth in 2025 Q3; management thinks AWS can continue growing at a similar clip as in 2025 Q3 for a while

AWS is growing at a pace we haven’t seen since 2022, reaccelerating to 20.2% year-over-year, our largest growth rate in 11 quarters…

…It’s very different having 20% year-over-year growth on a $132 billion annualized run rate and to have a higher percentage growth rate on a meaningfully smaller annual revenue, which is the case with our competitors. 

Backlog grew to $200 billion by Q3 quarter end and doesn’t include several unannounced new deals in October, which together are more than our total deal volume for all of Q3…

…Gartner has named AWS leader in its strategic cloud platform services Magic Quadrant for 15 consecutive years…

…Because of its advantaged capabilities, security, operational performance and customer focus, AWS continues to earn most of the big enterprise and government transformations to the cloud. As a result, AWS is where the preponderance of companies’ data and workloads reside and part of why most companies want to run AI in AWS…

…Moving next to our AWS segment. Revenue was $33 billion, up 20.2% year-over-year. This is an acceleration of 270 basis points compared to last quarter, driven by strong growth across both our AI and core services and more capacity, which has come online to support customer demand. AWS revenue increased $2.1 billion quarter-over-quarter and now has an annualized revenue run rate of $132 billion. AWS operating income was $11.4 billion, and reflects our continued growth, coupled with our focus on driving efficiencies across the business…

…We see the growth in both our AI area, where we see it in inference. We see it in training. We see it in the use of our Trainium custom silicon. Bedrock continues to grow really quickly. SageMaker continues to grow quickly…

…I think the other place we see a lot of growth in AWS also is just the number of enterprises who are — who have gotten back to moving from on-premises infrastructure to the cloud. And we continue to earn the lion’s share of those transformations. And I look at the momentum we have right now, and I believe that we can continue to grow at a clip like this for a while.

Amazon’s management thinks a lot of the value companies will derive from AI will come from agents and AWS is investing heavily in agents; management thinks companies will both create their own agents and use 3rd-party agents; management has launched Strands in AWS to make it easier for companies to build their own agents; management has launched AgentCore in AWS for companies who have built agents to deploy them in a secure and scalable way; Ericsson, Sony and Cohere Health are all users of AgentCore; Cohere Health is using AgentCore to deploy agents that reduces medical review times by up to 30% to 40%; AgentCore’s SDK (software development kit) has been downloaded over 1 million times; AWS has the coding agent Kiro, which attracted more than 100,000 developers in its first days of launch and that number has since doubled; AWS’s migration agent, Transform, has saved customers 700,000 hours of manual effort in 2025 9M; Thomson Reuters used Transform to transform 1.5 million lines of code per month to complete tasks faster than with other migration tools; customers have already used Transform to analyse 1 billion lines of mainframe code; AWS’s business agent, Quick Suite, has delivered 80% time savings and 90% cost savings to users; AWS’s contact center agent, Amazon Connect, is at a $1 billion annualised revenue run rate and has handled 12 billion minutes of customer interactions in the last year; customers of Amazon Connect include Capital One, Toyota, American Airlines and Ryanair

A lot of the future value companies will get from AI will be in the form of agents. AWS is heavily investing in this area and well positioned to be a leader.

Companies will both create their own agents and use agents from other companies. For those building their own, it’s been harder to build than it should be. It’s why we launched Strands to make it much easier to create agents from any foundation model that builders desire. For companies who successfully built agents, they’ve hesitated putting them into production because they lack secure, scalable runtime services or memory or observability, built specifically for agents. It’s why we launched AgentCore, a set of infrastructure building blocks that allow builders to deploy secure, scalable agents. Ericsson used AgentCore to deliver AI agents across their workforce, Sony used it to build a agentic AI platform with enterprise-level security, observability and scalability. And Cohere Health is using AgentCore to deploy agents that will reduce medical review times by up to 30% to 40%. AgentCore’s SDK has already been downloaded over 1 million times, and our builders are excited about it…

…For coding, we’ve recently opened up our agentic coding IDE called Kiro. More than 100,000 developers jumped into Kiro in just the first few days of preview and that number has more than doubled since. It’s processed trillions of tokens thus far, weekly actives are growing fast, and developers love its unique spec and tool calling capabilities.

For migration and transformation, we offer an agent called Transform. Year-to-date, customers have already used it to save 700,000 hours of manual effort. The equivalent of 335 developer years of work. For example, Thomson Reuters used it to transform 1.5 million lines of code per month, moving from Windows to open source alternatives and completing tasks or at times faster than with other migration tools. Customers have also already used Transform to analyze nearly 1 billion lines of mainframe code as they move mainframe applications to the cloud.

For business customers, we’ve recently launched Quick Suite to bring a consumer AI-like experience to work, making it easy to find insights, conduct deep research, automate tasks, visualize data and take actions. We’ve already seen users churn months long projects into days, get 80% plus time savings on complex tasks and realize 90% plus cost savings…

…For contact centers, we offer Amazon Connect which creates a more personalized and efficient experience for contact center agents, managers and their customers. Connect has recently crested $1 billion annualized revenue run rate with 12 billion minutes of customer interactions being handled by AI in the last year and is being used by large enterprises like Capital One, Toyota, American Airlines and Ryanair.

AWS has added 3.8 gigawatts of capacity in the last 12 months, more than any competitor; AWS now has double the capacity it had in 2022, and is on track to doubling capacity again by 2027; management expects to add 1 gigawatt of capacity in 2025 Q4; management is growing AWS capacity very aggressively because they see the demand; as soon as capacity is added to AWS, it is monetised

We’ve been focused on accelerating capacity the last several months, adding more than 3.8 gigawatts of power in the past 12 months, more than any other cloud provider…

…We’re now double the power capacity that AWS was in 2022, and we’re on track to double again by 2027. In the last quarter of this year alone, we expect to add at least another 1 gigawatt of power. This capacity consists of power, data center and chips, primarily our custom silicon, Tranium and NVIDIA…

…You’re going to see us continue to be very aggressive in investing in capacity because we see the demand. As fast as we’re adding capacity right now, we’re monetizing it. It’s still quite early and represents an unusual opportunity for customers in AWS.

Project Rainier, a massive AWS AI compute cluster consisting of 500,000 of AWS’s in-house Trainium 2 chips, is now online; AI startup Anthropic is using Project Rainier to build and deploy the next generation of its leading AI model; management expects Anthropic to use up to 1 million Trainium 2 chips by end-2025; Trainium 2 is currently fully subscribed, and is a multi-billion dollar business that grew 150% sequentially in 2025 Q3; Trainium is currently used by only a small number of very large customers, but management expects more customers to use Trainium once Trainium 3 comes online; the token usage of Amazon Bedrock, AWS’s fully-managed service for companies to leverage frontier models to build generative AI apps, is mostly on Trainium; even as AWS scales Trainium, management continues to order significant amounts of chips from NVIDIA, AMD, and Intel; management sees Trainium as 30%-40% more price-performant than other options; management thinks that as companies start to scale production AI workloads, they will care a lot about price performance, and this will lead to strong demand for Trainium; Trainium 3 should preview at end-2025, with full volume coming in early-2026; there are many large and medium-sized customers who are interested in Trainium 3; management thinks that AWS will always have multiple chip options for customers and that has been true for every major technology building block; management thinks the chip team behind Trainium, Annapurna, is really strong; management expects Trainium 3 to be 40% better than Trainium 2; it was not easy to build Project Rainier to be able to scale from 500,000 chips to 1 million; Project Rainier is specific for Anthropic

We’ve recently brought Project Rainier online, our massive AI compute cluster spanning multiple U.S. data centers and containing nearly 500,000 of our Trainium2 chips. Anthoropic is using it now to build and deploy its industry-leading AI model Claude, which we expect to be on more than 1 million Trainium2 chips by year-end. Trainium2 continues to see strong adoption, is fully subscribed is now a multibillion-dollar business that grew 150% quarter-over-quarter.

Today, Trainium is being used by a small number of very large customers but we expect to accommodate more customers starting with Trainium3.

We’re building Bedrock to be the biggest inference engine in the world and in the long run, believe Bedrock could be as big a business for AWS as EC2, and the majority of token usage in Amazon Bedrock is already running on Trainium.

We’re also continuing to work closely with chip partners like NVIDIA, with whom we continue to order very significant amounts as well as with AMD and Intel. These are very important partners with whom we expect to keep growing our relationships over time…

…Because Trainium is 30% to 40% more price performant than other options out there, and because as customers, as they start to contemplate broader scale of their production workloads, moving to being AI-focused and using inference, they badly care about price performance. And so we have a lot of demand for Trainium. Trainium3 should preview at the end of this year with much fuller volumes coming in the beginning of ’26. We have a lot of customers, both very large, and I’ll call it, medium-sized who’re quite interested in Trainium3…

…We’re always going to have multiple chip options for our customers. It’s been true in every major technology building block or component that we’ve had in AWS. Really in the history of AWS, it’s never just one player that over a long period of time has the entire market segment and then it can satisfy everybody’s needs on every dimension…

…We’re different from most technology companies in that we have our own very strong chip team, and this is our Annapurna team. And you saw it first on the CPU side with what we built with Graviton which is about 40% better price performance than the other x86 processors, and you’re seeing it again on the custom silicon on the AI side with Trainium, which is about the same amount of price performance benefit for customers relative to other GPU options…

…As we think about Trainium3, I expect Trainium3 will be about 40% better than Trainium2 and Trainium2 is already very advantaged on price performance…

…It’s not simple to be able to build a cluster that has 500,000 plus chips going to 1 million. That’s an infrastructure feet that’s hard to do at scale…

…Project Rainier is something that is specific for Anthropic.

Rufus, Amazon’s AI shopping assistant, has 250 million active customers in 2025 9M; Rufus monthly users are up 140% year-on-year and interactions are up 210%; customers using Rufus are 60% more likely to complete a purchase; Rufus is pacing towards $10 billion in incremental annualized sales; management is very excited about agentic commerce; management thinks agentic commerce will be very useful for consumers who don’t know what they want to buy; management sees Rufus as a part of Amazon’s agentic commerce efforts; Amazon has a Buy For Me agentic feature where products will be surfaced for consumers, even items that Amazon does not stock but that other merchants have; management is also looking to partner with 3rd-party agents; search engines are a very small part of Amazon’s traffic today, and 3rd-party agents are an even smaller part; management thinks the current agentic commerce experience is not good for consumers; management thinks agentic commerce will expand the amount of online shopping and this bodes well for Amazon

Rufus, our AI-powered shopping assistant has had 250 million active customers this year with monthly users up 140% year-over-year, interactions up 210% year-over-year and customers using Rufus during a shopping trip being 60% more likely to complete a purchase. Rufus is on track to deliver over $10 billion in incremental annualized sales…

…As a business, we’re very excited about in the long term the prospect of agentic commerce. And it has a chance to be good for customers, it has a chance to be really good for e-commerce…

…If you know what you want to buy, there are a few experiences that are better than coming to Amazon. But if you don’t know what you want, it’s — physical store with a physical salesperson still has some advantages. Obviously, lots of people do it on Amazon all the time. But you very often want to ask questions and help — get help narrowing what you’re going to look for. And as you keep asking new questions, having a whole bunch of different options presented to you. And I think AI and agentic commerce are going to change the experience online where that experience where you’re narrowing what you want when you don’t know is going to get better online than it even is in physical environments…

…We obviously have our own efforts here in agentic commerce. We have Rufus, which I talked about in my opening comments, which is continuing to get better and better and used more broadly. And we have features like Buy for Me where we will surface on Amazon, even items that we don’t stock that other merchants have. And then if customers want us to go and buy it for them on those merchants’ websites, we will do that. And both of those have been successful for us. But we’re also having conversations with and expect over time to partner with third-party agents…

…Today, search engines are a very small part of our referral traffic and third-party agents are a very small subset of that…

…We have to find a way, though, that makes the customer experience good. Right now, I would say the customer experience is not. There’s no personalization, there’s no shopping history, the delivery estimates are frequently wrong, the prices are often wrong…

…I do think that the exciting part of this and the promise is that AI and agentic commerce solutions are going to expand the amount of shopping that happens online. And I think that’s really good for customers, and I think it’s really good for Amazon because at the end of the day, you’re going to buy from the outfit that allows you to have the broadest selection, great value and continues to deliver for you very quickly and reliably. And I think that bodes well for us.

Customers are talking to Alexa+ 2x more compared to the classic Alexa experience; customers are talking to Alexa+ for longer compared to classic Alexa; compared to classic Alexa, customers are using Alexa+ in Fire TV 2.5x more, to discover audio content 4x more, to engage with photos 4x more, and to complete shopping conversations (that end with a purchase) 4x more

We continue to be energized by the response to Alexa+. Compared to what we call the classic Alexa experience, Alexa+ customers are talking to Alexa 2x more. Those interactions are much longer, and they’re covering a broader range of topics. So using Alexa+ in Fire TV at 2.5x the rate of classic, using natural conversation to discover audio content 4x more, engaging with photos 4x more, and customers are completing 4x more shopping conversations that end in a purchase.

More than 1.3 million sellers have used Amazon’s generative AI capabilities to speed up the launch fo high-quality listings; 3rd-party seller unit mix was 62% in 2025 Q3 (62% in 2025 Q2)

Our millions of global third-party sellers continue to be important contributors to our vast selection, which helps customers find the items they need at competitive prices. We’re committed to building innovative services and features for our sellers, including our ongoing advancements in generative AI. Today, more than 1.3 million sellers have used our generative AI capabilities to more quickly launch high-quality listings. Better listings translate into better traction with customers. And in Q3, worldwide third-party seller unit mix was 62%, up 200 basis points from Q3 of last year.

The majority of Amazon’s capital expenditure (capex) in 2025 Q3 was for AWS’s technology infrastructure, including the Trainium chips

Now turning to our cash CapEx, which was $34.2 billion in Q3. We’ve now spent $89.9 billion so far this year. This primarily relates to AWS as we invest to support demand for our AI and core services and in custom silicon, like Trainium as well as tech infrastructure to support our North America and international segments. We’ll continue to make significant investments, especially in AI, as we believe it to be a massive opportunity with the potential for strong returns on invested capital over the long term. Additionally, we continue to invest in our fulfillment and transportation network to support the growth of the business, improve delivery speeds and lower our cost to serve. These investments will support growth for many years to come.

Apple (NASDAQ: AAPL)

Apple’s management sees Apple’s silicon as the heart of the company’s efforts in AI; management thinks the A19 Pro chip and M5 chip make Apple products the very best place to experience the power of AI; management has introduced dozens of new features in Apple Intelligence, including Live Translation, Visual Intelligence, Workout Buddy, Clean Up in photos, and more; management is seeing developers build on the foundation models on Apple’s devices; management expects to release the new, more personalised version of Siri next year; Apple is using Private Cloud Compute (PCC) to handle some of Siri’s queries, and the company continues to build that out; there were capital expenditures in FY2025 that were related to the build out of PCC; management intends to continue using internal foundation models together with other LLMs in building the personalised version of Siri

As we continue to expand our investment in AI, we’re bringing intelligence to more of what people already love about our products and services, making every experience even more personal, capable and effortless. At the heart of it all is Apple silicon, and we were thrilled to launch new products powered by the A19 Pro chip and M5. These incredibly advanced chips make Apple products the very best place to experience the power of AI.

With Apple Intelligence, we’ve introduced dozens of new features that are powerful, intuitive, private and deeply integrated into the things people do every day, features like Live Translation, which help users communicate across languages in real time; and Visual Intelligence, which opens new ways to learn about and explore the world. We also introduced Workout Buddy, a new experience that uses AI to provide personalized motivational insights based on a user’s workout data and fitness history. And these joined so many others from Clean Up in photos and new image creation tools to powerful writing tools. We’re also seeing developers take advantage of our own device foundation models to create entirely new experiences for users around the world. We’re also excited for our more personalized Siri. We’re making good progress on it, and as we’ve shared, we expect to release it next year…

…We’re obviously using PCC, our Private Cloud Compute, today for a number of queries for Siri, and we will continue to build it out. In fact, the manufacturing plant that makes the servers used for Apple Intelligence just started manufacturing in Houston a few weeks ago, and we’ve got a ramp plan there for use in our data centers and it’s robust…

…In ’25, we did have CapEx costs associated with building out our Private Cloud Compute environment in our first-party data centers. So you would have seen that in some of the CapEx investment in the year…

…[Question] Good to know that the personalized Siri is making good progress and on track for next year. Will you continue to use a three-pronged approach with your own foundation models and partner with other LLM providers and maybe potential M&A?

[Answer] We’re obviously creating Apple foundation models within Apple. We ship them on device and use them in the Private Cloud Compute as well. And we’ve got several in development. And so we also, from a — continually to surveil the market on M&A and are open to pursuing M&A if we think that it will advance our road map.

The Apple Watch Series 11 has the most comprehensive set of health features yet, and these health features are powered by AI and advanced machine learning; the latest Apple Watch now has hypertension notifications that were developed using large-scale machine learning models; AirPods Pro 3 can pair very well with Live Translation to deliver an incredibly new and exciting experience for users

Apple Watch Series 11 brings our users the most comprehensive set of health features yet. And Apple Watch SE 3 delivers advanced capabilities at an incredible value. AI and advanced machine learning are at the core of powerful health features like heart rate monitoring, fall detection, crash detection and more. With our latest Apple Watch lineup, we were proud to introduce hypertension notifications, developed using large-scale machine learning models. Hypertension is one of the leading risk factors for heart attack and stroke affecting more than 1 billion adults worldwide, and we expect to notify more than 1 million users of this life-threatening condition…

…With Live Translation powered by Apple Intelligence, AirPods deliver an incredibly new and exciting experience for users around the world.

Apple’s management has committed to invest $600 billion over the next 4 years (was $6500 billion in 2025 Q2; Apple has $190 billion in gross profit per year, for perspective) in the USA in areas such as advanced manufacturing, silicon engineering and artificial intelligence; Apple already built a new factory in Houston for advanced AI service

A great example is the work we’re doing in the U.S. where we’re committed to invest $600 billion over the next 4 years with a focus on innovation in strategic areas like advanced manufacturing, silicon engineering and artificial intelligence. These commitments build on our long-standing investments in America while supporting more than 450,000 jobs with thousands of suppliers across all 50 states. We built a new factory in Houston for advanced AI service, for example, which just started shipping its first products off the line, and we’re leading the creation of end-to-end silicon supply chain across the country.

It’s still too early to tell for sure, but management thinks Apple Intelligence has been a driver of demand for Apple devices, and the driving force will become greater over time

[Question] With all the hype now around AI, are you seeing evidence that AI capabilities or features are a material purchase consideration for consumers?

[Answer] I think that there are many factors that influence people’s purchasing considerations and so — and we don’t have a great in-depth survey yet on the current iPhone 17 because it’s very new in the cycle, and we give it some time to formulate. But I would say that Apple Intelligence is a factor, and we’re very bullish on it becoming a greater factor. And so that’s the way that we look at it.

Apple’s management will continue with Apple’s hybrid approach when it comes to data centers, of using its own data centers as well as those of 3rd-parties

[Question] In the wake of nearly every other large tech company massively increasing their CapEx in advance of AI demand and also mentioning that there’s scarce capacity, do you anticipate Apple altering its sort of long-standing hybrid approach to your own and third-party data centers?

[Answer] As we’ve talked about before, we are expecting increases in our CapEx spending related to AI investments. For example, as I mentioned earlier, we did end up having investments this year to build out our Private Cloud Compute environment. And we do believe this hybrid model has served us very well, and we continue to want to leverage it. And so I don’t see us moving away from this hybrid model where we leverage both first-party capacity as well as leverage third-party capacity. We’ll continue to want to build out Private Cloud Compute, as Tim outlined, as we have more usage there over time. But I think, in general, we want to continue to have this hybrid model.

ASML (NASDAQ: ASML)

Some of the recent uncertainties hampering ASML’s business has reduced, driven by AI; management thinks there will be continued investment in advanced Logic and DRAM because of AI; management thinks AI-demand will benefit a larger part of ASML’s customer base than previously expected; management is a little careful with how all the big AI-related announcements can eventually translate into real capacity, but nonetheless, has been preparing for growth; the broadening of ASML’s customer base because of AI is a big positive development for the company

I think we have seen a flow of positive news in the last few months that has helped to reduce some of the uncertainties we discussed last quarter. First, we continue to see strong news about commitment to AI. Which means, we think, investment in advanced Logic and DRAM. Second, and it’s very important for us, it looks like AI is going to benefit a larger part of our customer base. Third, we continue to make very good progress with our litho intensity, especially with EUV that continues to be adopted with DRAM and advanced Logic customers…

…If you look at the sum of the announcement, I would say this creates a pretty positive backlog of opportunity for AI moving forward…

…I think we are a bit careful with how the big announcement can translate into real capacity need on the ground.

I think the one thing I’d still like to stress one more time is we see the broadening of the customer base, I think, a very important news in that matter because whatever you do is the first set of news, I think we can all agree that we need to make sure that the market will not be supply limited. And this has always been a risk with a limited amount of customers supplying AI chips, both in logic and DRAM…

…We have said for a few quarters that we have been preparing for growth. So we were following those dynamic. And I think we know now that EUV most probably will be stronger next year. So we’ve been preparing for that. We have, as you know, also worked on longer-term capacity. So we continue to track basically the market carefully, having in mind that we want to be able to follow the demand. S

ASML recently invested in AI startup Mistral and management sees it as a strategic partnership that will help ASML improve the software in its systems and also speed up product development; ASML invested €1.3 billion in Mistral AI 

We entered into a partnership with Mistral AI. I think Mistral is really recognized on a number of fronts. They’re recognized for their business-to-business approach. They’re also recognized for the quality of their large language model. Particularly when it comes to software coding and software coding development. So, they’re recognized for that. That’s the reason why we entered into the partnership with them. Because many people look at ASML, look at our products and are really looking at hardware. But increasingly I think people appreciate the very significant software content that is within those systems. People really understand that if you get to the level of precision and the level of speed that we have in our scanners, but also quite frankly, what we need in metrology and inspection, it’s pretty clear that the software contingent therein becomes increasingly important.

So, that’s the reason why this is very strategic to us. Why it’s very strategic to improving the performance. Improving the precision and the speed of our tools as we bring them to our customers. So, therefore, this collaboration is truly a strategic choice for us. I would also say that, on top of the significance that it has for our products, AI is also a great way to improve the speed of our product development. To improve the speed of our time-to-market of any product development to our customers. That’s another big area that we’re collaborating with Mistral on.

So, all in all, we believe a very strategic partnership. We also, to underscore that strategic partnership, we were the lead investor for their Series C funding round. By being the lead investor we took approximately an 11% share in Mistral. We also have a seat on their Strategic Committee…

…ASML has invested EUR 1.3 billion in Mistral AI’s Series C funding round as lead investor. 

ASML’s management continues to see strong growth for the semiconductor market in the long-term, driven by AI; management thinks the shift of ASML’s customers towards advanced Logic and Memory chips will drive demand for advanced lithography and higher lithography intensity; management thinks the shift from 6F-square to 4F-square in DRAM will not cause the number of EUV layers to drop; it’s hard to tell how exactly how all the AI-announcements will translate into business for ASML in the next few years

[Question] Can I ask you to remind us of the long-term opportunities for ASML and a little bit the market you see there?

[Answer] We said that most probably AI will drive more advanced applications in semiconductors. So advanced DRAM, advanced Logic. This is happening and this is driving more advanced litho, higher litho intensity. We expect that to continue. As we just discussed, we see that 3D integration will become a new opportunity which we are going to pursue. As Roger explained very nicely, we also see that AI could create a lot of value in our products moving forward. So we continue to see a very strong opportunity on our technology roadmap… 

…[Question] There is a view that when you go into 4 F-squared from 6 F-squared for DRAM architecture, that’s actually negative for EUV, the EUV layer count comes down. Can you just help us understand that?

[Answer] The short answer is no. If we look at the number of EUV layer going from 6 F-squared to 4 F-squared, we do not expect the number of layers to drop. In fact, as 4 F-squared road map continues after the transition, we, in fact, expect the number of EUV layers to continue to grow. And I make that statement after many discussions with our customer. On top of that, what I’d like to add is 4 F-Squared has a bit of a more complex structure. So it’s, in fact, adding overall more litho mask, more advanced litho mask. So there is a benefit also to some extent, to advanced deep UV. So in any case, you still doubt about it. 4 F-squared is in no way a bad news for ASML…

…I think we wish we had a formula to translate all the announcement on what it means exactly for us in the next few years. But I think no one has that.

The entire semiconductor supply chain does not have very clear visibility into AI-driven demand

[Question] It feels like you wake up every day to another massive announcement from somewhere within the AI food chain. And you sort of — you spent a lot of time talking about how that to create a theoretical backlog for you, not yet orders. But I’m just wondering, you are a critical supplier into this market. You have the potential to be a backlog — to be a bottleneck rather for the market. Now of course, you don’t want that to be the case. You’re preparing for growth, et cetera. But do you feel like there’s sufficient understanding through the chain, whether it’s of where you sit or perhaps your customers?

[Answer] I think we wish we had a formula to translate all the announcement on what it means exactly for us in the next few years. But I think no one has that…

…[Question] Do you feel like your customers are giving you that heads up, right? Is there a sufficient acknowledgment through the chain?

[Answer] I think they do their very best. I will say it this way because they have the same challenge as we do.

Cloudflare (NYSE: NET)

A large digital media platform expanded its relationship with Cloudflare because the media company saw Cloudflare as the only company building the essential platform to protect and manage content for the emerging AI-driven web; the media company is looking to ramp up Cloudflare’s Pay Per Crawl service; the media company thinks Pay-Per-Crawl could turn Cloudflare from an expense-line into a revenue-generator

A Global 2000 digital media platform expanded its relationship with Cloudflare, signing a 3-year $22.8 million pool of funds contract for application services and workers. This contract marks the culmination of a powerful comeback story. We actually lost this customer to a competitor in 2016, but the Internet and Cloudflare evolved. We earned their trust back in 2023, starting with our Zero Trust portfolio. During 8 months of testing before signing this deal, our world-class security, unmatched product breadth and powerful Workers platform ran circles around the incumbent. But that’s not the whole story. The decisive factor of the win was AI. This customer looked at the landscape and correctly identified Cloudflare is the only company building the essential platform to protect and manage content for the emerging AI-driven web. This strategic win established us as the customer’s clear forward-looking partner and creates a direct on-ramp for Pay Per Crawl, which could transform Cloudflare from a vendor they pay for services into a powerful revenue generator for their business.

A web infrastructure platform signed a contract with Cloudflare for AI Crawl and Bot Management after seeing a huge surge in visits from AI scrapers and bots, leading to cost inflation with growth in revenue; management thinks the web infrastructure platform could become a Pay Per Crawl customer im the future

A global web infrastructure platform expanded its relationship with Cloudflare, signing a 14-month $1.2 million contract for AI Crawl Control and Bot Management. This customer is experiencing a massive surge in AI scrapers and malicious bots hitting their origin servers, inflating costs without revenue conversion and obscuring visibility into legitimate traffic. They selected Cloudflare for our innovative best-of-class bot blocking capabilities in addition to seamless expedited deployment by our deep platform integration. We’re already exploring a much larger opportunity with this customer for Pay Per Crawl.

Cloudflare’s management sees the company having emerged as a strategic partner to media companies in managing the new business model of the internet in the AI world; management thinks AI is a massive information consumption platform shift that will also change the business model of the internet from the previous long-standing model of (1) create content, (2) generate traffic, and (3) sell products or advertising; management thinks there are many questions that will arise with the new business model of the internet and they do not even know what this new business model will look like, but they think Cloudflare will be an important force in shaping the conversation; 80% of leading AI companies are relying on Cloudflare’s infrastructure; management sees Cloudflare as a thought leader in what the future business model of the Internet looks like; even companies such as the research departments of banks are speaking with Cloudflare to figure out a new business model of the internet in the AI world, and the same goes for brands and small businesses

We talked last quarter about how the rise of AI would impact media companies. Cloudflare has emerged as a strategic partner to these firms as they work through what the new business model of the Internet will be. But it goes beyond just media. Businesses of all shapes will be transformed by the rise of AI. I don’t think people yet appreciate how AI is another massive information consumption platform shift, just as we move from consuming information via a browser on a desktop to social media and then to apps on mobile devices, AI is another information consumption platform shift. It changes where and how we will consume and interact with information.

With the last 3 platform shifts, the business model of the Internet remains the same: create content, generate traffic and then sell things, subscriptions or ads. With AI, for the first time in a long time, the fundamental business model is going to change. Human eyeball traffic is unlikely to be the currency of the Internet’s future. We already can see glimpses of that future. It’s represented in SciFi. When George Jetson asks his helpful robot Rosie for a recipe for cookies, the response isn’t 10 blue links to hunt through. It’s a recipe for cookies. Most of us are increasingly living in some version of that future now with tools like ChatGPT, and it seems inevitable that more and more commerce will be facilitated by AI-powered agents working on our behalf. 

As that happens, new questions will arise. What happens to small businesses? What happens to brands? Brands, of course, are just shortcuts for humans to be able to assess quality and value. What do they mean in the world of agentic commerce? I don’t know what the future business model of the Internet will look like, who the winners and losers will be, but I do believe Cloudflare will help shape it. We estimate 80% of the leading AI companies already rely on us. A huge percentage of the Internet sits behind us. The agents of the future will inherently have to pass through our network and abide by its rules. And as they do, we will help set the protocols, guardrails and business rules for the Agentic Internet of the future…

…One of the things that has really set us apart is — and this is thanks to our over time, just significant investment in public policy and the side of the house that maybe doesn’t always get as much attention. But I think we have been thought leaders in thinking about what does the future business model of the Internet look like…

…At banks, the research departments they’re a little nervous because they’re seeing ticks down in the amount of research that people are paying for because the AI companies are slowing that up. So that’s open conversations with financial services companies. We’re seeing challenges with brands that are worried about what does a brand mean in the future of agentic commerce. We’re seeing challenges from small businesses. And I think one of the things that I am passionate about is how do we make sure that as this new paradigm, as this new platform emerges, how do we make sure that everybody has a fair shot to be able to participate in it.

Cloudflare’s management is seeing companies increasingly adopting Cloudflare’s Workers developer platform for running AI inference, and building AI agents and applications; management has always been investing behind demand for Cloudflare, not ahead of it; Workers is not facing any form of capacity constrain

Our Workers developer platform continues to deliver outsized growth with the world’s most innovative companies increasingly adopting Workers for running AI inference tasks as well as building AI agents and full stack applications…

…[Question] do you think that you’re capacity constrained in Workers?

[Answer] I don’t think we’re capacity constrained because of somewhat the nature of how we’ve architected Cloudflare and the philosophy of how we make CapEx and network investments. We always have tried to invest behind demand, not ahead of demand.

Cloudflare’s management believes Cloudflare can get the utilisation rate of its GPUs up to 70%-80%, given the company’s excellent track record with utilising CPUs; Cloudflare has been able to generate revenue from its hardware deployment even before it starts paying for the equipment

It’s been remarkable to see over the last 15 years, how our team has been able to squeeze as much as possible out of the CPU capacity that we have, where we can run that CPU capacity at 70% to 80% utilization and get more out of every CapEx dollar we spend. But what’s fascinating is we’re sort of speed running the last 15 years now with GPUs, where we’re figuring out how to make GPUs multi-tenant, how to make them load and unload models more quickly and driving the utilization of GPUs up substantially. And so that is still well below what we have with CPUs, but we see no reason that we can’t get GPUs also up to that 70%, 80% utilization…

…The supply chain within Cloudflare is so optimized to a large degree because we use off-the-shelf equipment and parts that we can deploy hardware, especially in Tier 1 cities and generate revenue even before we start to pay for the equipment. So not only do we have the flexibility that Matthew described really well at length, our reaction time to deploy hardware where we need it is really, really fast.

Cloudflare’s management sees the biggest competition for the company from winning inference workloads is the hyperscalers; management thinks Cloudflare can show much better TCO (total cost of ownership) than the hyperscalers when it comes to inference workloads; Cloudflare can become very sticky for inference workloads once customers realise there’s a different way to run these workloads from the traditional way of doing it with the hyperscalers; AI inference is still a tiny portion of Cloudflare’s revenue today, even though management is excited about its potential; management does not see any concentration risk in its AI-native business; management has found that the first product from Cloudflare that AI companies are often interested in is security-related, because the AI companies’ cost-to-serve queries is high, so they want to block out fraudulent queries; of the 80% of leading AI companies that rely on Cloudflare’s infrastructure, many of them are using Cloudflare’s security products; management thinks a particular strength of Cloudflare is being able to bring the inference workloads close to users, resulting in lower latency; management thinks that many inference workloads in the future will be run on the edge (i.e. on-device) and if it can’t be done, then it will be run on the network, which suits Cloudflare’s strength

[Question] On competition for Cloudflare in the enterprise for securing those inference workloads and winning those inference workloads in particular. Matthew, I would love to hear you comment how do you think competition is evolving in the enterprise as you build out some of the breadth and depth of your functionality?

[Answer] I think that the primary competition for inference workloads continues to be the hyperscalers. And it continues to be the model of do you want to do this work yourself and have to optimize yourself or do you want to hand it off to Cloudflare. And I think in the cases where we’re in the conversation, we’re able to show that there’s just a much better TCO, total cost of ownership, a much lower cost, much better performance when we manage that for you. And so there’s kind of a standard way people do things, which is the hyperscaler way. We’re having to teach them that there is a different way that’s out there…

…I think that we are finding, though, that once somebody learns that there’s a better way that Cloudflare is very, very sticky, and we keep those customers over the long term…

…Even though we’re excited about AI and AI inference, it is still a relatively de minimis portion of our overall revenue, growing fast, but not — I don’t see any current concentration risk that’s there. And what we’re seeing is actually sometimes it’s not the inference products that initially get interest from the AI native companies. It’s actually the security products. And the reason why is the cost of AI, every query can be so high that making sure that you don’t have fraudulent queries running through your system is critical in order to make sure that you can continue to operate cost effectively. And so many of the AI companies, we estimate that about 80% of AI companies use us in one way or another. But a lot of the times, that’s using us for actually securing some of our — really our Act 1 products. And then we are working on getting more and more of them to use the inference products as well.

In terms of what we can do that others can’t do, I think you’re absolutely right that being able to be close to users is important for a latency perspective. And that’s — and when you have human computer interaction, especially with something that is seems almost alive when you’re interacting with it. Every millisecond counts because it breaks that illusion if things slow down, especially as you get to things like voice communication and other things that need to have kind of a natural rhythm to them. And so I think we’re well positioned for that…

…It’s clear to me that there is something very, very real here that it is going to be transformative that a lot of inference will run on your handset or your driverless car directly there, but that if it can’t run there, it needs to run somewhere else, the next best place for it to run is in the network. And Cloudflare is the only network that gives you that capability on a global basis today. And I think that, that’s going to continue to allow us to win workloads regardless of what happens to AI generally.

Cloudflare’s management started the NET Dollar project because they think a common currency would be needed in agentic commerce transactions; management thinks NET Dollar fits well with the regulatory regimes of the US and other parts of the world; Cloudflare has other irons in the fire apart from NET Dollar when it comes to facilitating payments in agentic transactions; management believes there will be multiple different ways to pay in agentic transactions, and they want Cloudflare to be in the center of that

So as we have really interacted with AI companies, but also the merchants and media companies and the real long tail of the Internet, much of which sits behind us. What we realized was that as we move into a world of agentic commerce, we’re going to need a currency to pay for the commerce that is done between agents that is really designed specifically for that task. And that’s the spirit with which we started the NET Dollar project…

…I think we’re approaching it in a thoughtful way and are confident that we can execute in a way that is both going to help facilitate agent-to-agent commerce and be something that it fits well within any of the regulatory regimes that we have both in the U.S. and around the rest of the world…

…We want to be the Babel fish of AI, sort of the universal translator, whether you’re using MCP, the Anthropic protocol or Google’s version of it or Microsoft’s version of it, Cloudflare supports all of those. And so I think in addition to the excitement that we’ve seen around NET Dollar, I am equally excited about the partnerships that we’re doing with Coinbase around X402, with Visa, Mastercard, American Express, around how you can create agent-to-agent payments. And I think that Cloudflare is a network, and what you want networks to be able to do is facilitate the ability for connection to happen and do it regardless of what makes sense. So we think there are potentially some advantages to what we’re building with NET Dollar, but we’re not all in on any one of these things…

…We also believe that there are going to be multiple different ways to pay. There are going to be multiple different agentic protocols, and they are going to be hopefully many, many, many AI companies interacting with many media and businesses to create a more frictionless and AI-powered future of commerce. And I think that we see ourselves in the center of that.

Cloudflare’s management is seeing good progress with Pay Per Crawl; media companies have gotten markedly better deals with AI companies with Pay Per Crawl; 

I think you’re asking about the product around us thinking about how do we help media companies figure out a new business model for the future. I think that, yes, I think that’s going just extremely well. Like the number of media companies that are signed up and engaged is powerful. We’re hearing from them about how the deals that they are able to do with AI companies have gotten markedly better, and we are getting a lot of praise for that.

Mastercard (NYSE: MA)

Mastercard is building the foundation for agentic commerce in partnership with key players such as OpenAI and Google; the Mastercard Agent Pay feature enables agents to facilitate transactions over Mastercard’s payment network; Mastercard processed its first agentic payment in 2025; US Bank and Citibank cardholders can now use Agent Pay, with more US issuers able to use Agent Pay in November, followed by a global rollout in early 2026; merchants can use Agent Pay without any significant need for integration; Mastercard has a partnership with Walmart for Agent Pay; agents can use Mastercard’s inside tokens to deliver personalised agentic commerce experiences to consumers; management thinks the runway for agentic commerce is long

With our global acceptance reach, trusted brand and services capabilities, we’re instrumental in creating the foundation for agentic commerce. We’re now working with key players such as OpenAI on their agentic commerce protocol and with Google and Cloudflare to set industry standards, all to drive safety and security.

To Mastercard Agent Pay, we’re enabling agents to facilitate transaction over a Mastercard’s payment network in a secure and scalable way. You already have agents registered and have tools in place for easy onboarding as others are ready. Our first agentic transaction took place on our network this quarter at a pivotal moment in payments, and that’s just the start. U.S. Bank and Citibank cardholders can now use Agent Pay. The rest of our U.S. issuers will be enabled in November with a global rollout to follow early next year.

And the beauty of it all, we’ve made it easy for merchants across the globe to benefit on day 1 with the same trust and security they are used to do from us today. Our acceptance framework enables any Mastercard merchant to participate without significant development or integration, a no-code approach…

…We have strong partnerships with the players I just mentioned and many more, including Walmart, to accelerate the adoption of agentic commerce using cards through Mastercard Agent Pay…

…Agents through Mastercard’s inside tokens can make agentic commerce even more personalized. By harnessing our proprietary data, we will be able to provide agents with predictive insights to help drive smarter decisions and recommendations.

The shift we’re seeing in commerce is creating further opportunity for our capabilities, more consulting, more loyalty, more security and so on. The runway for agentic focused services in consumer and business use cases is long, and we’re well positioned to capture this opportunity.

Mastercard’s management is seeing consumer search behaviour change because of AI chatbots; management thinks agentic commerce is a significant paradigm shift for the payments ecosystem because the agent is now an extra party that has entered the loop and this increases complexity for merchants; in agentic commerce, it’s important to determine the identity of an agent, and this is what Mastercard Agent Pay can do; in agentic commerce, the consumer identity also needs to be determined, as in a traditional online transaction; there are tricky aspects on agentic commerce to solve, such as handling a challenged transaction, and this is something Mastercard can do; management thinks agentic commerce will be very hard to handle for local payment networks, and this will be an opportunity for Mastercard to win share; the transition from physical payment to online payment unlocked new suite of services Mastercard could provide, and management expects a similar thing to happen with the transition to agentic commerce

What we’re seeing is behavioral change, driven and powered by generative AI and bots and so forth, where search behavior is changing. So, that’s on the consumer side, if we start right there. So, consumers are migrating their search increasingly so to their favorite chatbot and they’re asking their queries there, and they get potentially better answers, who knows…

…It’s really quite a significant paradigm shift for the payment ecosystem, because in the payment ecosystem, what happens is there’s now an extra party that has entered the realm, and that is the agent. So, that comes with a lot of those aspects you just talked about in your question, is there’s legal questions, there’s a security question…

…Some of the things that need to happen in a world of agentic commerce is, the first is, is this a real bot? Is this a bot that we believe matches up to Mastercard’s safety and security standards? So, we will certify and register bots out there. So that’s what Mastercard Agent Pay does. So, nothing really new from us on a perspective, but it’s a new party. Not really visible to the consumer in that way, but certainly driving some complexity potentially for merchants, for issuers, for every other party because that is just a new flow for the transaction…

…The merchant needs to know that the agent on the other side that we have certified is actually the agent. So, we have to pass through that information and ensure that the circle closes. We’re doing that. Well, there’s still the question of what is in focus today very much so is the consumer, the person they claim to be. So, consumer authentication needs to continue, but it now needs to flow through a somewhat more complicated transaction. So, all of this is happening…

…If you have asked an agent to buy you something in a chat, and then in the end, you challenge that transaction, who can prove who’s right. Is it the consumer? Is it the merchant? What happens? What do you do on return policies and various other things. Those are all complexities that we’re pretty good at solving in today’s world, and they were pretty busy solving in the future world, and that comes down to some of the aspects that you’ve talked about in your question. Where is the legal and regulatory framework on this yet? This is not something that’s specifically contemplated, but that will evolve over time…

…On the point of challenging a transaction. We’ve bought a company a couple of years ago called Ethoca, and what they do is they provide transaction detail at the moment of a charge back to a consumer that says, “Hey, you actually did this transaction because you were here at this time doing the following.” And the same can be done with this audit trail that would be capturing out of the chat that I talked about earlier. That is one example…

…One thing that I think is a pretty obvious opportunity is, this is going to be very hard to do for local payment networks. So, if you look around various kind of local payment systems that exist in Europe, in Asia and so forth. Big markets for us is an opportunity for us to continue to drive up our switching ratio as we’ve done in years, and this gives us another, kind of, field to execute on. I think that’s the first thing to say…

…You think back about the days where everything was in store and what kind of services portfolio we had and the opportunities we had to apply services and drive differentiation for us versus others. And then it went online. There was a whole different set of solutions that were suddenly needed to keep the online transaction safe. And agentic, it’s going to be even more opportunity for us to do that.

MercadoLibre (NASDAQ: MELI)

MercadoLibre’s management is very excited about the potential of infusing AI agents within MercadoLibre’s ecosystem; management recently launched Seller Assistant, a chatbot that provides personalized advice and recommendations to sellers; in MercadoPago, management just launched an AI assistant that can help users with a wide range of tasks; management thinks it’s still early in terms of determining OpenAI’s impact on e-commerce but what MercadoLibre needs to do is to develop agentic capabilities first so that it can be utilised if needed

We are extremely excited about the potential of Agent to enhance discovery, service and productivity within our ecosystem. There are several examples of things that we are doing on that regard. We just launched our own Seller Assistant, which is a conversational tool that gives sellers personalized advice and recommendations on how to manage data activity in our platform. In FinTech, as you probably know, we just launched our first AI assistant that can help our users with a wide range of tasks like making or scheduling money transfer through a conversation platform, asking for questions on the user’s operation and so on…

…[Question] I wanted to hear from you how you are thinking about OpenAI’s recent move into e-commerce?

[Answer] We need to continue to focus ourselves in building the best agentic experience within our platform, and that will give us optionality on what to do next and how to move forward. I think it’s early to make comments on OpenAI and their partnership with Etsy, Shopify, and so on. We need to understand how this will develop in the long run, what role agent will play in the relationship with consumers. And eventually, decide if there’s something different that we need to do for sure. We need to put the technology in place in order to have an agentic experience in MercadoLibre and in Mercado Pago in the near term.

Meta Platforms (NASDAQ: META)

Meta’s management is building an industry-leading amount of compute to be ready for whenever superintelligence arrives; if superintelligence takes longer than expected, the extra compute can be used to accelerate Meta’s core business; Meta’s core business has been able to profitably use much more compute than what’s available; management is seeing very high demand for compute; the worst case for building compute now is that Meta will be growing into the compute that it’s building; management recognises the possibility that Meta could overshoot on building compute capacity, and if so, it will lead to the worst case scenario

We’re also building what we expect to be an industry-leading amount of compute. Now there’s a range of time lines for when people think that we’re going to get superintelligence. Some people think that we’ll get there in a few years. Others think it will be 5, 7 years or longer. I think that it’s the right strategy to aggressively frontload building capacity so that way we’re prepared for the most optimistic cases. That way, if superintelligence arrives sooner, we will be ideally positioned for a generational paradigm shift in many large opportunities. If it takes longer, then we’ll use the extra compute to accelerate our core business which continues to be able to profitably use much more compute than we’ve been able to throw at it. And we’re seeing very high demand for additional compute, both internally and externally. And in the worst case, we were just slow building new infrastructure for some period while we grow into what we build…

…Now I mean, it’s, of course, possible to overshoot that, right? And if we do… the kind of the very worst case would be that we effectively have just prebuilt for a couple of years, in which case, of course, there would be some loss and depreciation, but we’d grow into that and use it over time.

AI recommendation systems are improving the content delivered across Facebook, Instagram, and Threads; AI recommendation systems have led to 5% more time spent on Facebook in 2025 Q3, and 10% on Threads; AI recommendation systems have led to 30% more time spent on video in Instagram in 2025 Q3; improvements in Meta’s AI recommendation systems will also benefit the company with the coming growth of AI-generated content; Facebook is now surfacing twice as many Reels published that day than at the start of 2025; management expects to evolve Instagram’s recommendation systems in 2026 to surface broader content that cater to diverse interests of each person; Meta has produced promising results in creating foundational ranking models and management expects to significantly scale up data and compute for training recommendation models in 2026 to yield better recommendations; management expects Meta to leverage LLMs (large language models) in 2026 to improve understanding of content by the recommendation systems; ranking optimisations made in 2025 Q3 alone led to a 10% increase in time spent on Threads

Across Facebook, Instagram and Threads, our AI recommendation systems are delivering higher quality and more relevant content, which led to 5% more time spent on Facebook in Q3 and 10% on Threads. Video is a particular bright spot with video time spent on Instagram up more than 30% since last year…

…Improvements in our recommendation systems will also become even more leveraged as the volume of AI-created content grows. Social media has gone through 2 eras so far. First was when all content was from friends, family and accounts that you followed directly. The second was when we added all of the creator content. Now as AI makes it easier to create and remix content, we’re going to add yet another huge corpus of content on top of those. Recommendation systems that understand all this content more deeply and can show you the right content to help you achieve your goals are going to be increasingly valuable…

…On Facebook, our systems are now surfacing twice as many Reels published that day than at the start of the year.

Looking to 2026, we expect to advance our recommendation systems across several dimensions. On Instagram, one focus is evolving our systems to surface content across a broader set of topics that cater to the diverse interest of each person. This follows a similar approach we’ve implemented on Facebook that has driven good results. We also expect to make significant progress on our longer-term ranking innovations in 2026. We’re seeing promising new results from our research efforts to create foundational ranking models and expect the new model innovations we’re developing as part of this will enable us to significantly scale up the amount of data and compute we use to train our recommendation models in 2026, yielding more relevant recommendations.

Another large focus next year is leveraging LLMs to improve content understanding. We expect this is going to enable our systems to more precisely label the keywords and topics within videos and posts, which will allow our systems to both develop deeper intuition about a person’s interest and retrieve the content that matches them…

…The ranking optimizations we made in Q3 alone drove a 10% increase in time spent on Threads.

Meta’s advertising business has benefited from improvements in AI ranking systems; the unification of different models into simpler, general models have led to improvements in the advertising business in 2025 Q3; management rolled out Lattice, its unified model architecture for advertising ranking models, to app ads in 2025 Q3 and drove a 3% gain in conversions; since the introduction of Lattice and other improvements in 2023, Meta has reduced the number of ads ranking and recommendation models by around 100, and the reductions have led to performance improvements; management expects Meta to achieve additional gains as it consolidates another 200 models over the coming years; management is innovating on run time models used for advertising inference; a new run time advertising ranking model was piloted in 2025 Q3 that uses more compute and data than prior models, and it drove a lift in conversions on Instagram of more than 2%; management has improved the performance of the Andromeda model architecture in 2025 Q3, driving a 14% increase in advertising quality on Facebook surfaces

Our ads business continues to perform very well, largely due to improvements in our AI ranking systems as well. This quarter, we saw meaningful advances from unifying different models into simpler, more general models, which drive both better performance and efficiency…

…We are driving performance gains through ongoing improvements in our larger scale ads ranking models. For example, we continue to broaden the adoption of Lattice, our unified model architecture. In Q3, we rolled out Lattice to app ads, which drove a nearly 3% gain in conversions for that objective. 

Since introducing Lattice back in 2023, along with other back-end improvements, we have now cut the number of ads ranking and recommendation models by approximately 100 as we consolidated smaller and more specialized models into larger ones that use the Lattice architecture to generalize learnings across surfaces and objectives. We continue to observe performance improvements as we combine models and expect to drive additional gains as we consolidate another 200 models over the coming years into a smaller number of highly capable models…

…We’re innovating on our run time models we use downstream of them for ads inference. For example, we began piloting a new run time ads ranking model in Q3 that leverages more compute and data than our prior models to select more relevant ads. In testing, we’ve seen this new model drive a more than 2% lift in conversions on Instagram.

We also significantly improved performance of Andromeda in Q3 by combining models across retrieval and early-stage ranking into a single model, driving a 14% increase in ads quality on Facebook surfaces.

Meta’s end-to-end AI-powered advertising tools, which are under Advantage+, is now handling $60 billion in annualised run rate revenue; management rolled out a streamlined campaign creation flow for Advantage+ lead campaigns in 2025 Q3, so end-to-end automation is turned on from the beginning; the number of advertisers using at least 1 of Advantage+’s video generation features grew 20% sequentially in 2025 Q3; management has added more generative AI features to Advantage+ to help advertisers optimise and improve ad creatives; management introduced AI generated music in Advantage+ in 2025 Q3; management continues to think a fully automated AI advertising product, where advertisers just have to tell the system what its objectives are, and the AI figures out everything else, is still important; advertisers who run lead campaigns using Advantage+ are seeing a 14% lower cost per lead; a lot of advertisers only use Advantage+ for a portion of their campaigns, so management thinks there are share gains to be made

Now the annual run rate going through our completely end-to-end AI-powered ad tools has passed $60 billion…

…In Q3, we completed the rollout of our streamlined campaign creation flow for Advantage+ lead campaigns. So now advertisers running sales app or lead campaigns have end-to-end automation turned on from the beginning, allowing our systems to look across our platform to optimize performance by automatically choosing criteria like who to show the ads to and where to show them. The annual run rate of revenue running through our end-to-end automated solutions has now reached $60 billion following the implementation of the new streamlined creation flow, as we continue to see more advertisers leverage the performance benefits of our solutions.

Within our Advantage+ creative suite, the number of advertisers using at least 1 of our video generation features was up 20% versus the prior quarter as adoption of image animation and video expansion continues to scale. We’ve also added more generative AI features to make it easier for advertisers to optimize their ad creatives and drive increased performance. In Q3, we introduced AI generated music so advertisers can have music generated for their ad that aligns with the tone and message of the creative…

…I mean there’s one opportunity that we just usually talk about on these calls, but hasn’t come up as much here is just the ability to make it so that advertisers are increasingly just going to be able to give us a business objective and give us a credit card or bank account and like have the AI system basically figure out everything else that’s necessary. Including generating video or different types of creative that might resonate with different people that are personalized in different ways, finding who the right customers are. All of these — all of the capabilities that we’re building, I think, go towards improving all of these different things. So I’m quite optimistic about that…

…Advertisers who run lead campaigns using Advantage+ are seeing a 14% lower cost per lead on average than those who are not…

…A lot of advertisers only use our end-to-end automated solutions for a portion of their campaigns so we can grow share there. And to capture that opportunity, we’re focused on driving continued performance improvements and addressing some of the key use cases that we still need in order to grow adoption.

Meta AI has more than 1 billion monthly actives, with usage increasing as the underlying models improve; the majority of Meta AI’s responses to queries in the US now show related Reels; users have created over 20 billion images with Meta AI; the launch of Vibe within Meta AI in September has led to a 10x increase in media generation in Meta AI; Meta AI is still powered by Llama 4

More than 1 billion monthly actives already use Meta AI and we see usage increase as we improve our underlying models…

…We’re increasingly leveraging first-party content into Meta AI results with the majority of Meta AI’s responses to Facebook Deep Dive queries in the U.S. now showing related Reels. We’re also seeing a lot of traction with media generation. People have created over 20 billion images using our products. And since launching Vibes within Meta AI in September, we have seen media generation in the app increased more than tenfold…

…A lot of people use Meta AI today. I mean, as I said in my comments upfront, there’s more than 1 billion people who use it on a monthly basis. And what we see is that as we improve the quality of the model, primarily for post-training Llama 4 at this point. We are — we continue to see improvements in usage.

Meta sees more than 1 billion active threads happening everyday with business accounts across its messaging platforms; management thinks Meta’s Business AI will help tens of millions of businesses scale the conversations and improve sales at low cost; business messaging continues to be a significant opportunity for Meta; Click-to-WhatsApp ads revenue was up 60% year-on-year in 2025 Q3; management has broadened Business AI access in the initial test markets of Philippines and Mexico, and strong usage has been seen, with millions of conversations between people and Business AIs taking place since July; in the US, management is rolling out the ability for merchants to add their Business AIs to their website 

Every day, people have more than 1 billion active threads with business accounts across our messaging platforms, ranging from product questions to customer support. Our business AIs will enable tens of millions of businesses to scale these conversations and improve their sales at low cost and the better our models get, the better this is going to work for all businesses…

…Business messaging remains a significant opportunity for us. We’re seeing strong growth across our portfolio of solutions, including with Click-to-WhatsApp ads, which grew revenue 60% year-over-year in Q3.

We’re also making good progress on our business AI efforts, where we’ve been focused on building a turnkey AI that helps businesses generate leads and drive sales. We’ve been opening access in recent months to more businesses within our initial test markets, the Philippines and Mexico. And we’ve seen strong usage with millions of conversations between people and Business AIs taking place since July. This month, we expanded availability within WhatsApp and Messenger to all eligible businesses in Mexico and the Philippines, respectively. In the U.S., we’re also starting to roll out the ability for merchants to add their Business AIs to their website so we can support the full sale funnel from ad to purchase.

Retention at Vibes is looking good so far, with usage growing fast weekly; management sees Vibes as new content type enabled by AI; the launch of Vibe within Meta AI in September has led to a 10x increase in media generation in Meta AI

This quarter, we also launched Vibes which is the next generation of our AI creation tools and content experiences. Retention is looking good so far. And its usage keeps growing quickly week over week…

…I think that Vibes is an example of a new content type enabled by AI, and I think that there are more opportunities to build many more novel types of content ahead as well…

…And since launching Vibes within Meta AI in September, we have seen media generation in the app increased more than tenfold.

The response to Meta’s 2025 line of AI glasses has been great; sales of the new Ray-Ban Meta glasses and Oakley Meta Vanguards are both good; the new Meta Ray-Ban Display glasses that come with the neural band as an interaction touch-point, sold out within 48 hours; management wants to invest to increase manufacturer of the Meta Ray-Ban Display glasses; management thinks there’s huge opportunity ahead with the Meta Ray-Ban Display glasses; management thinks that if the smart glasses continue on their current trajectory, then Meta’s ongoing investments in Reality Labs (via operating losses) will generate a good return; the return on investment of the smart glasses will come from both the hardware sales and new AI-enabled services that are layered on top; management will continue investing in virtual reality hardware products, such as the Orion

At Connect, we announced our 2025 line of AI glasses, and the response so far has been great. The new Ray-Ban Meta glasses and Oakley Meta Vanguards are both selling well as people love the improved battery life, camera resolution, new AI capabilities and the great design.

And there’s our new Meta Ray-Ban Display glasses, our first glasses with a high-resolution display and the Meta Neural Band to interact with them. They sold out in almost every store within 48 hours with demo slots fully booked through the end of next month. So we’re going to have to invest in increasing manufacturing and selling more of those. This is an area where we are clearly leading and have a huge opportunity ahead…

…[Question] On wearables, in particular, do you think you’ll be able to sell enough hardware to recoup your investment?

[Answer] The work on Ray-Ban Meta and the Oakley Meta product is going very well. I think, yes, I mean, at some point, if these continue going as well as it has been, then I think it will be a very profitable investment. I think that there’s some revenue that we get from basically selling the devices and then some that will come from additional services from the AI on top of it. So I think that there’s a big opportunity. Certainly, the investment here is not just to kind of build just the device. It’s also to build these services on top. Right now, a lot of people get the devices for a range of things that don’t even include the AI even though they like the AI. But I think over time, the AI is going to become the main thing that people are using them for and I think that that’s going to end up having a big business opportunity by itself.

But as products like the Ray-Ban Meta and Oakley Metas are growing, we’re also going to keep on investing in things like the more full field of view, product form of the Orion prototype that we showed at Connect last year. So those things are obviously earlier in their curve towards getting to being a sustaining business. And our general view is that we want to build these out to reach many hundreds of millions or billions of people and that’s the point at which we think that this is going to be just an extremely profitable business.

Meta’s management is focused on preserving maximum long-term flexibility for Meta’s AI capex; Meta Superintelligence Labs’ compute needs account for the largest chunk of Meta’s capex growth in 2026; when management was planning for 2025’s capex, they had investments they thought would be paying off in 2026, and those are already paying off through the course of 2025; one of the ways management looks at the ROIC (return on invested capital) of AI capex is growth in conversions relative to impressions, and Meta is putting out conversion growth that is faster than impressions; the new model architectures Meta has been deploying in its advertising systems has enabled Meta to deploy more data and compute to drive ads performance management expects this to continue in 2026; management wishes they had more compute capacity today than what’s available and they know that at least some of the capacity can be put towards positive ROI use-cases in the core business

Our primary focus is deploying capital to support the company’s highest order priorities including developing leading AI products models and business solutions. As we make significant investments in infrastructure to support this work, we are focused on preserving maximum long-term flexibility to ensure we can meet our future capacity needs while also being able to respond to how the market develops in the years ahead. We’re doing so in several ways, including staging data center sites so we can spring up capacity quickly in future years as we need it as well as establishing strategic partnerships that give us option value for future compute needs…

…I will say that the growth in 2026 CapEx relative to 2025 comes from growth in each of the core areas, MSL, core AI as well as non-AI spend. So all of those areas are growing, but the MSL AI needs are growing the most…

…[Question] Can you help us a little to understand some of the early quantifiable signals you’re seeing on AB tests from some of these improvements to come that sort of make you most excited and give you confidence you’re going to get ROIC from all this CapEx?

[Answer] In terms of the core AI pipeline, I think, we talked about last year when we were going into the 2025 budget process, we had a road map of resource investments across both head count and compute that we thought would pay off in 2026. And it’s really a very broad range of sort of different ads ranking and performance efforts. And we’re continuing to see that those have paid off through the course of the year. There is a long list of specific efforts, but 1 of the measures that we look at to monitor this is how are we driving ad performance, how are conversions growing?

Conversions is a complex metric for us because advertisers optimize for so many different conversions on different values. But when we control for that and look at value-weighted conversion rates, we’re seeing very strong year-over-year growth in conversion — weighted conversions continue to grow faster than impressions.

We also talked about some of the new model architecture over the course of the year and the degree to which the new model architecture is enabling us also to take advantage of having more data and more compute to drive ads performance. So we expect that, that’s going to be a continued story in 2026. We are, in fact, at the beginning of our 2026 budgeting process now, and we see a similar list of revenue investments that we’re excited to be able to invest in. And so we think that, that’s going to be a big part of our ability to continue to drive strong revenue performance throughout the year…

…We’re certainly seeing that we wish we had more capacity today than we do. We would be able to put it towards good use certain not only with the MSL team appreciate having more capacity, but we’d be able to put it towards good and ROI-positive use in the core business as well.

Meta’s management  has repeatedly seen a pattern of Meta building compute capacity based on an aggressive assumption, only to see even higher demand for compute; Meta’s core business keeps having the ability to use more compute in profitable ways than what’s available

To date, we keep on seeing this pattern where we build some amount of infrastructure to what we think is an aggressive assumption. And then we keep on having more demand to be able to use more compute, especially in the core business in ways that we think would be quite profitable, then we end up having compute for.

Meta does not use its large models for inference work because that is too expensive; Meta gets the large models to transfer knowledge to smaller models for inference work

We don’t use our larger model architectures like GEM for inference because their size and complexity would make it too cost prohibitive. The way that we drive performance from those models is by using them to transfer knowledge to smaller lightweight models that are used at run time.

Meta’s management  is unsure of the margin-profile of the new products Meta may develop with AI

[Question] You mentioned the prior 2 content cycles, and obviously, you’ve been able to generate very attractive margins on them. As we get into the AI cycle, obviously, some concerns on the investment. But can you talk a little bit about how you’re thinking about tools that could be coming out for users? I know there’s some new competition. And then secondly, how do you think about margins in this content cycle? Any reason to think they would be different versus prior cycles.

[Answer] I think it’s too early to really understand what the margins are going to be for the new products that we build. I mean, I think certainly, every — each product has somewhat different characteristics. And I think we’ll kind of understand how that goes over time. I mean, my general goal is to build a business that maximizes value for the people who use our products and maximizes profitability, not margin. So I think we’ll kind of just try to build the best things that we can and try to deliver the most value that we can for most people.

Meta’s management  thinks being the best at a given capability in the AI world will drive the greatest returns; management thinks it’s unlikely that one company will become the best at all capabilities; management wants Meta to develop novel capabilities with AI

I think the art of product development here is looking at the list of technology capabilities and figuring out what new products are going to be useful and prioritizing those. But fundamentally, I would sort of expect this exponential curve in new technology capabilities that are going to become available. And the other thing that I expect is that I think being the best in a given area will drive great returns rather than — this is not like a check-the-box exercise of like, okay, we can generate some kind of content and someone else can. I think that like the company that is the best at each of these capabilities, I think, will get a large amount of the potential value for doing that. So there are lots of different capabilities to build. I’m not sure that any one company is going to be the best at all of them. I doubt that’s going to be the case. But a lot of what we’re trying to do is not like not kind of do some things that others have done. We’re really trying to build novel capabilities.

Meta’s management  thinks a lot of AI apps today are still really small, but there’s huge opportunity

But if you look at it today, the companies that are building apps, I mean, a lot of the apps are still relatively small. And I think that’s obviously going to be a huge opportunity.

Meta’s management  thinks AI is different from past technological developments because AI allows new capabilities to be introduced fast, and new products and businesses can be built around these capabilities

I think what we haven’t really seen as much in the history of the technology industry is the rate of new capabilities being introduced because around each of these capabilities, you can build many new products that I think each will turn into interesting businesses.

Microsoft (NASDAQ: MSFT)

Microsoft and OpenAI have a new agreement; Microsoft’s investment in Open AI has 10x-ed in value; under the new agreement, Open AI has a $250 billion contract with Azure while Microsoft has model and product IP rights to 2032; management does not think AGI will be achieved any time soon, but a lot of value from AI can still be derived

We closed a new definitive agreement with OpenAI, marking the next chapter in what is one of the most successful partnerships and investments our industry has ever seen…

…Already, we have roughly 10x-ed our investment. OpenAI has contracted an incremental $250 billion of Azure services, our rev share, exclusive IP rights and API exclusivity for Azure continue until AGI or through 2030. And we have extended the model and product IP rights through 2032…

…I don’t think AGI as defined at least by us in our contract is ever going to be achieved anytime soon. But I do believe we can drive a lot of value for customers with advances in AI models by building these systems.

Azure has the most expansive data center fleet for the AI era and is adding capacity at scale; Azure will increase AI capacity by >80% in FY2026 and will double its total data center footprint in 2 years as management sees strong demand; Azure announced the most powerful AI data center in the world in 2025 Q3 and it will start operations in 2026 and scale to 2 gigawatts; Azure has the world’s first large-scale cluster of NVIDIA GB300s; Azure is building a fungible GPU fleet that’s continuously modernised for all stages of the AI lifecycle (from pretraining to inference) and for workloads that go beyond generative AI; management thinks Azure has the best ROI (return on investment) and TCO (total cost of ownership) for customers; Azure increased the token throughput of GPT-4.1 and GPT-5 by 30% per GPU in 2025 Q3 (FY2026 Q1); Azure is supporting sovereign AI needs; Azure has customers in 33 countries who are developing their AI capabilities within local borders, such as OpenAI and SAP in Germany; Azure has Azure AI Foundry to help customers build own AI apps and agents; Azure AI Foundry offers enterprises access to 11,000 models (including GPT-5 and Grok 4) which is more than any competitor; Azure has 80,000 customers; Azure AI Foundry also provides other tools beyond models for developers to customize and manage AI applications and agents; real production-scale AI deployments are driving Azure’s overall growth; Azure took share again in 2025 Q3 (FY2026 Q1)

We have the most expansive data center fleet for the AI era, and we are adding capacity at an unprecedented scale. We will increase our total AI capacity by over 80% this year and roughly double our total data center footprint over the next 2 years, reflecting the demand signals we see. Just this quarter, we announced the world’s most powerful AI data center, Fairwater in Wisconsin, which will go online next year and scale to 2 gigawatts alone.  And we have deployed the world’s first large-scale cluster of NVIDIA GB300s. We are building a fungible fleet that’s been continuously modernized and spans all stages of the AI life cycle from pretraining to post training to synthetic data generation and inference. And it also goes beyond GenAI workloads to recommendation engines, databases and streaming. We’re optimizing this fleet across silicon systems and software to maximize performance and efficiency. 

It’s this combination of fungibility and continuous optimization that allows us to deliver the best ROI and TCO for us and our customers. For example, during the quarter, we increased the token throughput for GPT-4.1 and GPT-5, two of the most widely used models by over 30% per GPU.

We also have the most comprehensive digital sovereignty platform. Azure customers in 33 countries are now developing their own cloud and AI capabilities within their borders to meet local data residency requirements. In Germany, for example, OpenAI and SAP will rely on Azure to deliver new AI solutions to the public sector…

…We are building Azure AI Foundry to help customers build their own AI apps and agents. We have 80,000 customers, including 80% of the Fortune 500. We offer developers and enterprise access to over 11,000 models, more than any other vendor, including as of this quarter, OpenAI’s GPT-5 as well as xAI’s Grok 4…

…Beyond models in Foundry, we are providing everything developers need to design, customize and manage AI applications and agents at scale. Our new Microsoft Agent Framework helps developers orchestrate multi-agent systems with compliance, observability and deep integration out of the box…

…These kinds of real production scale AI deployments are driving Azure’s overall growth. And once again, this quarter, Azure took share.

Ralph Lauren used Azure AI Foundry to build a conversational shopping experience; Open Evidence used Azure AI Foundry to build a clinical assistant; KPMG used the Microsoft Agent Framework in Azure AI Foundry to connect agents with internal data

For example, Ralph Lauren used Foundry to build conversational shopping experience in its app, enabling customers to describe what they’re looking for and get personalized recommendations. And OpenEvidence used Foundry to create its AI-powered clinical assistant which surfaces relevant medical information to physicians and help streamline charting…

…KPMG used the framework to modernize the audit process, connecting agents to internal data with enterprise-grade governance and observability.

Microsoft has 900 million MAU (monthly active users) of AI features across its products; Microsoft’s family of Copilot apps now has 150 million MAU (was 100 million in 2025 Q2); management sees Copilot becoming the UI (user interface) for agentic AI; a chat feature released in Microsoft 365 just 9 months ago already has tens of millions of users; adoption of chat is up 50% sequentially in 2025 Q3 (FY2026 Q1), and usage intensity is increasing; management introduced Agent Mode in 2025 Q3 (FY2026 Q1), which can turn prompts into full Powerpoint slides or Excel spreadsheets; Agent Mode is ranked best-in-class by 3rd-party benchmarks; adoption of Microsoft 365 Copilot is growing super fast; more than 90% of the Fortune 500 are using Microsoft 365 Copilot; a number of large companies each purchased over 15,000 Microsoft Copilot seats in 2025 Q3 (FY2026 Q1); Lloyds Banking Group deployed 30,000 Microsoft Copilot seats in 2025 Q3 (FY2026 Q1), saving each employee 46 minutes daily; enterprises are coming back to purchase even more seats of Microsoft 365 Copilot after the first purchase; PwC employees interacted with Microsoft 365 Copilot over 30 million times in 6 months, saving millions of hours on employee productivity

We now have 900 million monthly active users of our AI features across our products. And our first-party family of Copilots now has surpassed 150 million monthly active users across the information work, coding, security, science, health and consumer.  

When it comes to information work, we continue to innovate with Microsoft 365 Copilot. Copilot is becoming the UI for the agentic AI experience. We have integrated chat and agentic workflows into everyday tools like Outlook, Word, Excel, PowerPoint and Teams. Just 9 months since release, tens of millions of users across Microsoft 365 customer base are already using chat. Adoption is accelerating rapidly, growing 50% quarter-over-quarter, and we continue to see usage intensity increased. 

This quarter, we also introduced Agent Mode, which turns single prompts into export quality Word documents, Excel spreadsheets, PowerPoint presentation and then iterate to deliver the final product much like agent mode in coding tools today. We’re thrilled by the early response, including third-party benchmarks that rank it best-in-class…

…Customers continue to adopt Microsoft 365 Copilot at a faster rate than any other new Microsoft 365 suite. All up more than 90% of the Fortune 500 now use Microsoft 365 Copilot. Accenture, Bristol-Myers Squibb, EY Global and the U.K.’s Tax and Payments and Customs Authority all purchased over 15,000 seats this quarter. Lloyds Banking Group has deployed 30,000 seats, saving each employee an average of 46 minutes daily. And a large majority of our enterprise customers continue to come back to purchase more seats. Our partner, PwC, alone added 155,000 seats this quarter and now has over 200,000 deployed across its global operations. In just 6 months, PwC employees interacted with Microsoft 365 Copilot over 30 million times, and they credit this agentic transformation with saving millions of hours on employee productivity.  

Microsoft’s management is observing a growing list of software companies, including Adobe and Asana, building their own agents that connect with Copilot; management is seeing customers building their own agents that connect with Copilot; the number of agent users doubled sequentially in 2025 Q3 (FY2026 Q1); management has announced App Builder, a new Copilot agent that turns prompts into apps and agents in Microsoft 365

We are seeing a growing Copilot agent ecosystem with top ISVs like Adobe, Asana, Jira, LexisNexis, SAP, ServiceNow, Snowflake and Workday, all building their own agents that connect to Copilot. And customers are also building agents for their mission-critical business processes and workflows using tools like Copilot Studio and integrating them into Copilot. The overall number of agent users doubled quarter-over-quarter. And just yesterday, we announced App Builder, a new Copilot agent that lets anyone create and deploy task-specific apps and agents in minutes grounded in Microsoft 365 context.

Github Copilot is the most popular AI-pair programmer now with >26 million users; tens of thousands of developers at AMD use GitHub Copilot and they are saving months of developer time; Github now has 180 million developers, and is growing its fastest rate ever; 80% of new developers start on Github with Copilot; GitHub Copilot had 500 million pull requests merged over the past year; management has released Agent HQ; management sees GitHub Copilot and Agent HQ as the organising layer for all coding agents 

GitHub Copilot is the most popular AI pair programmer now with over 26 million users…

…Tens of thousands of developers at AMD use GitHub Copilot, accepting hundreds of thousands of lines of code suggestions each month and crediting it with saving months of development time…

…GitHub is now home to over 180 million developers and the platform is growing at the fastest rate in its history, adding a developer every second. 80% of new developers on GitHub start with Copilot within the first week. Overall, the rise of AI coding agents is driving record usage with over 500 million pull requests merged over the past year.

And just yesterday, at GitHub Universe, we introduced Agent HQ. GitHub Copilot and Agent HQ is the organizing layer for all coding agents, extending GitHub privatives like PRs, issues, actions to coding agents from OpenAI, Anthropic, Google, Cognition, xAI as well as OSS and in-house models. GitHub now provides a single mission control to launch, manage and review these agents, each operating from its own branch with built-in controls, observability and governance.

Half of Microsoft’s cloud and AI-related capex in 2025 Q3 (FY2026 Q1) are for long-lived assets that will support monetisation over the next 15 years and more, while the other half are for CPUs and GPUs, driven by strong AI- and Azure-related demand; there is a difference between Microsoft’s total capital expenditure and cash expenditure because of the use of finance leases; Microsoft’s AI capital expenditure for CPUs and GPUs are backed by signed-contracts and the useful lives of the GPUs are quite matched with the duration of the contracts; Microsoft’s AI capital expenditure for long-lived assets are not backed by contracts, but management is confident these assets will be useful over their lifespans; when building AI infrastructure, management’s priority is for Microsoft’s internal workloads, such as Copilot and AI research

Capital expenditures were $34.9 billion, driven by growing demand for our Cloud and AI offerings. This quarter, roughly half of our spend was on short-lived assets, primarily GPUs and CPUs, to support increasing Azure platform demand, growing first-party apps at AI solutions, accelerating R&D by our product teams as well as continued replacement for end-of-life server and networking equipment. The remaining spend was for long-lived assets that will support monetization for the next 15 years and beyond, including $11.1 billion of finance leases that are primarily for large data center sites. And cash paid for PP&E was $19.4 billion. As a reminder, the difference between total CapEx and cash paid for PP&E is primarily due to finance leases as well as the normal timing of goods received, but not yet paid…

…Increasingly, we talked about this short-lived assets, both GPUs and CPUs, Again, we talk about all these workloads are burning both in terms of app building. Now when that happens, short-lived assets generally are done to match sort of the duration of the contracts or the duration of your expectation of those contracts. And so I sometimes think when people think about risk, they’re not realizing that most of the lifetimes of these and the lifetime of the contracts are very similar. And so when you think about having revenue and the bookings and coming on the balance sheet, the depreciation of short-lived assets, they’re actually quite matched, Mark…

… We’re continuing to do that also using leases. Those are very long-lived assets, as we’ve talked about 15 to 20 years. And over that period of time, do I have confidence that we’ll need to use all of that, it is very high…

…Because when you think about real priorities that you have to fill first, it’s obviously the increasing usage and adoption and sales we’ve seen of M365 Copilot and the usage of Copilot chat, which we’ve seen very different patterns, which we’re encouraged by. It’s the adoption of security features. It’s the GitHub momentum.  And so when you’re thinking about it, that is where and it is a priority for us to allocate resourcing there first. And so you are right to ask how do I think about that. We’ve worked very hard to try to mitigate it as best we can, but we have been short in Azure, and we’ve been clear on it. And I would say the other 2 priorities that I haven’t mentioned maybe as much before is also just making sure our product teams and the AI talent that we’ve been able to hire into the company really over the past 1.5 years have access also to significant capacity because we’re seeing it make the product better in a loop that is adding great benefit today into products people are using today for real-world work. And so we are making that a priority to make sure our research teams have that as well as our product engineering teams. And yes, it does impact Azure directly. That is the place where you see that prioritization. But I think it’s probably hard for me to give an exact number, but it is safe to say that the number could be higher.

Azure grew revenue by 40% in 2025 Q3 (FY2026 Q1) (was 39% in 2025 Q2); Azure’s core infrastructure business had better than expected growth; Azure’s AI services revenue was in line with expectations; Azure was capacity-constrained in 2025 Q3 (FY2026 Q1) despite bringing more capacity online; management expects Azure to be capacity-constrained through at least FY2026; management will continue to balance capacity-additions between Azure’s revenue growth, and Microsoft’s internal-needs for compute; the demand signals that management is seeing is accelerating faster than they expected; management is seeing demand increasing across many places and they are investing in capacity with confidence in usage patterns and in bookings

In Azure and other Cloud services, where we continue to see accelerating demand, revenue grew 40% and 39% in constant currency. Results were ahead of expectations, driven by better-than-expected growth in our core infrastructure business, primarily from our largest customers. Azure AI services revenue was generally in line with expectations, and this quarter, demand again exceeded supply across workloads, even as we brought more capacity online…

…In Azure, we expect Q2 revenue growth of approximately 37% in constant currency as demand remains significantly ahead of the capacity we have available. And while we’re accelerating the amount of capacity we’re bringing online, we will continue to balance Azure revenue growth with the growing needs across our first-party apps and AI solutions, our own R&D efforts and the end-of-life server replacements. Therefore, we now expect to be capacity constrained through at least the end of our fiscal year…

…Demand signals across bookings, RPO and product usage are accelerating faster than we expected. We’re investing in infrastructure, AI talent and product innovation to capture that momentum and expand our leadership position…

…Demand is increasing. It is not increasing in just one place. It is increasing across many places. We’re seeing usage increases in products. We are seeing new products launch that are getting increasing usage, and increasing usage very quickly. When people see real value, they actually commit real usage. And I sometimes think this is where this cycle needs to be thought through completely is that when you see these kind of demand signals and we know we’re behind, we do need to spend. But we’re spending with a different amount of confidence in usage patterns and in bookings, and I feel very good about that. 

Azure is expected to grow revenue by 37% in 2025 Q4 (FY2026 Q2) in constant currency, driven by demand that remains significantly ahead of capacity; management now expects capital expenditure in FY2026 to have a higher growth rate than in FY2025 (previous guidance was for capital expenditure growth in FY2026 to moderate from FY2025’s level) because of an increase in spend on GPUs and CPUs

For Intelligent Cloud, we expect revenue of USD 32.25 billion to USD 32.55 billion or growth of 26% to 27%. In Azure, we expect Q2 revenue growth of approximately 37% in constant currency as demand remains significantly ahead of the capacity we have available… As a reminder, there can be quarterly variability in the year-on-year growth rates depending on the timing of capacity delivery and when it comes online as well as from in-period revenue recognition depending on the mix of contracts…

…Capital expenditures. With accelerating demand and a growing RPO balance, we’re increasing our spend on GPUs and CPUs. Therefore, total spend will increase sequentially, and we now expect the FY ’26 growth rate to be higher than FY ’25. 

Microsoft’s management thinks AI models, even when they become more powerful over time, will have spiky intelligence (being really good at only certain areas), and software systems such as GitHub Agent HQ or M365 Copilot or Azure AI Foundry will be needed to smooth out the spikiness

I think your question touches on something that’s pretty important, which is how are these AI systems going to truly be deployed in the real world and make a real difference and make a return for both the customers who are deploying them and then obviously, the providers of these systems. And I think the best way to characterize the situation is that even as the intelligence capability increases, let’s even say, exponentially like model version over model version, the problem is it’s always going to still be jagged, right? I think the term people use is the jagged intelligence, even — or spiky intelligence, right? 

So you may even have a capability that’s fantastic at a particular task, but it may not uniformly grow. So what is required is in fact, these systems, whether it is GitHub Agent HQ or the M365 Copilot system. Don’t think of this as a product. Think of it as a system that in some sense smooths out those jagged edges, and really helps the capability…

…If I am in M365 Copilot, I can generate an Excel spreadsheet. The good news is now an Excel spreadsheet does understand Office JS, has the formulas in it. It feels like, wow, it is a great spreadsheet created by a good model. The more interesting thing is I can go into agent mode in Excel and iterate on that model. And yet, it will stay on rail. It won’t go off rail, it will be able to do the iteration. Then I can even give it to the analyst agent, and then it will even make sense of it like a data analyst would of our Excel model. The reason I say all of that is because that’s the type of construction that will be needed even when the model is magical, all powerful. I think we will be in this jagged intelligence phase for a long time. So one of the fundamental things that these — whether it’s GitHub, whether it’s security, whether it’s M365, the 3 main domains we’re in, we feel very, very good about building these as organizing layers for agents to help customers.

And by the way, that’s the same thing that we want to put into Foundry for our third-party customers. So that’s kind of how people will build these multi-agent systems.

Microsoft’s management believes that AI software can grow the overall revenue-pie for Microsoft, in a similar manner as how cloud computing expanded the overall server market

I should also say one of the things I like about Copilot is, I mean, Copilot ARPU is compared to M365 ARPUs, right? It’s expansive. The same thing that happened between server and cloud like we used to always say, well, is it zero-sum, it turned out that the cloud was so much more expansive to the server market. The same thing is happening in AI because first, you could say, hey, our ARPUs are too low when it comes to M365 or you could say we have the opportunity with AI to be much more expansive. Same thing with tools, right? I mean, tooling — the tools business was not like a leading business, whereas coding business is going to be one of the most expansive AI systems. And so we feel very good about being in that category. 

To deal with customer-concentration risk from OpenAI, in the event OpenAI cannot follow-through on its spending-commitments, management is (1) building fungible data centers that can serve a broad base of customers including Microsoft itself, (2) only selectively building out data centers for OpenAI, and (3) having internal needs for AI infrastructure, such as Copilot; management walked away from building certain capacity for OpenAI (which Oracle won the contract for) because they wanted to avoid customer-concentration, and they did not want to build capacity that was specific to only one company

[Question] We seem to be entering into a new era where the contractual commitments from a small number of AI natives are just incredibly large, not only in absolute terms, but sometimes relative to the size of the companies themselves. For instance, contracts worth hundreds of billions of dollars that are 20x their current revenue scale. Philosophically, how do you evaluate the ability of those companies to follow through on these commitments?

[Answer] It’s great to have the hit first-party apps in the beginning because you can build scale that then if it’s a fungible and that’s where the key is. You don’t want to build for a digital native in — as if you’re just doing hosting for them. You want to build. That’s where — I think some of the decision-making of ours is probably getting better understood. What do we say yes to, what do we say no to. I think there was a lot of confusion, hopefully by now, anyone who switched on would figure this out. And so that’s, I think, one thing we’re doing on the third party. But the 1 — first party is probably where a lot of our leverage comes and it’s not even about one hit app on our first-party even. Our portfolio of stuff which I just walked through in the earlier answer, gives us, again, the confidence that between that mix, we will be able to use our fleet to the maximum. And remember, these assets, especially the data centers and so on are long assets, right? There will be many refresh cycles for any one of these when it comes to the gear. So I feel that once you think about all those dimensions, the concentration risk gets mitigated by being thoughtful about how you really ensure the build is for the broad customer base…

…When you think about concentration risk or delivering to any customer, you have to remember that because we’re talking about this very large flexible fleet that can be used for anyone and for any purpose, 1P, 3P, and including our commercial cloud, by the way, which I should be quite clear on, it is pretty flexible in every regard…

…[Question] There’s talk that another hyperscaler came in and took away the business that was rightfully Microsoft’s. I’m sure that there is a different point of view here. I’m wondering if you could offer some perspective.

[Answer] Just always goes back to, I think, the core principle, which is build a fleet that is fungible across the planet and works for third-party and first-party and research. So that’s essentially what we have done. And so when some demand comes in shape, that don’t fit that goal, where it’s too concentrated, not just by customer, by location, by type of skewing, right? I think Amy mentioned some very key things. When you think about the margin profile of a hyperscaler, you’ve got to remember this, the AI accelerator piece, but there’s compute, there’s storage. And so if all of the demand just comes for just one [ meter ] that’s really not a long-term business we want to be in. That’s even from a third party. We have to balance it with all of our first-party stuff because that’s after all a different margin stack for us. And then we have to fund our own R&D and model capability because in the long run, that’s what’s going to differentiate us. And so I look at all of those. We sort of use all of that to make sure we are saying yes to all the demand that we want, we say no to some of the demand that may be something that we could serve, but it’s not in our long-term interest. And so that’s sort of the decision-making we have done, and we feel very, very good about the decisions. In some sense, I feel even each time we say no to, the day after, I feel better.

Netflix (NASDAQ: NFLX)

Netflix has been using ML (machine learning) and AI (artificial intelligence) for years to recommend titles to viewers; management thinks Netflix’s data, products, and business processes, gives the company a great position to leverage AI; Netflix is beta-testing a conversational search experience for titles that is powered by GenAI (generative AI); Netflix is using GenAI to localise promotional assets; Netflix productions are starting to use GenAI tools when creating content; management has created guidelines for content producers when using AI tool; Netflix is using AI to test new ad formats

For many years now, ML and AI have been powering our title recommendations as well as production and promotion technology. Given our significant data assets and at-scale products and business processes, we are very well positioned to effectively leverage ongoing advances in AI…

…We’re leveraging GenAI to further enhance the member experience by improving the quality of our recommendations and content discovery features. One example is our beta testing of a conversational search experience that allows members to use natural language to explore the catalog and discover the perfect title for that moment. Another is the way we’re using GenAI to localize promotional assets in a variety of languages so titles can more easily travel to audiences who will love them around the globe…

…For example, in Happy Gilmore 2, filmmakers used GenAI coupled with ML and Eyeline’s proprietary volumetric capture technologies to de-age characters 6 during the opening flashback scene. And the producers of Billionaires’ Bunker used various GenAI tools during pre-production, including for pre-visualization to explore wardrobe and set designs. To help our creative partners use these new technologies responsibly, we recently released production guidance for creators…

…In Q4, we are using AI to test new ad formats, to generate the most relevant ad creative and placement for members, and for faster development of media plans. With these advancements, we’ll be able to test, iterate, and innovate on dozens of ad formats by 2026. 

Netflix’s management thinks that video-generating AI apps such as Sora will mostly impact UGC (user-generated content) platforms in the near term; management thinks AI will mostly help great story-tellers better tell their stories, but it will not make lousy story-tellers great, just like how listeners still gravitate largely towards human-created music rather than AI-created music

[Question] What are your thoughts on the impact from Sora 2 and other new AI content creation apps in terms of increased competition from short-form video, do you think it creates new competition from an engagement standpoint?

[Answer] What we’ve seen so far from these content creation apps is that it’s likely to have a lot more impact on UGC creators the most in the near term. In other words, AI content replacing viewing of existing user-generated content, that starts to make sense. Before we do, it takes a great artist to make something great. Writing and making shows and films well is a rare commodity, and it’s only done successfully by very few people. So AI can give creatives better tools to enhance their overall TV movie experience for our members. But it doesn’t automatically make you a great storyteller if you’re not. So if music is a leading indicator of all this, AI-generated music has been around for a long time, and there’s a lot of it. And it’s a pretty small part of total listening and established artists like Taylor Swift continue to be more popular than ever. So even in a world filled with AI music, AI seems to be mostly a tool for musicians to take — to make — to take their sound in new directions. And so we’re confident that AI is going to help us and help our creative partners tell stories better, faster in new ways, we’re all in on that. But we’re not chasing novelty for novelty sake here, and we’re investing in what we believe delivers value for creators and members alike. So we’re not worried about AI replacing creativity, but we’re very excited about AI creating tools to help creativity.

PayPal (NASDAQ: PYPL)

PayPal has partnerships with Perplexity, Google, and OpenAI for agentic commerce; PayPal has its own agentic commerce service where they can access consumers through multiple LLMs (large language models) with one integration; management thinks agentic commerce will take time but that consumer behaviour will shift; the presence of agentic commerce has not changed any of PayPal’s priorities; management thinks PayPal is well positioned to win in payments for agentic commerce from the merchant perspective (with the agentic commerce service), the consumer perspective (with the largest wallet ecosystems), and the LLM perspective (it would take a long time for LLMs to build the merchant ecosystem that PayPal has already built); some investment from PayPal would be needed for the agentic commerce partnerships

We continue to partner with leaders across the agentic space, including Perplexity earlier this year. And in September, we announced our expansive multiyear partnership with Google to create new AI shopping experiences. This morning, we announced a significant partnership with OpenAI to expand payments and commerce in ChatGPT, including adding PayPal branded checkout for shoppers and payment processing for merchants using Instant Checkout. This is a big win for PayPal and our customers. Today, we also announced our own agentic commerce services, which help merchants sell through multiple AI platforms, including Google, OpenAI and Perplexity. Merchants will have one integration to access consumers through multiple LLMs. Agentic commerce will take time, but we do believe consumer behavior will shift. PayPal is building for that future…

…Our strategy we’ve laid out very clearly is that we want PayPal to be available anywhere and everywhere that consumers want to pay. And we want merchants to be able to sell to consumers anywhere and everywhere. And we’ve talked about this even back at Investor Day where we laid out we want it to be online. We want it to be in-person and we want it to be agentic. And so agentic is just an evolution of this strategy…

…We actually think we’re extremely well positioned to win here. Let me just lay out a couple of the different components. So first, on the merchant side, merchants are going to need to figure out how to integrate with each of these LLMs. And that’s hard because there’s multiple LLMs that are out there. And whether you’re a large enterprise or a small business, you really don’t have the bandwidth to go figure out how to integrate with each and every one of these LLMs, make your catalog available, understand the identity and fraud protection that comes with each of these different elements. And so what we announced today was our PayPal agentic commerce services… We give them seller protection. We give them the ability to scale across all the different LLMs.

From the consumer standpoint, we’re, again, very well positioned. We’ve got the largest wallet ecosystems that are out there and our ability to give consumers the trust, the safety, the buyer protection and the ability to get access and make purchases on any of the LLMs they want to is a huge win. They get to use the wallet that they know and love and have a great end-to-end experience, which includes not only the purchase through the LLM, but also then all the things that happen afterwards, whether it’s package tracking or customer service or returns. So that’s again, a big win for consumers…

…For the LLMs themselves, it would take over a decade if they wanted to go and try to build the same kind of merchant ecosystem of the head, the torso and tail of merchants that PayPal has established over the last couple of decades. And so instead, they get to partner once with us and get access to tens of millions of merchants with identity, authentication, fraud protection and payment processing on a global scale…

…These partnerships do entail some level of investment, whether that’s in product and tech or around co-marketing, things that really drive usage and habituation around the product. And I mentioned in my prepared remarks that we would be reinvesting — begin reinvesting some of our margin dollars in the fourth quarter to really amplify some of our product initiatives. And between the push into agentic and that some of those investments are likely to be a near-term headwind to how fast TM dollars or earnings grow next year.

Taiwan Semiconductor Manufacturing Company (NYSE: TSM)

Demand from AI continues to be very strong and management wants to invest to support TSMC’s customers’ growth; management now expects capex for 2025 to be US$40 billion to US$42 billion, slightly higher than previous expectation for US$38 billion to US$42 billion (2024’s capex was US$29.8 billion); most of the capex for 2025 will be for advanced process technologies; TSMC’s capital expenditure is always in anticipation of growth in future years

As the structural AI-related demand continues to be very strong, we continue to invest to support our customers’ growth. We are narrowing the range of our 2025 CapEx to be between USD 40 billion and USD 42 billion as compared to USD 38 billion to USD 42 billion previously. About 70% of the capital budget will be allocated for advanced process technologies, about 10% to 20% will be spent for specialty technologies, and about 10% to 20% will be spent for advanced packaging, testing, mask making and others.

At TSMC, a higher level of capital expenditures is always correlated with higher growth opportunities in the following years…

TSMC’s management thinks recent developments in the AI market are very positive; management sees explosive growth in token volume, and they think this shows increasing consumer AI model adoption and thus more leading-edge silicon demand; TSMC is using AI internally to improve productivity, and management thinks enterprise AI is another source of demand; management is seeing the emergence of sovereign demand for AI; management has received very strong demand signals from TSMC’s customers and the customers’ customers; management’s conviction in the AI megatrend is strengthening

Recent developments in AI market continue to be very positive. The explosive growth in token volume demonstrated increasing consumer AI model adoption which means more and more computation is needed, leading to more leading-edge silicon demand. Companies such as TSMC, we are leveraging AI internally to drive greater productivity and efficiency to create more value. As such, enterprise AI is another source of demand. In addition, we continue to observe the rising emergence of sovereign AI. We are also happy to see continued strong outlook from our customers. In addition, we directly received very strong signals from our customers’ customers, requesting the capacity to support their business. Thus, our conviction in the AI megatrend is strengthening, then we believe the demand for semiconductor will continue to be very fundamental.

TSMC’s management is disciplined when planning for capacity; TSMC’s lead-time has now increased to 2-3 years because of heightened complexity in process technologies; management thinks TSMC has the deepest and widest look at demand in the semiconductor industry; when planning for AI capacity, management is talking to TSMC’s customers’ customers, which is different from past capacity-planning exercises for other platforms such as smartphones and PCs, where TSMC would talk to only its customers

In order to raise a structural increase in the long-term market demand profile, TSMC employs a disciplined [ in the ] capacity planning system. Externally, we work closely with our customers and our customers’ customer to plan our capacity. We have more than 500 different customers across all the market segments. In addition, as process technology complexity increases, the engagement lead time with customer is now at least 2 to 3 years in advance. Therefore, we probably get the deepest and widest look possible in the industry…

…[Question] Now cloud AI is granting a lot faster than the prior opportunities like smartphones and PCs. Yes, I think the demand for cloud AI is also may be harder to forecast. So just wanted to maybe get a bit more color from you that now to the prior rounds of capacity expansions, what is TSMC doing differently versus before?

[Answer] I believe we are just in the early stage of the AI application. So very hard to make the right forecast at this moment. What do we do differently? There’s a big difference because right now, we pay a lot of attention to our customers’ customer. We talk to and then discuss with them and look at their applications, maybe in the search engine or in social media application. We talk with them and see how they view the AI application to those functions. And then we make a judgment about what AI going to grow. And so this is quite the difference. As compared with before, we only talk to our customers and have an internal study. This is different.

TSMC’s A16 process technology is best suited for specific HPC (high-performance computing) products, which means it is best suited for AI-related workloads; A16 is scheduled for volume production in 2026 H2

We also introduced A16 feature in our best-in-class super power rail, or SPR. A16 is best suited for specific HPC product with compressed signal route and dense power delivery networks.

TSMC’s management now sees the possibility of the revenue CAGR from AI accelerators in the five years ending 2029 to be higher than previous guidance of mid-40s percent because demand is “insane”

[Question] I think we gave a guidance of mid-40s data center AI growth CAGR earlier this year until 2029. Anything that you see which should kind of change that number?

[Answer] The demand actually continue to be very strong in a more — more stronger than we saw the 3 months ago, okay? So in today’s situation, we have talked to customers and then we talk to customers’ customer. So the CAGR previously we announced is about mid-40s, but it is still it’s a little bit better than that. We will update you probably in beginning of next year. So we have a more clear picture. Today, the number are insane.

TSMC’s management continues to see very strong demand for CoWoS (chip on wafer on substrate), driven by AI; management is working hard to narrow the gap between supply and demand for CoWoS; advanced packaging is already close to 10% of TSMC’s revenue

Talking about the CoWoS capacity, all I can say is continue the 3 months ago, we are working very hard to narrow the gap between the demand and supply. We are still working to increase the capacity in 2026. The real number, we probably update you next year. Today, all I want to say about the AI everything related, frontend and backend capacity is very tight. We are working very hard to make sure that the gap will be narrow, but what I can say is we are working very hard…

…Advanced packaging revenue is approaching close to 10% and is significant in our revenue, and it’s important for our customer.

TSMC’s management thinks AI’s growth will still be very positive for TSMC even without access to the China market

I have confidence on my customers, both in graphic or in ASIC, they are all performing well. And so if the China market is not available, but I still think the AI’s growth will be very dramatically and as I said, very positive, and I have confidence that our customers’ performance, and they will continue to grow, and we will support them…

…[Question] So even with immediate obscurity from China for the time being you are still confident that a 14% CAGR or even higher can be achieved in the coming years?

[Answer] You are right.

The amount of TSMC’s wafer-content in a 1 gigawatt AI center differs according to each project

When customers say that 1 gigawatt, they need about — invest about $50 billion, how much of TSMC’s wafer inside? We are not ready to share with you yet because it’s different from different projects… 

…I just want to say that right now, it’s not only 1 chip. Actually, it’s many chip together to form a system, right?

It makes no difference to TSMC’s revenue and gross margin whether it’s helping its customers manufacture GPUs or ASICs (application specific integrated circuits) for AI

[Question] From a TSMC angle, does it matter whether it’s — that demand is coming through a GPU or an ASIC? Does it have an impact on your revenue or gross margin mix?

[Answer] whether with it’s GPU or it’s an ASIC, it’s all using that our leading-edge technologies. And from our perspective, we are working with our customers, and we all know that they are going to grow strongly in the next several years. So no differentiation in front of TSMC. We support all kinds of types.

Tesla (NASDAQ: TSLA)

Tesla’s management thinks Tesla is the leader in real-world AI; management thinks Tesla vehicles have the highest intelligence density of any car; there’s no other company apart from Tesla that is designing AI chips as well as vehicles

I think it’s important to emphasize that Tesla really is the leader in real-world AI. No one can do what we can do with real-world AI. I have pretty good insight into AI in general. I think that Tesla has the highest intelligence density of any AI out there in the car, and that is only going to get better…

…I don’t think there really isn’t anyone that’s doing this — the entire stack all the way through real world — kind of calibrating against the real world where you’ve got cars and robots in real world that like we know what the chip needs to do, and we know what — just as importantly, we know what the chip doesn’t need to do…

…Obviously, you can do reasoning on the server, that takes whatever. But then in a car, you need to make real-time decisions. So putting all that into the computer that’s in the car, that’s the challenge…

…I’m confident in saying that Tesla has — Tesla AI has the highest intelligence density. When you look at the intelligence per gigabyte, I think like Tesla AI is probably, in order of magnitude better than anyone else. And it doesn’t have any choice because that AI has got to fit in the AI for computer.

Millions of existing Tesla vehicles can become fully autonomous with a software update; management now has clarity on achieving full autonomy; version 14 (v14) of Tesla’s FSD software is broadly available now and current users have been amazed by it; Tesla’s Robotaxi service is now operating in 2 markets; Robotaxi’s coverage area in Austin has expanded by 3x since the initial launch; management thinks Tesla’s Robotaxi fleet bleeds in with other vehicles, unlike those of its competitors with many extra sensors; management thinks that demand for Tesla vehicles will expand significantly as people experience FSD at scale; FSD’s adoption is making decent progress, with the total paid FSD customer base being 12% of the current flee; Tesla groups Robotaxi’s costs within the Services and Other revenue line; management expects to have no safety drivers for Robotaxi in large parts of Austin by end-2025, even when operating with an abundance of caution; management expects Robotaxi to be in 8-10 metro areas by end-2025; management expects Robotaxi to be in Nevada and Florida and Arizona by end-2025; Robotaxis in Austin without anyone in the driver seat have covered more than 0.25 million miles; Robotaxis in Bay Area have crossed more than 1 million miles; customers are happy with Robotaxi and there are no notable issues; total miles driven by supervised FSD has crossed 6 billion and the overall safety remains excellent; Tesla will be working on a V14-light version FSD software that is compatible with Hardware 3; a big reason why autonomy is safer than human driving is because a large part of human driving accidents are caused by texting-while-driving; the autonomous driving software shipped to customers and Robotaxi are very similar; updated editions of V14 FSD will have reasoning capabilities

We have millions of cars out there that with a software update become full self-driving cars…

…We see now as a clarity on achieving full self-driving, unsupervised full self-driving…

…With version 14 of the — of self-driving, which people — you can see the reactions of people online. They’re quite amazed. Actually, anyone in the U.S. can get version 14 if they just go and select, I want the advanced software in their car. So if you’re listening right now and you’d like to try it out, just go in Settings and say, I want the advanced software, and you will get version 14…

…We’re now operating our Robotaxi in 2 markets, Austin and most Bay Area cities. We’ve already expanded our coverage area in Austin 3x since the initial launch and are on pace to continue expanding further.

Unlike our competitors, our Robotaxi fleet blends in the markets we operate in since they don’t have extra sensor sets or peripherals, which make them stick out. This is an underappreciated aspect of our current vehicle offerings, which are all designed for autonomous driving.

We feel that as experience — as people experience the supervised FSD at scale, the demand for our vehicles, like Elon said, would increase significantly.

On the FSD adoption front, we’ve continued to see decent progress. However, note that total paid FSD customer base is still small, around 12% of our current fleet. We’re moving — we’re working with regulators in places like China and EMEA to obtain approvals so that we can get FSD in those regions as well…

…Note that while small, our Robotaxi costs are included within Services and Other, along with our other businesses like paid supercharging, used car, parts and merchandise sales, et cetera…

…We are expecting to have no safety drivers in at least large parts of Austin by the end of this year. So within a few months, we expect to have no safety drivers at all at least in parts of Austin. We’re obviously being very cautious about the deployment. So our goal is to be actually paranoid about deployment because obviously, even one accident will be front page headline news worldwide. So it’s better for us to take a cautious approach here. But we do expect to have no safety drivers in the car in Austin within a few months. I think that’s perhaps the most important data point.

And then we do expect to be operating Robotaxi in, I think, about 8 to 10 metro areas by the end of the year. It depends on various regulatory approvals…

…We expect to be operating in Nevada and Florida and Arizona by the end of the year…

…We continue to operate our fleet in Austin without anyone in the driver seat, and we have covered more than 0.25 million miles with that. And then in the Bay Area, where we still have a person in the driver seat because of the regulations, we crossed more than 1 million miles. So — and we continue to see that the fleet — Robotaxi fleet works really well. Customers are really happy, and there’s no notable issues…

…Customers have used FSD supervised for a total of 6 billion miles as of yesterday. So that’s like a big milestone. And overall, the safety continues to be very good…

…Once the V14 release series is fully done, we are planning on working on a V14 light version for Hardware 3 probably expected in Q2 next year…

…The reason you’ve seen like there’s been an uptick in accidents pretty much worldwide is because people are texting and driving. So Autopilot actually dramatically improves the safety here because if somebody is looking down their phone, they’re not driving very well. So that’s really the game changer…

…In terms of like what we ship to customers versus Robotaxi, it’s mostly the same. Obviously, customers have some more features like they can choose the car wants to park in a spot or drive something like that, which is not super relevant for Robotaxi. But there’s only a few minor changes like those ones. But the majority of the algorithms and the architecture, everything is the same between those 2 platforms…

…We’ll be adding reasoning to — I don’t know, Ashok, is that like reasoning in like 14.3, maybe 14.4, something like that?… Yes, by end of this year for sure.

Tesla’s management is still very optimistic about the potential of Tesla’s Optimus autonomous robot; Tesla will unveil Optimus V3 in 2026 Q1; most of the real-world AI Tesla has developed for fully autonomous driving can be transferred to Optimus; management thinks Optimus can be a great surgeon; management thinks bringing Optimus to market is incredibly difficult; Optimus robots are already walking around Tesla’s offices; it’s really difficult engineering-wise to create the hands and fingers of Optimus that can mimic human hands and figures; it’s hard to manufacture Optimus at scale because the supply chain currently does not exist, so Tesla has had to be very vertically integrated and manufacture very deep into the supply chain; management thinks Tesla is uniquely positioned to win in autonomous robots because success in autonomous robots depends on 3 things, namely, scaled manufacturing technology, real-world AI, and a dextrous hand, and Tesla is the only company that can achieve all 3; management thinks Optimus can be 5x more productive than humans; many of the people working on Optimus in Tesla now were working on Tesla vehicles in the past; Optimus’s management reviews involve a tight loop between manufacturing and engineering design so that the overall manufacturing processes for Optimus can be good; Optimus 2 was impossible to manufacture; Tesla will have rolling changes for the Optimus design even after start of production

We’re also on the cusp of something really tremendous with Optimus, which I think is likely to be or has potential to be the biggest product of all time…

…We look forward to unveiling Optimus V3 probably in Q1. I think it will be ready for — to show off…

…The real-world intelligence we’ve developed for the car, most of that transfers to Optimus. So it’s a very good starting point…

…Optimus will be an incredible surgeon, for example, I imagine everyone had access to an incredible surgeon…

…Bringing Optimus to market is an incredibly difficult task to be clear…

…We do have Optimus robots that walk around our offices at our engineering headquarters in Palo Alto, California, basically 24 hours a day, 7 days a week. So any visitors that come by, you actually — you can stop one of the Optimus robots and ask it to take you somewhere, and it will literally take you to that meeting room or that location in the building…

…It’s difficult to create a hand that is as dextrous and capable as the human hand, which is an incredible — the human hand is an incredible thing that the more you study the human hand, the more incredible you realize the human hand is and why you need 5 — 4 fingers and a thumb, why the fingers have certain degrees of freedom, why the various muscles are of different strengths, the fingers are of different lengths. And it turns out actually that those are all there for a reason. And so making the hand and forearm, because most of the actuator — just like the human hand, the muscles that control your hand are actually primarily in your forearm. The Optimus hand and forearm is an incredibly difficult engineering challenge. I’d say it’s more difficult than the rest of — from an electromechanical standpoint, the forearm and hand is more difficult than the entire rest of the robot…

…Trying to make 1 million Optimus robots per year, that manufacturing challenge is immense, considering that the supply chain doesn’t exist. So with cars, you’ve got an existing supply chain. With computers, you’ve got an existing supply chain. With a humanoid robot, there is no supply chain. So in order to manufacture that, Tesla actually has to be very vertically integrated and manufacture very deep into the supply chain, manufacture the parts internally because there just is no supply chain…

…If I put myself in the position of a start-up trying to make a humanoid robot, I’m like, I don’t know how to do it without an immense amount of manufacturing technology. So — that’s why I think like Tesla is in almost a unique — I think a unique position when you consider manufacturing technology scaling, real-world AI and a truly dextrous hand. Those are generally the things that are missing when you read about other robots that just don’t have those 3 things. So I think we can achieve all those things — those 3 things with an immense amount of work. And that is the game plan…

…Optimus at scale is the infinite money glitch. It’s like this is — it’s difficult to express the magnitude of — like if you’ve got something like that — like if Optimus, I think, probably achieve 5x the productivity of a person per year because it can operate 24/7, it doesn’t even need to charge. It can operate tethered. So it’s plugged in the whole time…

…4-plus years back, we were in a finance meeting with Elon and Elon said, hey, our car is a robot on wheels. And that’s where we started developing. In fact, most of the engineering team, which is working on Optimus has come from the vehicle side. And that’s why when we talk about manufacturing prowess, we have the wherewithal because the same engineers who worked back in the day on drive units are working on actuators now. So that’s where we can — if there is any company which can do it at scale, that is going to be us…

…The Optimus reviews at this point are there’s the engineering review and then there’s the manufacturing review being done simultaneously with an iterative loop between engineering design and manufacturing because then we see — we design something and we say like, oh man, that’s really difficult to make. We need to change that design to make it easier to manufacture. So we’ve made radical improvements to the design of Optimus while increasing the functionality but making it actually possible to manufacture. 

Like I’d say, Optimus 2 is almost impossible to manufacture, frankly…

…The hardware design will not actually be frozen even through start of production. There will be continued iteration because a bunch of the things that you discover are very difficult to make. You only find that pretty late in the game. So we’ll be doing rolling changes for the Optimus design even after start of production.

Tesla’s A14 chip is manufactured by Samsung; Tesla is going to manufacture its A15 chip with both TSMC and Samsung; the A15 chip has 40x better performance than A14 because Tesla designed the hardware to address all the pain points in software; the A15 chip deleted a lot of components that were in the A14 chip and this has greatly improved the performance of the A15 chip; management thinks Samsung’s US fab is slightly more advanced than TSMC’s US fab; management wants to have an oversupply of A15 chips because the chips that do not go into vehicles and Optimus can be used for Tesla’s data centers; Tesla uses a combination of its A-series chips and NVIDIA chips for AI training; Tesla is not looking to replace NVIDIA, but management notes that NVIDIA’s chips need to accommodate a wide range of use cases, which disadvantages it against Tesla’s self-designed chips which need to accommodate only Tesla’s use cases; management thinks Tesla’s A15 chip will have the best performance per watt and best performance per dollar for AI

Samsung is worth noting, does manufacture our AI4 computer and does a great job doing that. So now with the AI5, and here’s I need to make a point of clarification relative to some comments I’ve made publicly before, which is we’re actually going to focus both TSMC and Samsung initially on AI5…

…By some metrics, the AI5 chip will be 40x better than the AI4 chip, not 40%, 40x because we have a detailed understanding of the entire software and hardware stack. So we’re designing the hardware to address all of the pain points in software…

…With the AI5, we deleted the legacy GPU or the traditional GPU, which is — it’s in AI4. But AI5 does not have — we just deleted the legacy GPU because it basically is a GPU. So we also deleted the image signal processor. And there’s like a long list actually of deletions that are very important. As a result of these deletions, we can actually fit AI5 in a half reticle and with good margin for the traces from the memory to the Tesla Trip accelerators, the ARM CPU cores and the PCI-X sort of the PCI blocks. So this is a beautiful chip. I’ve hoarded so much life energy into this chip personally. And I’m confident this will be — this is going to be a winner next level…

…Technically, the Samsung fab has slightly more advanced equipment than the TSMC fab. These will both be made in the U.S., one — TSMC in Arizona, Samsung in Texas…

…Our goal — explicit goal is to have an oversupply of AI5 chips because if we have too many AI5 chips for the cars and robots, we can always put them in the data center…

…We already use AI for training in our data centers. So we use a combination of AI5 and NVIDIA hardware. So we’re not about to replace NVIDIA to be clear, but we do use both in combination, AI4 and NVIDIA hardware. And the AI5 excess production, we can always put in our data centers…

…The challenge that they have is that they’ve got to satisfy a large range — a lot of requirements from a lot of customers, but Tesla only has to satisfy requirements from one customer, that’s Tesla. That makes the design job radically easier and means we can delete a lot of complexity from the chip. Like I can’t emphasize how important this is. So like when you look at the various logic blocks in the chip, as you increase the number of logic blocks, you also increase the interconnections between the logic blocks. So you can think of it like there’s highways, like how many highways do you need to connect the various parts of the chip. And especially if you’re not sure how much data is going to go between each logic block on the chip, then you kind of end up having giant highways going all over the place. It’s a very — like it becomes an almost impossibly difficult design problem. And NVIDIA has done an amazing job of dealing with almost an impossibly difficult set of requirements. But in our case, we’re going for radical simplicity…

…I think AI5 will be the best performance per watt, maybe by a factor of 2 or 3 and the best performance per dollar for AI, maybe by a factor of 10.

Tesla has a world simulator for reinforcement learning for autonomous driving that is indistinguishable from actual video; Tesla will be increasing the parameter count for its autonomous driving AI model by an order of magnitude

Our world simulator for reinforcement learning is pretty incredible, like — when you see the Tesla Reality Simulator, it’s — you can’t tell a difference between the video that’s generated by the Tesla Reality Simulator and the actual video, it looks exactly the same. So that allows us to have a very powerful reinforcement learning loop to further improve the Tesla AI.

We’re going to be increasing the parameter count by an order of magnitude. That’s not in 14.1. There are also a number of other improvements to the AI just that are quite radical. So it’s — this car will feel like it is a living creature. That’s how good the AI will get with the AI4 computer just before AI5.

Tesla’s management thinks Tesla vehicles can become a giant distributed AI inference fleet

We could actually have a giant distributed inference fleet and say like, well, if they’re not actively driving, let’s just have a giant distributed inference fleet. At some point, if you’ve got like tens of millions of cars in the fleet or maybe at some point, 100 million cars in the fleet, and let’s say they had at that point, I don’t know, a kilowatt of inference capability of high-performance inference capability, that’s 100 gigawatts of inference distributed with power and cooling — with cooling and power conversion taken care of. So that seems like a pretty significant asset.

The AI models Tesla and xAI are developing are very different, with Tesla’s models being much smaller

So the xAI, Grok is like a giant model that you could not possibly squeeze Grok onto a car. That’s for sure. It is a giant beast of a model. It’s — with Grok is trying to solve for artificial general intelligence with a massive amount of AI training compute and inference compute. So for example, Grok 5 will actually only run effectively on a GB300. That’s how much of a beast that Grok 5 is. So — whereas Tesla’s models are, I don’t know, maybe about less than 10% of the size, maybe closer to 5% the size of Grok. So yes, they’re really coming at the problem from very different angles. xAI and Grok are — they’re competing with Google Gemini and OpenAI ChatGPT and that kind of thing. So — and some of it is complementary. I mean for example, for Grok voice, being able to interact with Grok in the car is cool. Grok — for Optimus voice recognition and voice generation is Grok. So that’s helpful there. But they are coming at it from kind of opposite ends of the spectrum.

Visa (NASDAQ: V)

Visa has begun deployment of the next generation of VisaNet, its core processing platform; more than half of the new code base for VisaNet was written with the help of generative AI

We have begun deployment of the next generation of VisaNet, the core processing platform in our Visa as a Service stack. It offers a cloud-ready micro services distributed modular architecture that uses open languages and technologies, enabling easier scaling, configuration and faster feature deployment. Over half of the new code base was built with the assistance of generative AI, improving development speed, security and maintainability. We have specific modules in market today with plans to roll out additional modules and markets.

The Visa Scam Disruption product detects scam activity at the network level and uses AI to monitor merchants; Visa Scam Disruption has been launched for only a year, but it has helped law enforcers to dismantle more than $1 billion in fraud attempts from 25,000 scam merchants

We continue to enhance our risk management capabilities, including Visa scam disruption, which proactively detects scam activity at the network level that no single issuer, acquirer or a merchant could see alone and leverages AI-enhanced merchant monitoring, external intelligence feeds and our global expertise. Just a year since launch, we have worked closely with our clients and law enforcement to dismantle more than 25,000 scam merchants representing more than $1 billion in fraud attempts.

Visa is now powering live agentic transactions; Visa recently released the Visa Trusted Agent Protocol to help merchants verify agents and avoid malicious bots; there is minimal integration required from merchants to utilise Visa Trusted Agent Protocol; Visa recently launched its MCP (model context protocol) server, which allows AI systems to interface with the Visa Intelligent Commerce APIs; management thinks Visa is leading in setting standards for agentic commerce; with Visa Intelligent Commerce, management has put out a set of capabilities for AI-ready cards to allow consumers to easily set spending limits and conditions for agentic transactions; the Visa Trusted Agent protocol is an open standard; management thinks agentic commerce will accelerate adoption of traditional e-commerce and mobile commerce, and be a net positive for Visa in both the transactions-driven business, and the value-added services business; management thinks there will be 3 phases to agentic commerce, (1) consumers using agents for discovery then making purchases on merchant sites, (2) consumers using agents for discovery and making purchases with the agents, and (3) consumers empowering agents to search for things on their behalf and buy; management thinks Visa Trusted Agent Protocol can be the base layer for everyone in the agentic commerce ecosystem to leverage on ;see Point 32 for more on agentic commerce; 

I’m pleased to announce that we are now powering live agentic transactions and recently released a merchant agent toolkit to make it easy for developers to embed our solutions into workflows and agentic processes. Just 2 weeks ago, we announced the Visa Trusted Agent Protocol, a framework that enables safer agent-driven checkout by helping merchants verify agents and avoid malicious bots. And since it’s built on existing messaging standards, minimal integration is required for merchants…

…We recently launched our MCP server, providing access for AI systems to interface with our Visa Intelligent Commerce APIs…

…In this third wave of agentic commerce, we’ve been leading in terms of our role of setting the standards. I think one great example of that is Visa Intelligent Commerce, where we put out a set of capabilities for AI-ready cards, leveraging tokenization, AI-powered personalization, leveraging our data token service. We put out a set of standards with payment instructions that are going to allow customers like you and I to easily set spending limits and conditions to provide clear guidance for agent transactions and also our payment signals, which are going to share those data payloads in real time with Visa, enabling us to help set transaction controls, manage disputes and Chargebacks and those types of things…

…I think what differentiates the Visa Trusted Agent Protocol is 2 things. One is it’s open. It’s an open set of standards, and we think that an open framework is critical to drive mass adoption in the way that’s needed for agentic commerce. And the second is it’s easy to integrate. We built it on existing web infrastructure so that it’s going to be easy for merchants to integrate into existing messaging standards and get up and running quickly…

…[Question] What extent you see agentic as more of a substitute for traditional e-commerce versus being additive to the TAM of the overall payments industry.

[Answer] I think the base case is it continues to accelerate the adoption of e-commerce and mobile commerce as we all know it. I think there’s an upside case on that where you could actually see users buying from a much larger and more diverse set of merchants than they do today in traditional e-commerce given the power of these agents and their ability to go out and search the world’s inventory based on whatever it is that you prefer for your agent. That might be value. That might be price. That might be inventory. That might be speed of delivery and so on and so forth. I think that could ultimately result in consumers buying more things from more merchants, which ultimately means more transactions on Visa. I also think there’s a significant upside in the delivery and the relevance of our portfolio of value-added services for the entire ecosystem, especially as you said, they have to work through a number of things that involve potential fraud and disputes and chargebacks and things like that…

…It’s still early days. And I think what you’re likely to see in the evolution of agentic commerce is not different or dissimilar to what we saw in e-commerce. I think early on, you’re seeing consumers use these agents and these platforms for discovery. They’re shopping. They’re looking for what might be available for any given gift I’m trying to buy or any clothing item that I might try to buy. But then I might jump to the actual merchant site to make the purchase.

Then the next step of what you’re starting to see is the integration of the buy capabilities into that shopping journey. We’re just starting to see that in the marketplace today. We’ve been working on that for many, many months with the ecosystem.

And then I think the ultimate kind of user experience and the promise of agentic commerce will be truly empowering agents to go out to search for things on our behalf and ultimately make purchases and buy things without human intervention. That, we haven’t really seen in the marketplace today, but we’re working very hard with the platform players to ensure that the capabilities are in place to enable that…

…I think it’s where the Visa Trusted Agent Protocol can form a base layer for everyone to build on and everyone to ultimately leverage.

Visa Protect for A2A (Account-to-Account), which enables consumers to pay businesses directly from their bank accounts, is using AI to reduce fraud in Brazil; Visa Protect for A2A’s pilot in Brazil scored nearly $500 billion of Pix volume from Visa’s bank partner over a 6-month period and identified over $90 million of fraud; the fraud could have been prevented with a detection rate of more than 80% with Visa Protect for A2A

Our award-winning product, Visa Protect for A2A, is delivering value with AI. Our pilot in Brazil scored nearly $500 billion of our bank partner’s Pix volume over a 6-month period and identified over $90 million of fraud, which could have been prevented with a detection rate of more than 80%. We believe Visa Protect for A2A can play an important role in Brazil by providing real-time fraud monitoring on Pix, helping to reduce fraud for our bank partners and ensure a safer payment experience for buyers and sellers.

Visa’s management thinks tokenisation is the critical building block for agentic commerce

Tokenization, I think, is the critical building block that ultimately will help Agentic commerce reach its promise. And if you go back — I know you asked about the Trusted Agent Protocol, but if you go back to the Visa Intelligent Commerce set of products and standards that we put out, tokenization as a platform is what enables the bulk of that functionality and ultimately is what’s going to enable us all to have safe, secure, trusted transactions with agents on our behalf. So tokenization, critical building block of that.

---

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, MercadoLibre, Meta Platforms, Microsoft, Netflix, PayPal, TSMC, Tesla, and Visa. Holdings are subject to change at any time.