Last month, I published More Of The Latest Thoughts From American Technology Companies On AI (2026 Q1). In it, I shared commentary in earnings conference calls for the first quarter of 2026, from the leaders of technology companies that I follow or have a vested interest in, on the topic of AI and how the technology could impact their industry and the business world writ large. 

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

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

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

MongoDB (NASDAQ: MDB)

MongoDB’s management sees 2 dimensions to the growth opportunity ahead, namely (1) organisations running core workloads on MongoDB, and (2) organisations moving agentic applications into production and choosing MongoDB as the core database; the 2 dimensions reinforce each other, as agentic applications are built on the data already residing within MongoDB

These conversations reinforce my conviction in both what we have built and the scale of the opportunity ahead. That opportunity has 2 dimensions. The first is core workloads where large customers run their most demanding, mission-critical workloads on MongoDB across on-prem, public clouds and hybrid environments. The second is AI, where enterprises, digital natives, frontier labs and AI natives alike are moving agentic applications into production and choosing MongoDB as the data platform to power them. As you heard from other software companies, these 2 opportunities are not distinct and, in fact, reinforce each other. Enterprises are starting to build agentic application on top of the very data already running on MongoDB.

MongoDB’s management is seeing accelerating AI adoption across the company’s base, with MCP (Model Context Protocol) server usage growing significantly; Voyage customers have doubled sequentially in 2026 Q1 (FY2027 Q1); Vector Search adoption is far outpacing MongoDB’s overall growth; Voyage AI embeddings entered public review in 2026 Q1 (FY2027 Q1) and it allows developers to deliver semantic search in minutes; MongoDB has delivered 10-plus integrations with LangChain for Vector Search and more

AI adoption of MongoDB technologies across our customer base continues to accelerate. MCP server usage is growing significantly. Voyage customers have more than doubled quarter-over-quarter and Vector Search adoption is far outpacing overall company growth…

…This quarter, automated Voyage AI embeddings entered public preview, removing weeks of infrastructure work and enabling developers to deliver semantic search in minutes…

…LangChain is the world’s most widely adopted agent framework with over 1 billion downloads. We delivered 10-plus native integrations with LangChain for Vector Search, hybrid retrieval, semantic caching and agent memory.

Several frontier AI labs have selected MongoDB for mission-critical use cases; it’s still early days for MongoDB regarding the frontier AI labs’ workloads, but management is optimistic about expanding with the labs over time; AI-native companies are choosing MongoDB as the foundation for their data layer and the right choice for the data layer is important because it is a chokepoint on rapid scaling; it’s still early, but MongoDB’s management is starting to see enterprises shift from experimenting with AI to deploying AI in production; customers are choosing MongoDB as the memory layer for AI agents; MongoDB’s current results are driven by core workloads, but management is seeing growing moment from AI and agentic workloads, and MongoDB is ready for agentic deployment at scale whenever it happens; management is seeing the frontier AI labs realising that MongoDB is a great data platform, after trying out alternatives such as Postgress; the frontier AI labs are using MongoDB for multiple use cases

Turning to AI. This opportunity spans 3 distinct segments. First is the frontier labs. Several of these have selected MongoDB for use cases that are mission-critical to the deployment of their products among the most demanding data workloads in the industry. The depth of engagement varies by lab and by workload, and it is still early. But we feel great about the use cases we are winning and the ability to expand within these customers over time.

Second is AI-native companies. These customers are choosing MongoDB as the foundation for their AI products from day 1 because the data layer determines if you can scale to support rapid growth…

…Third is enterprise deploying AI. It is still early here, but we are beginning to see customers move from experimentation into production, building AI application on top of the operational data layer already running their business…

…Customers choosing MongoDB as the memory layer for AI agents themselves, agentic workloads need memory, that’s transactional, high velocity and able to retrieve the right context at the right time…

…Our results today are driven primarily by core workloads, but we are seeing real and growing momentum from AI and agentic workloads and believe MongoDB is purpose-built to be generational data platform for the agentic era…

…I’m seeing it’s still early, Matt, just to be clear, because the security governance, observability, there are many, many aspects to the agents and what kind of outcomes they deliver if it is agents at scale. But we feel that we are ready…

…[Question] In your prepared remarks, you mentioned frontier labs and it sounded like it was labs plural. I know you choose your words very carefully in the prepared remarks. I guess, did I pick that up correctly, that Mongo might now be working with multiple frontier labs?

[Answer] Yes, it is plural, and it was chosen carefully. Thank you for noticing… As we work with them, and as they have tried, whether it’s a Postgres alternative or others, they have come to realize that. And these are truly at the forefront of innovation in AI space or driving innovation that MongoDB is just a great data platform for some of the workloads. And the point around — of course, we cannot go into specific details with our agreements with them on type of use cases, but they vary and there are multiple use cases depending on the lab, that we’re working with them, and it’s early, but we will continue to expand.

MongoDB’s management sees the company as the generational data platform for agentic AI for 5 reasons, namely (1) MongoDB is architecturally built for AI because rigid relational data schemas are not suitable for agentic coding and LLMs , whereas they play well with unstructured document databases, (2) MongoDB is a high-performance data platform that allow agents to read and write in real time, (3) MongoDB delivers retrieval accuracy that agents require for customer-facing applications, (4) MongoDB can run on-premises, on the cloud, on a hybrid format and (5) MongoDB is embedded in the tools that developers and agents are using; management sees 3 legs of the stool for an agentic workload, namely, the harness, the LLM (large language model), and the data layer; customers of MongoDB appreciate the integration with LangChain because this means the data layer works really well with the harness layer; MongoDB’s database was not designed with AI workloads in mind, but it turns out that the architecture is perfectly suited for AI workloads

We are seeing real and growing momentum from AI and agentic workloads and believe MongoDB is purpose-built to be a generational data platform for the agentic era. Built natively into the platform, MongoDB’s innovations in the core database, embeddings and vector capabilities are moving us beyond a system of record to becoming the real-time system of intelligence. That shift comes down to 5 core strengths.

Number one, MongoDB is architecturally built for AI in 2 key ways. First, our flexible schema is uniquely suited to how applications get built in the agentic era. A growing share of software is now created through prompt-driven development, natural language iteration rather than line-by-line authorship. Whether the prompt comes from a developer or an agent, the shape of the application shifts with each prompt and a rigid relational schema becomes a tax on every iteration compromising agility. In addition, LLMs are the lingua franca for AI, and they speak in unstructured documented shape data, the exact form MongoDB was built around…

…Second, MongoDB is a transactional, high-performance data platform built for how agents actually work. Agents don’t behave like traditional applications. They read, write and act continuously across multiple simultaneous threads with a single agent spawning subagents that each make independent reads and writes in real time. Analytical systems built for off-line processing weren’t designed for this, and it shows in the performance when you run agents on top of them. MongoDB 8.3 released this month takes that step one further, delivering up to 45% more reads, 35% more writes and 15% more ACID transactions over 8.0 without changing a line of application code.

Third, MongoDB is a data platform that delivers the retrieval accuracy agents need to be trusted while optimizing tokens and cost in production. For internal tools, occasional errors may be tolerable. But for customer-facing application such as clinical decision support, fraud detection, financial transaction, insurance transaction, accuracy is nonnegotiable. MongoDB delivers best-in-class retrieval through integrated Vector Search and Voyage embeddings and reranker models, purpose built to surface the most relevant context when agent needs it…

…Fourth, MongoDB runs wherever the agent needs to run across all 3 major clouds, on-prem and in hybrid environments…

…Fifth, MongoDB is embedded in the tools, developers and agents actually use to build agentic applications…

…The simplicity when we talk to customers is 3 legs of the stool for any agentic workload is harness, LLM and data layer. And if they are being used as in LangChain, they have significant traction. Even when I talk to some of the large banks, whether it’s on-prem or in the cloud, there’s significant traction on the harness layer. And then they say, okay, what about the data layer and data layer, MongoDB being a choice for the data layer just makes sense. So we have done many integrations with them, and we are seeing this being played out at some of the large enterprise customers who say, hey, CJ, I’m glad that the data layer as in MongoDB really works with the harness layer. And of course, we can choose whichever LLM we want…

…I would say that architecture, it is almost — our founder calls it really well that. We would rather be lucky than smart. And when we created MongoDB — this is from Dwight. We didn’t have AI workloads in mind, but this architecture is perfectly suited for AI workloads.

MongoDB’s management recently announced MongoDB Checkpointer for LangSmith; MongoDB Checkpointer for LangSmith collapses a dedicated Postgres instance per agent into a single, shared Atlas cluster; the MongoDB Plugin and agent skills on Claude Code’s marketplace was recently launched

We recently announced that MongoDB Checkpointer for LangSmith deployment, which collapses what used to be a dedicated Postgres instance per agent into a single, shared Atlas cluster, state, memory and operational data unified in one place. Last month, we also launched the MongoDB Plugin and agent skills on the Claude Code marketplace, where we are already seeing strong early traction with developers.

Endor Labs, an AI-native application security platform, chose MongoDB Atlas as its default database; Endor Labs is using Atlas and Atlas Search for mission-critical security workflows; MongoDB Atlas is lowering Endor Labs’ operational friction

For example, Endor Labs is an AI-native application security platform, protecting over 7 million applications across both human written and AI-generated code. Endor selected Atlas as its default database to support 225% year-over-year revenue growth. Endor uses Atlas and Atlas Search to power its mission-critical security workflows, including AURI, its new security intelligence layer for AI coding agents, allowing the company to reduce operational friction and accelerate delivery of its differentiated offerings.

Food delivery company Zomato has 25 million monthly active users; Zomato is using MongoDB Atlas to sell its AI-native customer support platform, Nugget, to other enterprises; Zomato chose MongoDB Atlas over DynamoDB and DocumentDB for its aggregation pipeline, right consistency and flexible schema; MongoDB Atlas has lowered Nugget’s support cost by 55% and raised human agent productivity by 40%

Zomato is a great example. The world’s second largest food delivery company with 25 million monthly active users built Nugget, an AI-native customer support platform, they are now selling to other enterprises on Atlas. After evaluating DynamoDB and DocumentDB, they chose Atlas for its aggregation pipeline, right consistency and flexible schema. Nugget now orchestrates 15 million conversations per month on MongoDB’s platform, reducing support cost by 55% and improving human agent productivity by 40%.

Adobe’s Journey Agent is using MongoDB Atlas for long-term memory; Atlas Search and Atlas Vector Search enables Adobe to achieve sub-100 millisecond hybrid search for Journey Agent to act in real time

Adobe’s Journey Agent is a clear example. A composite multimodal AI agent that unifies Adobe’s marketing suite and orchestrates end-to-end customer journeys for their global B2C user base with MongoDB as the agent’s long-term memory and reasoning layer. Adobe leverages the MongoDB platform, Atlas Search and Atlas Vector Search together to power the sub-100 millisecond hybrid search the agent needs to act in real time.

The growth of AI startup ElevenLabs was being choked by its data layer, and made the decision to move to MongoDB recently; Postgres databases are choking the growth of AI native companies that have adopted it

I shared the example of somebody like ElevenLabs at .local London a few weeks ago, they were using first-party database for operational data. They were using another software for search. And basically, most of those product lines were really choking as ElevenLabs was growing significantly, right? They are now at a $500 million ARR. So when I asked the team technically, the engineer who made that decision saw that the growth of the company as in that AI native company, ElevenLabs was being held up by the data layer. And us having Search, Vector Search and operational data in a single platform, they are — they made the decision to move to MongoDB not too long ago. And 2 things they said that really resonated with me, Ryan. Number one, they are like, gee, we should have done this a lot sooner. Otherwise, we would have not to deal with all these outages and other things they dealt with the previous platform. And number two, now choosing MongoDB even though they have scaled significantly on their ARR as an AI native company gives them peace of mind.

I’m hearing them from other AI native companies who also chose maybe a Postgres or something and Postgres completely choked on the performance. So that just gives me a lot of confidence that if AI native company where AI is the business or agentic layer is the business and they feel that they can scale with MongoDB.

Nu Holdings (NYSE: NU)

Nu Holdings is seeing AI-driven productivity gains, with engineering through up 50% year-on-year in 2026 Q1, weekly token consumption up 10x from the start of 2026 to March, and testing cycles becoming 90% faster; nearly 100% of Nu Holdings’ employees are utilising AI tools

AI is driving productivity gains across the company, with engineering throughput up 50% YoY, weekly token consumption nearly ten times higher than at the start of the year, and testing cycles 90% faster…

…We’re reaching close to 100% utilization of AI tools among our employees across all functions of the organization.

Nu Holdings’ management expects to launch new AI-native experiences to customers in 2026; Nu Holdings’ AI Private Banker functionalities currently have 15 million active users

Customer journeys are being rebuilt end-to-end, with new AI-native experiences expected to reach customers during 2026…

… Nu’s AI Private Banker functionalities — financial insights, payments, credit advice, and debt resolution — are now serving more than 15 million monthly active users. 

Nu Holdings’ proprietary foundation models, NuFormer, is already in production to make lending decisions for credit cards in Brazil and Mexico, and unsecured lending in Brazil; NuFormer can make a decision for each personal loan request in under a second

NuFormer, Nu’s proprietary set of foundation models, is in production today for credit card decisioning in Brazil and Mexico, and for unsecured lending in Brazil, with real-time AI valuation now pricing and approving every personal loan request individually based on its predicted NPV in under a second. 

Nu Holdings’ management sees 3 structural advantages the company has in AI, namely, (1) proprietary data from 135 million customers, (2) a cloud-native technology stack that’s built internally, and (3) a strong talent base

Nu’s AI Transformation is anchored by three structural advantages: first-party data at scale from 135 million transacting customers generating one of the largest and most differentiated financial datasets in the world; a proprietary cloud-native technology stack with core banking systems built internally and data unified across the company; and a world-class talent base of ten thousand employees from more than 50 nationalities across six countries. 

NVIDIA (NASDAQ: NVDA)

NVIDIA’s management capitalised on an inflection in inference demand by ramping its Blackwell systems; NVIDIA’s Data Center revenue again had very strong growth in 2026 Q1 (FY2027 Q1), driven by strong demand for Blackwell systems; the Blackwell systems are the fastest product ramp in NVIDIA’s history; management sees Blackwell systems as having the lowest token generation cost for inference; every hyperscaler, cloud provider, and model maker is using Blackwell; OpenAI’s latest GPT-5.5 model was trained with and is being served by Blackwell systems; Microsoft’s latest largescale AI data center, Fairwater, is powered by Blackwell GPUs; Amazon’s AWS will be adding more than 1 million Blackwell and Rubin (the next generation GPU) GPUs; Google Cloud will be offering Blackwell systems; Blackwell Ultra delivered the highest throughput in MLPerf inference results; management has improved the GB300 Blackwell system’s throughput by 2.7x and cost by 60% in just 6 months; management has line of sight to $1 trillion in Blackwell and Rubin revenue for 2025-2027

We capitalized on the inflection in inference demand by ramping Blackwell systems across our diverse end customer base. from hyperscalers to model makers to AI cloud providers and sovereign customers. In Q1, we also allocated capital effectively across R&D, investments in our ecosystem and share repurchases…

…Data Center revenue of $75 billion was up 92% year-over-year and 21% sequentially, driven by sustained strength in our Blackwell architecture and demand for GB300 NVL72 was particularly strong with frontier model builders and hyperscalers each having cumulatively deployed hundreds and thousands of Blackwell GPUs, marking the fastest product ramp in our company’s history. Grace Blackwell is the fastest training system as well as the lowest token generation cost at inference…

…Our Blackwell architecture is everywhere, adopted and deployed by every major hyperscaler, every cloud provider and every major model maker. Last month, we celebrated OpenAI’s launch of GPT-5.5, codesigned for, trained with, and served on Blackwell, currently positioned at the top of artificial analysis leaderboards. Microsoft’s Fairwater, the world’s most powerful AI data center is now live, ahead of schedule, powered by hundreds of thousands of Blackwell GPUs. Starting this year, AWS will add more than 1 million Blackwell and Rubin GPUs and are collaborating on Spectrum Networking. At Google, Blackwell will be offered to customers in the cloud, including confidential computing capability, a new foundation for secure high-performance AI…

…MLPerf inference results are in, and once again, we swept every benchmark as Blackwell Ultra delivered the highest throughput across the broad set of models and deployment scenarios. Full stack innovations drove the 2.7x increase in throughput and a 60% reduction in the cost per token on GB300 compared to just 6 months ago…

…We are continuing to work vigorously on our supply chain ecosystem to address the incredible demand we see ahead of us, giving us full confidence in the $1 trillion in Blackwell and Rubin revenue we foresee from 2025 through calendar 2027.

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

Spectrum-X, our end-to-end Ethernet platform purpose-built for AI, is now larger than all Ethernet network peers combined. InfiniBand has also had a very strong quarter, growing more than 4x year-over-year, driven by deployments of our next-generation XDR technology.

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

Back to our Data Center results. Hyperscale revenue of $38 billion was approximately 50% of Data Center revenue and increased 12% quarter-over-quarter. ACIE revenue was $37 billion and grew 31% quarter-over-quarter, including AI cloud revenue that more than tripled year-over-year. Our customers have enabled rapid stand-up of AI compute capacity. The number of partner data centers exceeding 10 megawatts has nearly doubled in just 1 year, now surpassing 80 sites. Sovereign revenue increased more than 80% year-over-year. NVIDIA AI infrastructure is now deployed across nearly 40 countries, representing $50 trillion in GDP.

NVIDIA’s management is seeing rising prices for renting the company’s previous Hopper and Ampere generations of GPUs

The value of NVIDIA AI infrastructure is rising. The price of renting an H100 has risen 20% year-to-date, while A100 cloud pricing is up nearly 15%. Benefiting from the versatility of our platform and continuous performance enhancements enhanced by our software stack, customers are generating profitable revenue beyond the depreciable life of their GPUs.

NVIDIA’s management is seeing the largest hyperscale workloads, across search, advertising, recommendation systems, and content understanding, continue to transition from CPUs to GPUs

First, from search and advertising to recommender systems and content understanding, the largest hyperscale workloads continue to transition from CPU to GPU-based accelerating computing.

NVIDIA’s management is seeing an inflection in the adoption of AI-native products and services, led by a transition to agentic AI; management is seeing incredible momentum with the AI model builders, with OpenAI’s Codex being a standout; there are a few hundred thousand AI agents today, but management sees a future world with billions of agents and they will all be using tools; management sees AI agents spinning off sub-agents, and each spin requires inference; management sees agents as having lower patience than humans

The adoption of products and services native to AI is inflecting. Since the advent of ChatGPT, we have witnessed mainstream AI transition from one-shot inference to reasoning and to now agentic…

…Growth in the model layer, particularly at Anthropic and OpenAI has been incredible with momentum continuing to accelerate, including breakout growth in OpenAI’s Codex since the launch of GPT-5.5…

…My sense is that the world is going to have billions of agents. Not today, I mean, we’re going to grow into it, but we’ll have billions of agents. And those billions of agents will all use tools. And those tools can be like PCs, just like us humans using PCs today. In the future, you’ll have an agent using PC and so if you kind of think along the lines of in the future, you pick your favorite number of agents at the moment. At the moment, call it, a few hundred thousand, but in the future, call it, eventually a few billion…

…Every one of those agents are going to spin off subagents. And every time they spin these off, you’re going to need to do inference…

…Agents use these tools and have — they have lower patience and tolerance than humans, and they want things to happen quickly.

NVIDIA’s management sees a $3 trillion to $4 trillion AI infrastructure opportunity by the end of 2029, driven by hyperscalers’ forecasted capex of over $1 trillion in 2027; management expects NVIDIA’s business to be growing faster than the growth in the hyperscalers’ capex; management expects hyperscalers’ capex to continue growing from here, because in the age of AI, compute equates to revenue, unlike in the SaaS (software-as-a-service) era

With analysts now forecasting hyperscale CapEx to exceed $1 trillion in 2027 and Agentic AI beginning to proliferate all industries, AI infrastructure spending is on track to reach $3 trillion to $4 trillion annually by the end of this decade…

…We should be growing faster than hyperscale CapEx. And the reason for that is illustrated by the segmentation that I just described. Our data center business has 2 large parts. It has more parts than that, but we combined it into 2 large parts for simplicity’s sake…

…The hyperscale CapEx that you were just talking about. And there are $1 trillion this year. I have every expectation it is going to grow from here for fundamentally good reasons. This is the way computing is going to work in the future. And if they don’t have the compute, they won’t have the revenues. It is very clear, compute is revenues, compute is profit. And so the world is changing. Software didn’t use to use — SaaS didn’t use to use as much compute, but AI requires a tremendous amount of compute.

NVIDIA has deepened its collaboration with Anthropic and will serve Anthropic’s AI compute needs through multiple cloud providers; management sees NVIDIA’s share of frontier AI models growing significantly; NVIDIA is the only platform that runs every frontier AI model

We have deepened our collaboration with Anthropic and are delighted to be a strategic partner to expand their compute capacity. We will support the company’s growth trajectory through AWS, Azure, CoreWeave, SpaceXAI and more. Now with the addition of Anthoropic too, OpenAI, Gemini, SpaceXAI, Meta MSL, Microsoft AI, TML, Reflection, Perplexity, Cursor, and other major frontier labs already building on NVIDIA. Our share of frontier AI models will grow significantly…

…NVIDIA is the only platform that runs every Frontier AI model.

NVIDIA’s management thinks the right metric to analyse the economics of NVIDIA’s GPUs is not the price paid, but the lifetime cost of the GPU in producing intelligence

Customers do not buy GPUs. They build AI factories and the right economic metric is not the purchase price of the GPU. It is the lifetime cost of an AI factory producing intelligence. Token per watt, tokens per dollar, uptime, utilization, time to production, software durability and asset life. NVIDIA excels at all of them.

NVIDIA’s management sees agentic AI as a growth opportunity for CPUs; NVIDIA’s Vera CPU can deliver 1.5x faster performance per core, 2x performance per watt, and 4x density per rack compared to x86-based CPUs; CPUs are a market NVIDIA has never addressed prior to Vera; management sees a total addressable market of $200 billion for CPUs in agentic AI; management has visibility to $20 billion in total CPU revenue in 2026 (FY2027); management sees 4 different use cases for Vera, which are Vera with the Rubin GPUs, Vera as a standalone CPU, Vera with CX-9 for storage, and Vera with CX-9 for security; the $200 billion CPU addressable market for Vera is specifically for Vera as a standalone CPU; management sees the Vera CPU as being supply constrained throughout the life of a Vera Rubin; an AI agent is a harness around an AI model, and this harness runs on a CPU, and the tools the harness utilises also runs on a CPU; Vera was designed to be an agentic CPU; traditional CPUs have many cores that are rentable, but agentic CPUs are designed to generate and process tokens and this is a strength of the Vera CPU; management sees the Vera CPU as the second largest driver of NVIDIA’s revenue beyond the $1 trillion in revenue-visibility management has for Rubin and Blackwell

Agentic AI and reinforcement learning represents new growth opportunities for CPUs. Building on the success of our Grace CPU, Vera is arriving just in time to meet this inflection. Built on custom ARM cores and codesigned end-to-end with Rubin GPUs and NVLink, Vera will deliver up to 1.5x faster performance per core, 2x performance per watt and 4x density per rack compared to x86-based alternatives. Vera CPU opens a brand-new $200 billion TAM for NVIDIA, a market we have never addressed before, and every major hyperscale and system maker is partnering with us to get it deployed. We have visibility to nearly $20 billion in total CPU revenue this year, setting us up to become the world’s leading CPU supplier…

…4 ways — let me just start with the one that you already know. The first way is Vera Rubin. And we’ll sell millions of Rubins, and every 2 of them is connected to a Vera. And of course, we price those 2 and they’re properly priced. And so that’s #1 use case. The second use case is Vera standalone CPU. The third is Vera with CX-9 and the software stack for storage. And then Vera in a — with CX-9 with a software stack for security and compute isolation and confidential computing. Okay, so each one of those use cases is built on Vera. And my sense is that we’ll be supply constrained throughout the entire life of Vera Rubin. There are 4 different use cases of it. And — but anyhow, the answer to your question is — of the $20 billion is a stand-alone…

…An agent is essentially what people call a harness. The agent has a harness that does the — and the harness could be OpenClaw, it could be Hermes, code — Claude Code is essentially a harness around Claude around the Opus model. OpenAI’s Codex is a harness around the GPT-5.5 model. And so these are harnesses. And these harnesses provide for things like IO, orchestration, memory management, tool use connected to tools, for example, browsers and things like that, C compilers, python compilers. And so the harness runs on CPU. And the tool use runs on CPUs. So for example, if the AI were to do a search or do a browser, use a browser that would run on the CPU…

…Vera was designed to be an agentic CPU. The CPUs of the past were designed to have many cores so that it could be easily rentable. People rented cores. Well, agents don’t rent cores. They just want the work to be done fast. The economics of the past was dollars per core. That’s the economics of cloud computing of the past. The economics of the AI of the future is tokens per dollar or dollars per token. And so what we need to do in the future is to generate tokens, process tokens as fast as possible, and that’s what Vera does incredibly well…

…[Question] Back at GTC, I believe you discussed $1 trillion visibility into both your Rubin and Blackwell platform revenue. But I believe that excluded things like LPX, Rubin, CPX and the Vera CPU racks. Can you maybe give us a sense about whether the Vera CPUs are going to be the biggest source of upside above and beyond that $1 trillion?

[Answer] In terms of incremental above the $1 trillion, I would say, one, the continued growing of share of the Frontier AI models. I’m expecting to grow more share. And so I’m expecting that to grow. Number two, we didn’t include any Vera CPU, stand-alone CPU in that number. And so I expect that to be the second largest. The TAM is, of course, quite large in agentic systems, and all of our customers are quite excited about Vera and we’re going to sell a whole bunch of Veras. And then third would be LPX, because as I explained earlier, LPX is designed as a — because of its SRAM architecture, it has the benefit of very low latency and very, very high interactivity, but it’s — also its throughput, its context processing ability is also quite limited.

NVIDIA’s next-generation GPU system, the Vera Rubin, is on track for shipment in 2026 Q3 (FY2027 Q3); Vera Rubin can deliver 35x higher inference throughput and 10x greater AI factory revenue compared to Blackwell systems; Google Cloud will be supporting 960,000 Rubin GPUs across multiple sites for customers; management thinks every single frontier AI model company will be adopting Vera Rubin once it’s launched, and that Vera Rubin will be even more successful than Blackwell even though they are unsure if Vere Rubin will ramp as quickly as Blackwell

We are on track to commence production shipments of Vera Rubin in the second half of this year starting in Q3. By integrating 7 purpose-built chips across 5 accelerated racks, Vera Rubin will deliver up to 35x higher inference throughput and up to 10x greater AI factory revenue compared with Blackwell. As an early adopter, Google’s A5X bare metal instances, which can support up to 960,000 Rubin GPUs across multiple sites can enable customers to run their largest AI workloads on NVIDIA’s optimized infrastructure…

…Every single frontier model company will jump on Vera Rubin from the get-go, and that wasn’t true before on Blackwell. And so Vera Rubin is off to a tremendous start and will surely be more successful than even Grace Blackwell…

…[Question] You mentioned GB300 is sort of the fastest ramp in the company’s history. How should we think about Vera Rubin against this benchmark. It’s obviously a new architecture at the silicon level, but in similar rack. Does that mean we should expect a similar slope to the Vera Rubin ramp as the GB300?

[Answer] It’s hard to say at this point what will be a faster ramp. But again, we have demand already planned, we’ve got POs. We’ve got almost all of our major customers ready to go, and these are very complex systems that we need to put together. So I think it’s just about the timing that it’s going to take for us to get that into market. Nothing else other than getting from production of all of the different systems that we have ready for order.

NVIDIA is yet to generate revenue from China and management does not know if the company’s AI chips will ever be allowed into China

While the U.S. government has approved licenses for H200 to be shipped to China-based customers, we have yet to generate any revenue, and we are uncertain whether any imports will be allowed into the country.

NVIDIA’s Physical AI revenue has exceeded $9 billion in revenue in the last 12 months; NVIDIA will power Uber’s robotaxi fleet in 30 cities and 4 continents by 2028; companies building industrial, surgical, and humanoid robotics are using NVIDIA’s technology; management thinks physical AI encompasses industries that have been untouched by IT (information technology) for the past 30 years, but they will soon be impacted by AI

Our physical AI continues to gain momentum, exceeding $9 billion in revenue over the last 12 months. Our partnership with Uber will power the robotaxi fleet across nearly 30 cities and 4 continents by 2028. And in robotics, leading companies across a range of industrial, surgical and humanoid applications are building on NVIDIA’s technology to develop and deploy at scale…

…When I talk about physical AI, and I talk about how the rest of the $100 trillion industry that has not been affected by — impacted by IT in the last 30 years. It’s about to be impacted by AI.

NVIDIA has increased inventory purchase commitments to $145 billion; NVIDIA is facing supply challenges

 In Q1, we increased total supply, inclusive of inventory purchase commitments and prepaid to $145 billion. While we are not immune to supply challenges, we remain confident in our ability to support the growth opportunity ahead with our intense focus, scale and long-standing partnerships with critical suppliers continuing to serve us well.

NVIDIA’s management thinks every base station in the future would be an AI-powered radio network

In the future, every single base station, every single radio network would become an AI-powered radio network.

Frontier AI companies are growing revenues in 1 month what older SaaS companies took a decade to achieve

Frontier AI companies, both Anthropic and OpenAI growing at an incredible pace. The fact that they can grow within 1 month, what some of the SaaS companies would have taken a decade to grow tells you something.

NVIDIA’s management thinks industrial AI will likely not be delivered via the cloud; the hyperscalers were happy to adopt AI first because they focused mostly on consumer applications where the stakes are lower but for industrial applications, AI needs to be really capable, safe, and productive before adoption can happen; right now, industrial AI has developed slower than consumer AI, but management thinks industrial AI will be even larger than consumer AI in the future

Many industrial companies, there’s no choice, but to put the computer where the context is, where the action is, you can’t put that in the cloud. It has to respond reliably, quickly every single time, can’t imagine a chip plant, a chip fab being connected to a cloud service provider, doesn’t make any sense…

…Hyperscale developed AI first for a lot of reasons. They have great computer science. They have excellent data center capability. And they also focus largely on consumer applications, which, if not perfect, is not the end of the world. It enhances the service — so long as it enhances the service. And so for many of the other applications, industrial applications, enterprise applications, until the AI is very capable and does really productive work and does it safely, and it could do it in a way that can actually generate impact and income, it doesn’t really get used. And so you expect the second category to develop slower than hyperscale, and you could see that in the numbers. However, long term, if you look at industrial and enterprise, clearly, that’s where future economics is going to be because it represents some $50 trillion, $80 trillion of the world’s economy. And so — and it’s going to be larger than that because of AI.

NVIDIA’s management thinks sovereign AI clouds will not want to use custom or semi-custom AI chips

The sovereign AI clouds. And so there’s a whole category of data centers that semi-custom chips just don’t apply because these data centers want to buy systems, they want to operate systems, they don’t want to design, they don’t want to build it themselves.

NVIDIA’s management sees the company taking market share in inference really quickly partly because of its new partnership with Anthropic; management thinks most of the inference taking place in AI data centers outside of the hyperscalers will be on NVIDIA

we are growing share in inference, and we’re growing share in inference very, very quickly. And the reason for that is this year, the number of frontier model companies grew. And so there’s Cursor and Perplexity and there’s some new model companies, TML and Reflection and the list goes on. And so the number of frontier model companies has grown, and we added Anthropic to our partnership this year. They’re expanding incredibly fast. We’ve partnered with them to secure computing capacity across Azure, AWS, CoreWeave, I forget who else we’ve already announced, but there’s a whole list of others that we are bringing online for them. And so the amount of capacity that we’re going to bring online for Anthropic this year and next year is going to be quite significant, very significant. And so we’re growing and our coverage of Anthropic has been largely 0 until just recently. And so we’re gaining share tremendously fast in inference…

…Everything that I’ve just explained in the inference question is really focused on hyperscale. Remember, there’s a whole second category of AI data centers that we serve almost uniquely. Now this segment is very fragmented. It requires a fairly integrated — a really well-integrated platform solution and a very large go-to-market. And that segment, all of the inference, 100% of that — the vast majority of that is NVIDIA.

NVIDIA’s management sees the LPX server rack as a specialty rack designed for low latency and high token rate but with low throughput

The LPX is designed for a low latency and high token rate. But its throughput is low. Its throughput is low. Its model size capacity is low. And its context processing, its ability to absorb a lot of context, for example, for software coding, for agentic workloads, its ability to absorb a great deal of context is lower. And so the challenge is simply, and I’ve explained before that the use case for LPX is not broad. It’s intended for somebody who has a fairly large portfolio of different types of token services. And for the high token rate, maybe these services are quite premium and the number of customers is not significant, but the token rate is very high.

Okta (NASDAQ: OKTA)

Okta’s management sees each AI agent in an organization as a new identity; AI agents are a rapidly-growing identity category, but they are the least governed; Okta brings agents under control by treating them as identities that can be managed and governed by existing identity management systems; management thinks that there will be more AI agents than humans over time, so the identity becomes increasingly important; all of Okta’s top 100 customers are deploying AI agents, but they are mostly doing it in a haphazard way in terms of security; management is seeing companies start to realise the importance of security for AI agents; management believes that companies will be getting their agentic capabilities from many different platforms; 90% of Okta’s customers have agents in production, but only 22% are confident in the governance of the agents

The future of technology is agentic. For Okta,, this represents a tremendous opportunity and an even greater responsibility. Every agent inside an enterprise is a new identity. Today, AI agents are the fastest-growing identity in the enterprise but also the least governed. Okta helps bring agents under control by treating them as first-class identities that can be managed and governed by their existing identity management system. We believe, over time, most large enterprises will have more agentic identities than human ones. This shift broadens the attack surface because every agent comes with credentials privileges, and the ability to act on a user’s behalf. In turn, this raises the strategic value of the identity layer because governing autonomous systems requires the kind of control, audit, continuous intent-driven authorization and real-time enforcement only an identity platform can deliver…

…I’ve spent the last 6 months, I’m on this goal to talk to in-person face-to-face with our top 100 customers, about 75 customers in. And when you mix that with a bunch of other conversations, here’s what’s going on, everyone is deploying agents in some way, shape or form. But they’re really just starting to think about and put in programs in place to lay out the rails of governed managed adoption. So a concrete example is you’ll have a development team that is using cloud code, but it’s connected to GitHub and their JIRA system with static tokens in the local developer box. So that company is viewing agents, but they’ve really done it in a haphazard nonsecure way. And what’s happening now is they’re figuring out those rails. They’re figuring out how they’re going to have secure connections, have a system to monitor where all the agents are, have the ability to support it for multiple platforms…

… I think what I’m seeing is that Boards and CEOs are saying, we know this agentic thing is real. We’ve got to put the guardrails in place for that. and we know that security is real, and we’re going to spend money on that. And it’s, the reality of it, Brian, is that it’s the fundamentals. It’s identity. 80% of breaches are go through identity. And you know you have to patch your systems. You know you have to have a good multilayered defense and Zero Trust so you can defend for multiple ways…

…there’s a few fundamental truth right that are going to play out. I think, one is that they’re going to get agentic capabilities from many, many companies. They’re going to have different platforms. They’re going to have hyperscaler platforms. They’re going to have Foundation model platforms. They’re going to have open source platforms. They’re also going to get agentic capability from apps. Salesforce is going to have there. Workday is going to have their ServiceNow is on and on…

…Customers have a problem today. They have a problem today where over 90% of them have agents in production, and only 22% of them are confident to have them governed.

Okta’s management sees 3 advantages the company has in securing AI agents, namely, (1) distribution, where Okta can extend its identity system to AI agents, (2) product breadth, where Okta is the only vendor that address both sides of the agent security problem, and (3) neutrality, where Okta allows customers to choose whichever cloud provider and agentic platform they want; Okta’s 3 advantages in securing AI agents are mutually reinforcing; Okta as a neutral identity layer, can help customers avoid vendor lock-in for agentic capabilities

To help our customers confidently secure this shift, we’re building on 3 unique advantages, each with powerful network effects: distribution, product breadth and neutrality…

…In the agentic era, identity becomes even more foundational. When a customer secures their agents with Okta, they are not taking on a new platform; they are extending the trusted foundation they already rely on with Okta. We’ve already seen how our customers benefit from this expansion in other parts of our business. Customers are finding value in Okta’s unified identity system as Okta in governance was once again the leading contributor among our new products. This distribution flywheel is evident in our results…

… Our second unique advantage is product breadth. We are the only vendor with solutions that address both sides of the agent security problem…

… The third unique advantage is neutrality, which is more important than ever. The AI landscape is opting rapidly. Customers need an identity solution that frees them to choose whatever technology serves their business best without fear of vendor lock-in. As the leading independent and neutral identity platform, Okta gives organizations the flexibility to do exactly that. In the same way, enterprises run workloads across multiple clouds, they are deploying agents across various platforms like OpenAI, Anthropic, Google, Microsoft, Salesforce and a growing set of open source frameworks. Managing and securing an autonomous workforce requires a neutral, independent identity layer that others can’t provide. In practice, cloud providers, model providers and agent platforms are partnering with Okta to securely manage agent identities as they continue to proliferate across the enterprise…

…These 3 advantages are unique and mutually reinforcing. The more organizations use Okta to secure their agents, the more identity signals flow into our platform and the stronger our governance and detection becomes, and our neutrality allows us to secure current and future agent frameworks for customers, allowing Okta to capture more of the addressable market…

… I think, one is that they’re going to get agentic capabilities from many, many companies. They’re going to have different platforms. They’re going to have hyperscaler platforms. They’re going to have Foundation model platforms. They’re going to have open source platforms. They’re also going to get agentic capability from apps. Salesforce is going to have there. Workday is going to have their ServiceNow is on and on. Everything is going to be agentic — have agentic capabilities. But we know they’re going to have a directory of these things or roster everything, a policy layer and they’re going to have to make sure they can connect to things. And so we’re seeing our customers — it’s a kind of a no-regrets move to pick this independent and neutral identity layer that can solve those fundamental problems without locking them in 

Okta has two product categories to address both sides of the agent security problem; Okta for AI Agents became generally available in April 2026 and provides enterprises with centralised visibility into agents with identity governance capabilities; Auth0 for AI Agents is for developers building AI agents and it helps developers ship secure agents inside their products; Okta had strong pipeline generation in 2026 Q1 (FY2027 Q1), driven partly by Okta for AI Agents and Auth0 for AI Agents; the opportunity for Okta for AI Agents is not limited to existing workforce customers, and it extends to every enterprise with a multi-platform AI strategy; Okta for AI Agents is integrated with ServiceNow and Amazon Bedrock; there is a lot of interest in Okta for AI Agents and Auth0 for AI Agents, but they are still early and are currently not contributing materially to the business; management believes Okta for AI Agents and Auth0 for AI Agents will become really big products; Okta can give agents specific access to different apps based on access management; the pipeline for Okta’s agentic products is bigger than anything management has ever seen; the pipeline for Okta for AI Agents is bigger than that for Auth0 for AI agents because companies are further along with deploying internal agents than building agents into products; management is already starting to see some pull-through of demand for Okta’s non-AI products because of Okta for AI Agents

Okta for AI agents, which became generally available last month, gives enterprises a single control plane to discover, govern and manage agents across their organization. It is the first and best implementation of the blueprint for the secure agentic enterprise, an industry framework for bringing agents under control by answering the three questions that have dominated my customer conversations over the past several months. Where are my agents, what can they connect to and what can they do? Enterprises need to maintain visibility and control over their sprawl of agents, ensuring they have governed identities, consistent access policies and ways to shut them down to secure every agent into end. Okta provides customers with centralized visibility into agents with identity governance capabilities, including ownership assignment and life cycle management while giving IT and security teams, critical security controls to deactivate rogue agents. For developers building AI agents, Auth0 for AI agents provides the identity foundation to ship secure agents inside their products. Auth0 for AI agents secures agents, APIs and users effortlessly for B2B, B2C and internal apps, all backed by the enterprise grade Auth they already trust. In tangible terms, pipe generation in Q1 was strong, driven in part by these 2 new products…

…Okta is the only modern identity platform purpose-built to sit above the agent ecosystem, and it federates with whatever identity provider a customer runs. That means the opportunity for Okta for AI agents is not limited to our existing workforce customers. It extends to every enterprise with a multi-platform AI strategy…

…We’ve entered into a partnership with ServiceNow that integrates their AI control tower product with Okta for AI agents…

…Okta for AI agents now integrates with Amazon Bedrock Agent core to provide customers with identity governance capabilities for their agents…

…They’re figuring out how they’re going to have secure connections, have a system to monitor where all the agents are, have the ability to support it for multiple platforms. And that’s why you’re seeing the record interest and the record pipeline for what we do with Okta for AI agents and Auth0 for AI agents. The reality is of these products, it’s still early. They’re not materially contributing to the business in Q1. In fact, we’re still being prudent in our guide. They’re not even — they’re a little bit in the guide, but not significant in the guide but it’s going to be big…

…So it’s very natural to say, who can really manage these connections and give me these governed rails for all these secure connections, where my agents are, what they’re doing, what can they do? It’s a natural fit for us. So I think as they build out this infrastructure, we’re in this really great position to have to be a super, super meaningful part of the business and TAM over the next several quarters and several years…

…We tell you who your agents are. There’s a directory of agents. We can scan multiple platforms and multiple systems and give you that source of truth of where your agents are and we can help you set a policy on what they can connect to. Agents can this from teams and they can read this from Slack, and they can read this information from Snowflake and they can you read this from GitHub. So it’s like a single sign-on or access management…

…[Question] You mentioned a building pipeline on AI. I wonder if you might hope with the size of this maybe relative to other products in the past

[Answer] The pipeline is bigger than anything we’ve ever seen…

…[Question] The difference between AI for agents in Auth0 versus Okta, the 2 different platforms. Maybe just help us appreciate the technology aspect of that? And is there like a big difference in size of pipeline between the 2? 

[Answer] They’re both healthy, the Okta pipeline is bigger. And I think that’s because it’s a little bit of a — I think the companies that are figuring out how to manage and deploy internal agents are further along than people building agents into their products and into their websites…

…We’re seeing that the products we’ve offered for AI agents in this blueprint, this vision we have for the industry and agents is raising the strategic level of conversations, which is pulling in other products and helping us displace legacy faster and sell more of our existing products and our newer products into new customers in the base than we would be otherwise. I say that because to make it clear that the AI agent products are still, is still immaterial, the contribution with Okta for agents going GA in April. They had a good quarter, but it’s still a small base. So the pull-through is real already though.

Okta’s management believes that no single company can address the agentic security market; Okta has entered into partnerships with AI leaders ranging from ISVs (independent software vendors) to AI vendors and hyperscalers; the ISVs include ServiceNow, while the AI vendors include Anthropic and OpenAI; Okta is partnering with Anthropic for its Project Glasswing cybersecurity initiative

Neutrality becomes even more important when it comes to technology partnerships and integrations, like the traditional cybersecurity landscape, no single company can address the agentic security market alone. That’s why we’ve partnered with AI leaders from ISVs to hyperscalers to frontier AI vendors, and I’d like to highlight a few of those partnerships today.  We’ve entered into a partnership with ServiceNow that integrates their AI control tower product with Okta for AI agents. Our partnership with Google brings centralized identity guidance and access control to Google’s agent gateway. Okta for AI agents now integrates with Amazon Bedrock Agent core to provide customers with identity governance capabilities for their agents. We were a launch partner for OpenAI’s release of GPT 5.5 trusted access for cyber. And finally, we’re collaborating with Anthropic in a number of ways to testing Anthropic’s preview model as part of Project Glasswing to a new integration between Okta Identity Security Posture Management and the Cloud compliance API.

Okta’s management is pricing agentic products as an increase to a user’s monthly price because (1) management is seeing customers want to consume agentic products via this pricing model, and (2) agents are currently mostly deployed on behalf of users; management thinks pricing models for agentic products will evolve over time and the software industry is still figuring it out; management is seeing that the average deal size for AI-specific deals is much larger than the average deal size for other types of deals; Okta does not have unlimited-consumption AI deals

And so the way we’ve done pricing for our products is exactly in line with how our products have been priced in the past. They’re priced on, it’s an uplift to a named user or it’s an uplift to a monthly active user. Now you might say, “Hey, Todd, but agentic — agents are this new thing and why are you pricing them on an active user or a named user price?” And that’s for two reasons. One reason is that’s the way customers want to consume it right now. And two, the majority of concrete use cases in the world right now for agents, it’s on behalf of the user. It’s an agent working on behalf of a software developer. It’s an agent working on behalf of a support rep. It’s an agent working on behalf of someone in accounting. So it’s very natural how they want to buy it and how they’re actually being used. So it’s an uplift on a named user, and it’s uplift on an active user.

Now we fully understand that, that’s going to evolve. And there will be more autonomous agents that have to be priced not by user base or not an extensive user. They have to be the unit has to be the number of agents. It’s a little bit tricky because it’s very hard to define the number of agents because some person might say, “Oh, I have 1,000 agents, but it’s really kind of 1,000 copies of the same agent or 1,000 instances of the same agent. In other cases, it might be literally 1 instance of an agent acting for many, many different use cases. So the industry is kind of figuring that out, and we’ll figure that over time how to monetize and price that now…

…The average deal size for these AI-specific deals is significantly larger than the average deal size for the rest of the company…

…[Question] You guys are doing deals where basically the contract is for an unlimited number of agents. The good thing is in those deals, I’m hearing that the spend is very, very high relative to your existing spend and other products. But the risk there is what if the customer doesn’t get to unlimited agents, so there’s downside renewal or other things that could happen. So how are you approaching that dynamic with customers in factoring in the contracts?

[Answer] There’s no unlimited. If there is unlimited, it’s time bound. So there have been some deals where we’ve done like a year, and then it’s like we’re going to figure out after a year what the — how the use case really unfolded and how to snap it back to the kind of normal pricing model. But there’s no — it’s not unlimited in the sense of time and volume.

Okta’s management is seeing the leaders of AI companies being worried about the durability of their revenues

If you look at the — particularly the AI landscape, I was having dinner with a bunch of CEOs of companies, different sizes, and everyone is super worried about their spend in their products and their revenue in their products being not durable because it’s token spend, and they worry about the products being used and then and maybe someone is going to look at the spin and stop spin the token spend

Okta’s governance-related product portfolio is still performing well; Okta’s privileged access product is not as mature as the governance portfolio

We’re very excited about our AI products. But governance continues to be a strength for us. We talked over the past couple of quarters about how governance has evolved from being primarily a cross-sell add-on product to also now being a land product. And we are seeing sizable land opportunities, starting with governance at some companies that are displacing systems that they’ve had in place. So we’re very excited about the enterprise readiness and robustness of our governance product and the rural deployments…

…Privileged access is further behind governance on that maturity curve. It came to market a little bit later. We’re continuing to invest heavily in it and we did an acquisition back Q3 at Axis to add capabilities to that. And we’re continuing to invest in that breadth of portfolio, kind of rounding out the identity security fabric in addition to all the momentum that we’re seeing with our great success in the AI product.

ServiceNow wanted kill switches for rogue AI agents; Okta can help sever connections and access that any rogue AI agent has

ServiceNow is, as you mentioned, super interesting. They are — their product strategy is they want to be the control tower for all AI agents. And what is, what they were really interested in was this kill-switches capability. When agents go awry and agents aren’t following the policy, how do you shut them down, and that can mean a lot of different things. That can mean actually stopping the running of the agent that can mean quarantining the agent at a network level, there’s many different strategies. The one thing we do really well and that they wanted from us is the ability to sever the connections, the access tokens, the actual logical connection at the authorization layer to the back-end resources, and we’re really good at that. That’s kind of the core of our product. What can these things connect to, what can they do.

Okta’s management thinks that cybersecurity in the future will take multiple companies to secure, and the large AI model providers cannot do it by themselves

I think in terms of the model providers, how they’re going to play in the broader cyber ecosystem, it’s going to take a village. I think we’ve seen that in cyber forever. I think consolidation in cyber never seems to work. All seems to be — gets to a certain point and then new threats emerge, and the companies that are trying to consolidate cyber have such a hard time integrating amongst themselves. It kind of fractures a part. And I think that will continue. I think cyber in the agentic world is going to take a village, and we’re going to have to make sure it’s integrated together and make sure we have layered defenses. And that’s why I think it’s really healthy to be coming into this conversation with this open mindset of, hey, we have our lane, we’re going to try to provide the best identity foundation in the world and then connect around that in a standard way that helps customers get great outcomes.

The cost of inference is real at Okta, and management thinks more companies using AI models will be scrutinising their inference costs in the near future; management is optimistic that Okta can manage inference costs and drive positive ROI (return on investment)

The cost thing you’re talking about is real, the inference costs and the AI tooling and what it’s driving in terms of expenses. And I think you’re going to see at Okta, and then over the whole industry over the next 6 to 12 months, you’re going to see a little bit more scrutiny in terms of what are you getting from all this, the inference cost you’re spending, how is it translating, which is not surprising given the amount it’s rising across the industry.  And we’re going to come out the other side with more balanced ROI-driven investment portfolio of how we spend these things. And we’re optimistic about how it’s going to work out very well for us.

Salesforce (NYSE: CRM)

Leading AI companies are all Salesforce customers, in particular, Slack customers; Slack was half of Salesforce’s $1 million-plus wins in 2026 Q1 (FY2027 Q1), up 80% year-on-year; Slack is AI startup Anthropic’s core operating system; Slackbot is also a MCP (model context protocol) client; Slack MCP has seen 1 million users in 6 weeks; Slack’s agentic work units (AWUs) was up 350% sequentially in 2026 Q1 (FY2027 Q1); management thinks that in 2 years, there will be more agents using Slack than people; management thinks agents need the context and data that resides in Slack; internal usage of Slackbot by Salesforce has led to 3.8 million hours of annualised productivity gains; Anthropic is one of the biggest users of Salesforce’s Sales Cloud; Slackbot has increased the productivity of Salesforce by around 3%; management sees Slack as the place where humans and agents work together; management sees the work graph of enterprises living in Slack, which is already one of the richest work contexts, becoming even richer over time; 3 million custom apps were built by the community on Slack in 2026 Q1 (FY2027 Q1), up 8x sequentially, and 250,000 of the custom apps were 3rd-party AI agents, which doubled sequentially; management sees Slack on a fast track towards being a $10 billion cloud

OpenAI, Anthropic, Google, companies building the future of AI, all of them Salesforce customers, all of them Slack customers, building these incredible new capabilities with Agentforce…

…Slack, which every AI company in the Bay Area here is using to run their business, including OpenAI and Anthropic, transforming our customers into agentic enterprise. Slack was nearly half of our 1 million-plus wins this quarter, up 80% year-over-year…

…Anthropic calls Slack its core operating system, and that’s what Slack is becoming for every enterprise. All of our apps are Slack first. So now a service agent can summarize a case, update the record, escalate to a human right in Slack. And Slackbot is also an MCP client, so you can tell it to create a purchase order in NetSuite or update a project in Jira, and it happens, no switching tools. We’ve seen 1 million users of Slack MCP in the first 6 weeks, and Slack AWUs grew nearly 350% quarter-over-quarter.

In 2 years, there’ll be more agents using Slack than people. Every one of those agents needs the context and the data and the insights directly from Slack. Every workflow needs the data. Every action needs the integration and every customer needs to see what’s happening across the entire business. We have the largest collection of trusted CRM context ever assembled between Data 360, Informatica, MuleSoft, Tableau manage and deliver all that context so that any agent can reason, act, and deliver real outcomes…

…Slackbot, which is embedded directly into the flow of work, is now our fastest adopted AI tool in Salesforce’s history, driving 3.8 million hours of annualized productivity gains for our employees…

…Anthropic is one of our biggest users of CRM of Sales Cloud…

…Slackbot is our personal assistant. It has increased the productivity of the whole company around 3% more or less…

…When we say agents and humans work together, you experience it in Slack. When you’re in a channel and suddenly in a lot of these — especially I see it now in my engineering channels, like half the time, somebody puts a question or a request on a Slack channel and the agent is listening and answering it, developers do a PR request in Slack. And then suddenly, the agent is picking up and trying to do it. They want status reports. So I think Slack is where people can really understand the manifestation and they’re all asking questions as a human and Slackbot is even a better way of articulating that in a packaged way…

…Because that work graph that will become one of the richest work context in the enterprise is getting richer and richer. So we build — I mean, the community built 3 million custom apps on Slack in Q1. That’s 8x quarter-on-quarter. I mean there is a huge boom. Out of those custom apps, there were 250,000 that were AI agents that were built, third-party AI agents, and that grew more than doubled in quarter-on-quarter, grew eightfold year-on-year…

…I’m not giving guidance by what I’m saying, but sales is a $10 billion cloud already. Service is a $10 billion cloud already. Data is already a $10 billion cloud. I think when we see the growth rate that’s happening inside Slack, you saw the ACV was incredible in the first quarter. This is going to be fast track from something we bought with less than $1 billion that I’m sure we’ll be talking in short order about Slack being a $10 billion cloud as well.

Agentforce ARR reached $1 billion in 2026 Q1 (FY2027 Q1) (was $800 million in 2025 Q4, up 169% year-on-year); Agentforce and Data 360 reached nearly $3.4 billion in ARR (annual recurring revenue) in 2026 Q1 (FY2027 Q1) (was $2.9 billion in 2025 Q4, up 200% year-on-year); 50% of Agentforce and Data 360  bookings in 2026 Q1 (FY2027 Q1) were from expansions by existing customers; management recently announced Agentforce Coworker, where every Salesforce application now comes with a built-in autonomous agent; bookings for A1E and A4X, Salesforce’s premium SKUs that include agentic capabilities, was up 60% year-on-year in 2026 Q1 (FY2027 Q1); top 10 customers by AWUs (agentic work units) in 2026 Q1 (FY2027 Q1) increased their total Salesforce spend by 1.5x in the last 12 months; Agentforce allows every user of Salesforce to create agents

We’re seeing incredible demand for Agentforce with ARR now greater than $1 billion. And combined with Data 360 and Informatica Cloud, we’ve delivered $3.4 billion in AI and Data ARR. 50% of Agentforce and Data 360 bookings were from existing customers expanding their commitment.

…Very excited about our new Agentforce Coworker, which we announced last week. If you haven’t heard about that, every single one of our Salesforce applications now comes with a built-in autonomous agent. No complex configuration. You just turn it on. It becomes your coworker, finding answers, taking action, getting work done fast. To give you an idea of the impact that Coworker will have, people search for information inside Salesforce 1 billion times a month. Coworker turns search into answers and answers into action…

…Agentforce ARR surpassed the $1 billion mark this quarter. Our largest applications, sales and service saw year-over-year seat growth with humans and agents both expanding on the platform. Bookings for A1E and A4X, our premium SKUs anchored in sales and service, including the value from our agentic capabilities, grew nearly 60% year-over-year. As customers adopt Agentforce, they expand across our platform. On average, our top 10 customers by Q1 AWU usage have increased their total Salesforce spend by 1.5x in the last year…

…Those of you who are Salesforce users, the millions of people who use Salesforce every day, the search bar is a critical part of how the application operates. Now Agentforce is that search bar. So you can not only search and aggregate and get insights into information throughout every single app we have, but also create agents, and those agents can appear in Slack and Microsoft Teams and other applications, even in an app that’s going to run directly on your phone called Salesforce Coworker.

Salesforce has processed 28.6 trillion tokens to-date in 2026 Q1 (FY2027 Q1), up 152% sequentially (was 19 trillion to-date in 2025 Q4); Salesforce has delivered 3.8 billion AWUs (agentic work units) to-date, up 111% sequentially (was 2.4 billion in 2025 Q4)

To date, we processed 28.6 trillion tokens, up 152% quarter-over-quarter and converted them into 3.8 billion, as I mentioned already, agentic work units for our customers, up 11% — sorry, up 111% quarter-over-quarter. 

Salesforce acquired Qualified in 2026 Q1 (FY2027 Q1); Salesforce has integrated Qualified’s sales development representative (SDR) agent, Piper, into Salesforce; more than 700 customers are already using Piper; Piper is deployed on Salesforce’s website and is engaging with 50% of the website’s traffic, delivering 45% more pipeline than traditional web agents

In Q1, we completed the acquisition of the Qualified and integrated Piper, their SDR agent, into Salesforce. Brought all those great Salesforce alumni back home. More than 700 customers are already using Piper. It’s an incredible success, and we deployed Piper on salesforce.com, as I mentioned. So you’re going to be able to use it firsthand. I think that’s so great. It’s engaging 50% of our traffic and qualifying thousands of leads and delivering 45% more pipeline than traditional web agents.

Salesforce’s management recently announced Headless 360, which makes all of Salesforce accessible through MCP (model context protocol) clients, APIs, and CLA (command-line agent) prompts; since Headless 360’s launch in April 2026, Salesforce has already processed 4.5 million MCP calls and 1 trillion API calls; management thinks Headless 360 expands Salesforce’s addressable market into previously unmonetised areas; management is excited about Headless 360 in 2 areas, namely, (1) Headless 360 making it easier to implement Salesforce with coding agents, and (2) customers getting more value out of Salesforce through Headless 360; management is not seeing customers build in-house applications with Headless 360 to replace Salesforce; Slackbot is an example of a Headless 360 experience; the Headless MCP server for Slack has done 50 million tool calls; Headless is a way for agents to connect to Salesforce with APIs, because agents require slightly different types of APIs than what human developers used when connecting with Salesforce in the past

This quarter, we also announced Headless 360. Again, making all of Salesforce accessible through our MCP clients, APIs, CLA prompts. Headless 360 bringing together the human agents and headless platforms so you can use Salesforce with any coding agent across any surface. It’s going to speed implementations, drive consumption, more actions, more workflow, more data, more intelligence, all compounding across Salesforce. We’re meeting our customers where they are. Since launch in April, we’ve already processed 4.5 million MCP calls into our platform. Q1 alone, we processed nearly 1 trillion API calls, incredible…

…Looking ahead, the Headless 360 strategy that Marc walked through expands our addressable market into surfaces we’ve never previously monetized…

…I think what’s so exciting about Headless is 2 things. One, it’s having a real impact on making it easier to implement with Salesforce. So building out with Salesforce has now become easier than ever because we’ve seen these coding agents, Claude and Codex from OpenAI. As you use these things, what you realize is you need to be able to connect the underlying APIs, which you do through this layer that’s called MCP. And if you can connect those into the coding agents, it makes it faster than ever to implement and deploy Salesforce. And I think we’re seeing that show up in the numbers. Just this quarter alone, Agentforce customers in production grew by 50%. So I think we’re starting to see a little bit of that impact as not just our customers, but also our global SIs across the entire platform, absolutely implementing Data 360, implementing Agentforce, implementing a service. All of this — life sciences, all of this now becomes really just a conversation. So that’s one end.

But the other end is really what we heard from Miguel, which is this is really changing how people get value and consume Salesforce. In my experience, we’re not seeing people take this capability and the coding agents, for example, and try to build all of this stuff themselves. What they want to do is they want to take this capability and they want to use Salesforce in different ways and get more value out of it. So rather than logging into this discrete application and this application and this application to get an answer to one question that might span multiple applications or multiple kind of sources of information, you can now just take these MCP servers and plug them into any tool that you want…

…If you’re a Slack customer, you can get to it right with Slackbot. That’s really a Headless experience as well…

…We announced the Headless MCP server for Slack and Slack has done 30 million — 50 million tool calls…

…When you’re a builder, when you’re out there building something, and this is especially true today because there’s now an ocean of builders that have been created as a result of this coding agent boom. When you go to build something for your business, you, at some point, are likely going to want to connect to Salesforce that is what we see. And it doesn’t matter what platform you’re doing it on. You can be building something on a competitive platform to Salesforce or on Google or AWS or one of our partners. But at some point, you’re going to want to connect into Salesforce. And that’s why those APIs have always been hugely, hugely used. But when you are building with an agent, you need a slightly different type of API. That’s what we call MCP. And so by really putting those MCP servers out and saying, yes, this is how we want people to build.

Informatica has been a successful acquisition, performing the heavy lifting and data management that customers need to move agentic workloads from pilot to production; Informatica has helped drive an acceleration in revenue growth at Salesforce; Informatica’s bookings growth has accelerated significantly since being acquired by Salesforce

Informatica has an amazing acquisition. It performed incredibly well this quarter. It’s doing the heavy lifting and data management that every customer needs to move from pilot to production…

…And now with Informatica as part of Data 360, we’re already unlocking synergies with revenue growth accelerating since the acquisition. This is the flywheel we laid out at our Investor Day, and it’s working. Those signals show up in the headline numbers…

…Informatica was a business that was growing single digit, both on bookings and revenue. In just 2 quarters, we have significantly reaccelerated that the bookings of the chart beyond anybody’s expectation because data is king.

Salesforce deployed Agentforce on its support website 15 months ago and it has already handled 4 million inquiries autonomously; Agentforce now handles 2x what human agents are handling on Salesforce’s support website; Agentforce Sales worked 220,000 leads for Salesforce autonomously in 2026 Q1 (FY2027 Q1), generating a $42 million pipeline; Agentforce Coworker is able to quickly answer questions that would have taken an hour to do so in the past

Since we deployed Agentforce on help.salesforce.com and on 1-800-NO-SOFTWARE, well, only 15 months ago, it’s autonomously handled now 4 million inquiries. It’s now double what human agents are handling…

…Over 25 years, Salesforce has generated tens of millions of leads. We never called back. In Q1 alone, Agentforce sales worked 220,000 leads autonomously, generating $42 million in pipeline, awesome…

…Agentforce Coworker was able to pull together and navigate our complex sales and ERP data to answer questions that just yesterday would have been 60 minutes of swivel chairing between screens and systems. It was pretty cool to see that.

Wine company Vivino is using Agentforce to support 74 million users with just 37 reps; Agentfore has helped Vivino reduce resolution time of customer queries by 70%; cyber security company McAfee has replaced ServiceNow with Salesforce’s Agentforce IT Service; Florida Prepaid is using Agentforce to autonomously handle 75% of business hour calls, and 100% of after-hour calls; cyber security company Fortinet is using Agentforce Sales for predictive lead scoring; Agibank built a sales development representative (SDR) agent with Agentforce Sales

Vivino, the world’s largest wine company supporting 74 million users with only 37 reps, kind of hard to believe, but it’s possible because its agent, Vivina, autonomously handles order status, lookups, account questions more autonomously slashing resolution time by 70%. McAfee has selected our new Agentforce ITSM product or what we call Agentforce IT Service to replace ServiceNow. They are using it for everything, ticket deflection, hardware provisioning, incident management. Florida Prepaid, a college savings plan provider with more than 200,000 accounts is using Agentforce voice to autonomously handle 75% of business hour calls and 100% of after-hour calls…

…Cybersecurity leader, Fortinet using Agentforce sales to power predictive lead scoring. Financial leader, AgiBank now built an SDR agent that instantly qualifies leads on WhatsApp.

Indeed is using Headless 360 to build and deploy Agentforce agents directly from Cursor; Just Eat is using Headless 360 to bring agents into WhatsApp for engaging 350,000 partners across 15 countries; Adecco is excited that agents they are building outside of Agentforce can now leverage Salesforce because of Headless 360; Anthropic’s usage of Slack through 2026 Q1 (FY2027 Q1) has grown 5x partly because they are using Sales Cloud via Headless 360; the presence of Headless 360 has made Sales Cloud even more strategic for Anthropic

With Headless 360, Indeed is building and deploying Agentforce agents right from Cursor and Just Eat Takeaway, one of the leading online food delivery platforms in Europe, we just had them speak to our entire management team with such an amazing story, is using Headless 360 already to bring agents into WhatsApp and other channels, engaging with 350,000 partners across 15 countries…

…Adecco, great customer across the board. They use pretty much every cloud. They went into Data Cloud and Agentforce last year. They did a big commitment in Q1, at the beginning of Q1. They are basically design and AELA, wall-to-wall. They have amazing recruiter agents going there, millions of transactions. They’re moving into voice. When we announced Headless, they called us and they are like, “Wait a minute, this is — let me try to understand what you’re doing.” So now because they are also using other platforms to develop other agents. So they have agents with some of the AI labs that they’re also trying to access our data. Are you saying that now these agents that we are building outside Agentforce can also leverage Salesforce? And we said, exactly, we did it for that. So now there’s going to be a lot of new agents that are going to be accessing our platform…

…Anthropic is one of our biggest users of CRM of Sales Cloud. And obviously, Slack, their usage through Q1 has exploded fivefold because now they are using Sales Cloud from a Headless perspective, and they are approaching it from Coworker, from other applications from Slack, they’re hitting Sales Cloud. So Sales Cloud has become more prominent and more strategic for them than ever because of Headless.

PenFed Credit Union handles 500 transactions every second, and 160 million member transactions annually, and wanted to deliver hyper-personalisation for customers; PenFed Credit Union chose Salesforce to enable the hyper-personalisation and now has 76 agents across various functions; PenFed Credit Union chose Salesforce because it has the products, engineers, and reputation that PenFed Credit Union was looking for in a vendor; PenFed Credit Union built Agent Wingman with Salesforce; in 2026 (FY2027), Agent Wingman will help PenFed Credit Union (1) save $1.6 million, (2) lower call handle time by 10%, (3) lower after-call work time by 50%, and (4) lower held calls by 40%; PenFed Credit Union has agents listening to a phone call with members for transcription; PenFed Credit Union only developed its agentic vision about 2 years ago

[PenFed Credit Union CEO] When we’re competing against 8,000 other firms, we got to deliver hyper-personalization and every transaction, we do about 500 transactions a second, 160 million member transactions a year. They have to be right anywhere in the world real time. So we built our entire platform over the last few years. We went from about 400 platforms down to literally 12 strategic partners. Our call center, our mobile, our web, and our branches all run on Salesforce. Every additional partner or tech siloed capability is a tax on innovation, it’s a tax on speed, and it’s a tax on security. So by building it around Salesforce, I really think it’s taking me 25 years to realize Jim Collins’ Flywheel Effect, we have 76 agents now running across operations, mortgages, IT, HR. All of our areas are adopting it to make our employees be more productive. We like to say they’re bionic employees now. We’re not losing employees. We’re able to add more volume at scale, industrialized scale with the same number of people, and we’re very proud of that…

…How is the decision really made? First of all, does the firm, in this case, Salesforce have the product and service that we need? Second, do you have the engineers, the architects, the professionals to work with my team in order to bring that vision to reality? And then lastly, even if another firm had those first 2, who is the firm standing behind it that can be there through good times and bad times that’s going to stand behind that product or service. When you line up all 3, that’s where a good trusted partnership exists. That’s why we went with Salesforce. So we work with your team literally hand in hand. We said we want to streamline processes. We want to take out latency in the code. We want to do X, Y or Z. Your team was there in the trenches at every level, engineers, architects, building out the vision. But then it’s not just pie in the sky on the white [ sheet ], it’s implementable. We have 76 agents running side by side with our employees. 

A good example is in our call centers. We have Agent Wingman. I’m an aviator, so I think they named it because I like Wingman. Agent Wingman is going to save me nearly $1.6 million this year, has decreased our call handle time 10% this year, 50% reduction in after-call work time and 40% reduction in held calls. So better experience for the member…

…I want my employees to do the knowledge work, building trust in the relationship, not entering what just happened on the phone call. We have agents that listen to the phone call, transcribe it. The human is still in the loop. They approve what was just talked about, but then it’s 360, if the transaction occurred in the branch, web, mobile. So the next person that deals with that consumer, that member, they know exactly the relationship. They know what we might want to sell them next or what they need next for their daughter, their graduation…

…We had the vision when we saw what was possible 2 years ago. You can build it quickly. The most important thing is having the right partner and not to have too many partners. Too many partners slow things down.

UCLA Health has been working with Salesforce for some time; UCLA Health recently consolidated into a single instance of Salesforce’s Health Cloud; UCLA Health recently launched its first experiment with Agentforce, which is a customer-facing virtual concierge; UCLA Health was very cautious about launching a customer-facing virtual concierge

[UCLA Health executive] We’ve been working with Salesforce for quite a few years. But most recently, we’ve consolidated into one single instance of Health Cloud, and we’ve built on top of that with Marketing Cloud, Data 360, and most recently launched our first experiment with Agentforce, and that’s a customer-facing chatbot that just — it’s — right now, it’s only scraping our website to act as a little bit of a virtual concierge to direct patients to where they need to go. It’s helping with find a provider. It’s helping with general inquiries. It’s helping with clinical trials…

…I would say it took a while for us to sort of dip our toe in the water in the customer-facing space. We’re doing a lot on the back end when it comes to research, but this really has an impact on our operations. And we took a lot of precautions. This particular product really helped us from a testing perspective. There were a lot of protocols in place that allowed us to validate every step that we were taking. And that offered a lot of certainty for senior leadership to kind of sign off on the first experiment that we took here.

The use of AI coding tools by Salesforce employees has doubled the amount of features and codes shipped in 2026 Q1 (FY2027 Q1) compared to a year ago; Salesforce’s engineering team has been kept at 15,000 for the past 2 years because of the higher efficiency of the engineers through the use of AI coding tools

In Q1, AI coding tools enabled us to double the amount of features and codes shipped year-over-year, while simultaneously reducing incidents and defects…

…Srini is here at the table. He’s got what about 15,000 engineers, and you’ve had the 15,000 engineers for about 2 years, it’s been mostly flat, right? And I would say that the reason it’s been mostly flat is because we have been using AI to create more efficiency for our engineers. And especially this year, now with these new coding agents, we’re seeing even more dramatic capability.

The biggest way for Salesforce to monetise AI is by selling Flex Credits

The biggest way that we have to monetize AI is with customer-facing use cases by selling Flex Credits, by putting fuel in the tank 6 of the top 10 deals, 6 of the top 10 deals were AELAs, unlimited enterprise license agreement, where we threw in a bunch of Flex Credits and customers are deploying use case after use case, channel after channel.

Salesforce has been able to protect its margins despite investing in AI because it’s not hiring more engineers as a result of higher productivity; Salesforce’s headcount is growing only because of an expansion of the sales team, and that is because AI agents cannot actually sell; Salesforce’s margins are protected despite the company spending a lot with OpenAI and Anthropic

Srini is here at the table. He’s got what about 15,000 engineers, and you’ve had the 15,000 engineers for about 2 years, it’s been mostly flat, right? And I would say that the reason it’s been mostly flat is because we have been using AI to create more efficiency for our engineers. And especially this year, now with these new coding agents, we’re seeing even more dramatic capability. So that’s a key part of our margin story is that we’re not hiring more engineers. We’re not hiring more GA. We’re mostly expanding only in one area.

You can see head count has grown, but it’s mostly growing in Miguel’s area in sales because I think we all realize the one thing that we’re doing here with you selling and communicating that agents are not exactly doing that. They can qualify, okay? They can provide service. But in sales, we still scale because there are so many different parts of the market that we have to get to. So that will be a critical part of expanding our company, but at the same time, expanding our margins…

…It’s not that we’re not spending a lot with OpenAI. We are. We’re using their platform. We’re using Codex, their coding tool. We’re using Anthropic. We’re using their platform and their coding tool Cowork. We’re using both of these platforms.

Sea Ltd (NYSE: SE)

Sea’s management has used AI in Shopee’s search and recommendation systems to improve product discovery; Shopee has AI content tools for sellers to create better product listings, which has helped the purchase conversion rate improve by 14% year-on-year in 2026 Q1; AI-powered advertising personalisation and targeting contributed to Sea’s 80% advertising revenue growth in 2026 Q1; management is exploring an AI shopping assistant for buyers that can deliver personalised recommendations and cost savings; management is building an AI agent for sellers that can be a business advisor; the AI shopping assistant and AI agent for sellers are both in the early stages

We have taken a practical resource-oriented approach, embedding AI into our operations to drive better outcomes for our users and greater efficiency across our platform. This is already making a meaningful impact. AI-powered enhancements to our search and recommendation algorithms have led to better product discovery. Our AI-generated content tools are helping sellers create more compelling product listing. These efforts supported a 14% improvement in purchase conversion rate year-on-year in the first quarter. And AI-driven personalization and targeting helped to contribute to the strong year-on-year ad revenue growth we saw this quarter…

…For buyers, we are testing an AI shopping assistant that leverages purchase history and preferences to deliver personalized recommendations and optimize savings.  For sellers, we are building an AI agent that acts as a virtual business adviser, providing diagnostic and actionable insights on shop performance. Both are in early stages with plans to roll them out more widely over time.

Around 80% of Sea’s customer queries are now handled by its AI chatbot; AI has reduced Sea’s customer service cost per contact by 30% year-on-year in 2026 Q1 while maintaining satisfaction

Around 80% of customer queries are now handled by our AI chatbot. AI usage helped reduce customer service cost per contact by around 30% year-on-year, while maintaining high satisfaction rate.

Tencent (OTC: TCEHY)

Tencent has made significant progress in its Hunyuan large language model in the last 6 months; Tencent has overhauled its foundation model team and system and processes for pretraining and reinforced learning; management has moved away from from chasing public model benchmarks that can be gamed and has chosen to evaluate Tencent’s models with the latest exams, human tests, product feedback and in-house tasks; management launched Hunyuan 3 Preview in April; Hunyuan 3 Preview was designed to deliver comprehensive intelligence with cost efficiency; Tencent reduced Hunyuan 3 Preview’s inference costs significantly by designing inference together with the model; Hunyuan 3 Preview is already deployed across 131 of Tencent’s products, including Yuanbao, QQ, and WorkBuddy; Hunyuan 3 Preview has been ranked 1st on OpenRouter by token usage since April 28, even after its free period ended on May 8; the Hunyuan team is already working on a larger parameter model; Hunyuan 3 Preview is a smaller model, but is still very capable; Hunyuan 3 Preview is significantly better than Hunyuan 2 for agentic work; Hunyuan 3 Preview’s total token usage is at least 10x compared to earlier generations; Hunyaun 3 is currently not fully integrated into Weixin because it depends on Weixin’s own evaluation on what’s the best model for users; the adoption of Hunyuan 3 Preview in actual use cases has been much better than management expected 

Over the last 6 months, we have made significant progress on our Hunyuan large language model..

…We started the initiative by completely overhauling our foundation model team, centering around newly added elite AI researchers and engineers with deep expertise in large language models. Our new team is young, energetic and cohesive, enabling us to make progress quickly in this highly dynamic AI era. 

In February, we reengineered the system and process for pretraining and reinforced learning from the ground up. We rearchitected the infrastructure to support robustness, scalability and efficiency across pretraining, data and reinforcement learning. On data, we expanded our data set significantly and strengthened our data collection, cleansing and synthesis capabilities with a focus on data quality. On training, we upgraded the process for pretraining and supervised fine-tuning, and we scaled up reinforcement learning. And for evaluation, we’re moving away from chasing public benchmarks that can be gamed. Instead, we evaluate our model through the latest exams, human tests, product feedback and in-house tasks to see how the model actually performs in the real world.

In April, we launched Hunyuan 3 Preview. When we set out to build this model, the purpose was to build a cost-efficient and solid model for diverse applications and derisk scaling toward larger models. The core design principles behind Hunyuan 3 Preview was to deliver comprehensive intelligence and cost efficiency, optimizing it for real-world deployment. We moved beyond narrow expertise and towards comprehensive intelligence such as integrating reasoning, long context understanding, instruction follow, dialogue, coding and tool-use capabilities. And by codesigning inference with model, we’re able to reduce costs significantly so that the intelligence is economical enough to be used at scale. Hunyuan 3 Preview has delivered on these expectations.

The model has already become a leading reasoning model in China and has proven effective in real-world software engineering and other productivity agent tasks. Internally, the model has been deployed across 131 widely used internal products, including Yuanbao, QQ and WorkBuddy, providing valuable feedback and iterative improvement vehicle design process. And externally, Hunyuan 3 Preview has been well received by users and developers in real applications. It has ranked first among all models available on OpenRouter by token usage since April 28 and continued its lead even after its free period ended on May 8…

…Our Hunyuan team is already working on a larger parameter model, leveraging our infrastructure and learnings from Hunyuan 3 by aggregating bigger and better data sets and scaling more powerful reinforcement learning, we can strengthen the model’s contextual understanding, enhance its agent capabilities in areas, including coding and increase the model’s general intelligence. Through codesigning and collaborating with other Tencent product teams, we are optimizing data set selection and focusing reinforcement learning for high-value use cases…

…We have given a pretty comprehensive overview of Hunyuan 3. And as you can see from the prepared remarks, it’s more intelligent and it’s actually very strong in terms of reasoning despite being a smaller model. And at the same time, it has significant improvement vis-a-vis Hunyuan 2 on agent capabilities…

…The total token usage is actually at least 10x compared to Hunyuan, so that’s the clear indication that Hunyuan 3 is actually well designed…

…In terms of the integration into the Weixin workflow, I think it will be a step-by-step process. And Weixin itself actually sort of have been always using some part of their products, Hunyuan 2 and they upgraded already to Henyuan 3. And in some cases, they use different models and they evaluate different models and evaluate what’s the best model to use for their users, right? So as Henyuan 3 continue to be getting better and better, then they will be adopting more…

…if you look at how this is received in the actual use cases, it’s actually better than our expectation by quite a bit.

Tencent’s management thinks agentic AI is a breakthrough use case for AI; management thinks agentic AI first delivered value in coding through enhanced productivity and is now shifting to more workloads and occupations; management thinks Tencent’s apps, such as Weixin, Yuanbao, and more, are great avenues for users to control AI agents; in the future, management will enable AI agents to access Tencent’s Mini Programs as AI skills; management sees Tencent having a lead in agentic AI deployment through the leading DAU (daily active users) of WorkBuddy; Tencent’s agentic products, CodeBuddy and WorkBuddy, are still early in their lifecycle but currently have strong organic growth and high retention rates; the high usage of Tencent’s agentic products is a virtuous feedback loop for the company, as more usage leads to insights for product development, which leads to more agentic usage, and as agentic usage grows, token usage in Tencent Cloud also grows; management thinks the breakthrough of agentic AI as a use case is a very recent phenomenon

It has become increasingly evident that agentic AI represents a breakthrough use case after AI chatbots have become popular. Agents are more valuable in uplifting productivity from initial use cases supporting programmers in creating code, such as with our product CodeBuddy to now catering to a wider range of workloads and occupations such as with Claws and WorkBuddy. These breakthroughs were made possible by more powerful models and by the hardness infrastructure that allows models to utilize tools and act as interfaces that enable users to manage agents effectively.

Our platform inherently has many benefits of hosting AI agents as users can control AI agents through our communications and browsing interfaces such as Weixin, WeCom, QQ, Yuanbao and QQ Browser in addition to third-party applications…

…And in the future, AI agents will be able to access our Mini Programs ecosystem using Mini Programs codes as AI skills.

Tencent has established an early lead in agentic AI deployment evidenced by the leading DAU of our product, WorkBuddy. While early in adoption cycle, CodeBuddy and WorkBuddy are already achieving strong organic growth and high retention rates among active users and paying users. The high time spent and high-frequency interaction with AI agents among early adopters act as a virtuous feedback loop to Tencent, enable us to identify and provide complementary software and services, which in turn drives increased AI agent usage among a broader enterprise and prosumer user base. As users utilize more AI agents for more complex tasks, paying user conversion increases, resulting in rapid growth in token usage on Tencent Cloud in recent weeks…

…The upturn in sort of productivity AI is really something that’s happened not in the last few quarters or even last few months, but last few weeks. And I think that’s true globally actually, that really, it’s since late in or since the end of the first quarter that the Agentic AI has broken through in terms of its ability to create code, in terms of its ability to make people more productive.

Tencent’s management believes Tencent Video has competitive advantages in creating animated series, partly because of the use of generative AI for storyboarding and producing animation

We believe Tencent Video possesses competitive advantages in creating animated series, including our ability to cross over IP from China literature and our games into animated IP and our use of technology tools such as Unreal Engine and generative AI for storyboarding and producing the animated content. Tencent Music subscription revenue increased 7% year-on-year, driven by growth in ARPU and subscribers.

Tencent’s management has improved the content recommendation model for video accounts, which has led to a 20% year-on-year increase in total time spent on video accounts; management has upgraded the developer toolkit architecture for Mini Programs to enable users to better leverage AI plug-ins; Weixin Search’s query volume was up 25% year-on-year in 2026 Q1, driven by foundation model powered ranking and broadening AI search coverage to include image-based queries

We scaled up the number of parameters and enhanced the algorithm for video accounts content recommendation model, enabling deeper understanding of users’ interest to recommend more personalized and relevant content and total time spent on video accounts increased over 20% year-on-year. For Mini Programs, we’ve upgraded the developer toolkit architecture so users can better leverage AI plug-ins, including CodeBuddy to create and debug Mini Programs…

…Total query volume on Weixin search increased over 25% year-on-year, benefiting from foundation model powered ranking and broadening AI search coverage to include image-based queries.

Tencent’s management sees AI being really helpful for game production in areas such as accelerating 3D asset production and animation, improving the player experience, and delivering better graphics; the use of AI in game production can be directly revenue-generating, and management has seen this happen; management sees Tencent as a global leader in utilising generative AI to improve game production; management’s objective with generative AI in the games business is to speed up content creation and generate incremental revenue; management is not intentionally using AI in the games business to expand margins, even though operating leverage should happen in the games business if AI is applied correctly to boost revenue

AI provides increasingly helpful tools, facilitating our game developers to deliver more content and enhanced experiences. Currently, AI for games is most beneficial in areas, including accelerating 3D asset production and animation, enriching player experiences with intelligent in-game guides and delivering more realistic graphics via AI rendering techniques…

…Generative AI enables us to produce more content faster. And that content is, in some cases, to enhance the overall player experience. But in some cases, it results in direct monetization. For example, if the content is a virtual outfit. And so that’s what we are doing, and that’s what we are seeing. And we think that we’re a China leader and to some extent, even more so a global leader in terms of deploying that capability and achieving that benefit. And the objective at this point is really faster content creation and incremental revenue generation. We’re not prioritizing margin expansion per se. It’s more that as we deliver the revenue uplift that we’re seeing and if we can keep headcount fairly stable, then I suppose mathematically, that combination would tend to result in higher margins over time, but that’s sort of a happy output rather than the intention of the process.

Tencent’s AI Market Plus automated campaign management solution, powered 30% of total advertising spend; management has upgraded Tencent’s runtime advertising recommendation models with a unified transformer-based architecture; Tencent’s video accounts ad impressions grew rapidly year-on-year in 2026 Q1

Our automated campaign management solution, AI Marketing Plus powered around 30% of total marketing services spending from advertisers with us in the quarter. We upgraded our runtime advertising recommendation models with a unified transformer-based architecture. This upgrade provides deeper understanding of user context and the intent while balancing model complexity with system efficiency. By inventory, video accounts ad impressions grew rapidly year-on-year, supported by increased total time spent video views and ad load. We released more inventory of rewarded ads, which deliver high click-throughs for advertisers.

Within the Fintech and Business Services segment, Business Services revenue grew 20% year-on-year in 2026 Q1, driven by higher demand and a better pricing environment for cloud services; Tencent Cloud benefited from AI-related demand across GPUs, CPUs, and storage; management had upgraded Tencent Cloud’s AI agentic solutions, which led to rapid usage growth and token monetisation; Tencent Cloud’s international business increased revenue by 40% year-on-year in 2026 Q1; Tencent Cloud finally has sufficient GPUs to serve all the external demand it’s seeing; previously, management had prioritised Tencent’s internal AI use cases for its AI compute but newer AI compute capacity will be focused on meeting external demand for Tencent Cloud

Turning to Business Services. Revenue in the first quarter grew 20% year-on-year, driven by increased demand and better pricing environment for our cloud services alongside rising technology service fees generated from mini shops e-commerce. For Tencent Cloud, AI-related demand contributed to increased revenue year-on-year across GPU, CPU and storage. We upgraded Tencent Cloud’s AI agent solutions with proprietary security infrastructure, skill hubs and interfaces, contributing to rapidly increasing usage and initial token monetization. Tencent Cloud’s international business grew its revenue over 40% year-on-year as we expanded our global footprint and captured demand for our Platform-as-a-Service solutions, including media processing services and TDSQL cloud database…

…For Tencent Cloud, where until now, we actually haven’t had sufficient GPUs to begin to service the external demand, the KPIs will be more revenue and market share related…

…We’ve already made the choice and paid the price in that we have prioritized a multiplicity of internal services ahead of Tencent Cloud…

…And the reason why we have been able to support all of these at once is because we have not been active in leasing out GPU capacity in Tencent Cloud. Now looking through the rest of this year, as the supply of China design GPUs progressively ramps up, then we’ll be remedying that situation, and we will be making more capacity available in Tencent Cloud and consequently driving up Tencent Cloud’s rate of expansion. But that’s where the trade-off has been made that we have been consciously late to monetize the AI opportunity through Tencent Cloud because we’ve been simultaneously supporting a number of AI initiatives internally.

Tencent’s operating capex in 2026 Q1 was up 18% year-on-year and up 84% sequentially because of higher server investments; non-operating capex was down 36% year-on-year (was RMB 1.1 billion in 2025 Q1); free cash flow was up 20% year-on-year, and up 67% sequentially

Operating CapEx was RMB 31.2 billion, up 18% year-on-year and 84% quarter-on-quarter as we accelerated investment in server infrastructure. Nonoperating CapEx was RMB 0.7 billion. Free cash flow was RMB 56.7 billion, up 20% year-on-year, driven by growth in games, gross receipts and advertising billings, partly offset by higher server infrastructure and compute spending. On a Q-on-Q basis, free cash flow was up by 67%, reflecting seasonally higher game gross receipts and the timing of certain seasonal accounts payable settlements, partly offset by higher server infrastructure and compute spending.

Tencent’s management thinks it’s still too early to determine the impacts that agentic AI can have on the e-commerce industry, but they don’t see agentic AI as a risk to Tencent’s advertising business

[Question] With agents increasingly potentially replacing the traditional click-throughs on the web pages and also the apps, could management share your view on the future advertising pricing and also the resulting impact on advertiser budget?

[Answer] It’s certainly more of an issue potentially for e-commerce companies than it is for us because users actively choose and desire to spend their time watching short videos or listening to music or consuming content or chatting with their friends versus generally speaking, when users spend time on e-commerce, it’s because they’re trying to find the lowest price. It’s not because they necessarily enjoy that process. So to the extent that AI agents play a bigger role in the future in facilitating price comparison, then it’s possible that users will spend less time on e-commerce sites and be less exposed to ads than they are today, while the AI agents can scan infinite listings and therefore, not influenced by ads the way that human beings with a finite attention span are influenced. All of that said, there’s been many prior iterations of price comparison services, including search engines and the big e-commerce companies are generally thrived despite the existence of those price comparison services. So I think it’s premature for us to sort of have a definitive view at this point on how it will affect our friends in the e-commerce industry. But we don’t see it as a primary risk for Tencent.

Tencent’s management continues to see Tencent increasing capex substantially in 2026, especially in 2026 H2, to meet AI-related demand; Tencent’s AI-related capex in 2026 will be focused on AI chips designed by Chinese companies; the KPIs management is looking at to determine the ROI (return on investment) of AI-related capex includes (1) revenue and profit for the advertising and games businesses, (2) intelligence, usage, and token consumption for the new AI products, and (3) revenue and market share for Tencent Cloud; Tencent Cloud finally has sufficient GPUs to serve all the external demand it’s seeing; in management’s eyes, the ROIs on AI-related capex have both near-term and long-term components, with advertising being a near-term example and Hunyuan being a long-term example; previously, management had prioritised Tencent’s internal AI use cases for its AI compute but newer AI compute capacity will be focused on meeting external demand for Tencent Cloud

We are seeing increased demand, both from internal products as well as from external users of our model for our AI-related services. And we had previously guided that we’ll be increasing CapEx this year versus last year, and we’re now more affirmative, more confident in that guidance. And we and you should expect a substantial increase in CapEx, especially in the second half of this year as more China designed ASICs become available to us month by month through the year…

…At a high level, for our existing activities such as advertising and games, the KPIs would be more revenue and profit related. For our new AI products, the KPIs would be more capabilities, how intelligent is our foundation model and usage, how much token consumption is happening on world body related. And then for Tencent Cloud, where until now, we actually haven’t had sufficient GPUs to begin to service the external demand, the KPIs will be more revenue and market share related…

…AI includes a range of sort of shorter cycle investments as well as longer cycle investments. And so if we buy GPUs and we deploy them into our ad tech, then that’s a relatively short-cycle investment. The GPUs yield better targeting, higher click-through rates and higher revenue and profit on a pretty accelerated basis. On the other hand, when we deploy GPUs into our Hunyuan foundation model, that’s something which we view as important for our franchise and where we’re taking a longer-term view…

…We’ve already made the choice and paid the price in that we have prioritized a multiplicity of internal services ahead of Tencent Cloud…

…And the reason why we have been able to support all of these at once is because we have not been active in leasing out GPU capacity in Tencent Cloud. Now looking through the rest of this year, as the supply of China design GPUs progressively ramps up, then we’ll be remedying that situation, and we will be making more capacity available in Tencent Cloud and consequently driving up Tencent Cloud’s rate of expansion. But that’s where the trade-off has been made that we have been consciously late to monetize the AI opportunity through Tencent Cloud because we’ve been simultaneously supporting a number of AI initiatives internally.

Tencent’s management thinks society is still at a very early stage in terms of AI diffusion; management thinks many new kinds of products will appear, beyond agentic AI; management believes that it’s much more important to find high-value use cases in AI as compared to focusing on gathering users because AI is expensive to produce for each user, unlike the internet which supports infinite scaling of users; management thinks building a subscription model for consumer AI in China is very difficult compared to the USA because the USA’s living standards are high and its population has a habit of paying high prices for subscriptions; management thinks the consumer AI market in China will not be a winner-takes-all market; management thinks it’s still early days for monetisation of AI in e-commerce and advertising even in the USA

In terms of how we think about the different products, we felt this is actually sort of a very early stage in terms of AI diffusion, right? And we would see many different products coming up going forward. Initially, it was chatbot and everybody felt chatbot is actually the king of the product. And then suddenly, you have a coding that came up and this becomes sort of even more eye-catching and less significant use case because it’s very high value, right? And now we are seeing sort of agentic capability proliferating right? And I think that would actually allow AI to be diffused to different industries, and you have many different agents coming up, which can help you to do work, right? And there’s going to be new products coming up. So I think that would continue to propagate.

And I think to some extent, right, you actually have to — in the AI world, you actually have to find a high-value use case as opposed to sort of just purely focused on DAU because the difference between the AI revolution and Internet is that this is about intelligence and intelligence manifest its value in sort of how much people are willing to pay for it. And at the same time, the intelligence is not free, right? In the Internet world, you basically sort of have mostly existing information. And then you also create some new information and content, but then that’s a fixed cost and then sort of the variable cost for delivering is actually very small, right? You only have to pay for bandwidth. and the compute sits on people’s devices, right? And as a result, you can almost like go for infinite scaling. But in this case, right, every single delivery of a DAU actually cost you quite a bit. And as a result, you can’t just apply the same logic as Internet and apply it to AI. And I would say the ability to find high-value use cases is going to be as important, if not more important than just sort of blindly get a lot of use DAU and user time…

…In terms of the 2C monetization, I would say it’s actually not easy, right? If you look at global standard in the Western market when the paid service is actually very well penetrated and the living standard is actually very high. So the subscription price in the Western market is multiple times of what the equivalent service in China is like, be it music service or be it video service. The paying penetration is probably in the single digit, right? And — and when you sort of applied it to China, I think the subscription model is not going to be that big for the China market…

…I think the more important implication is that when you have to have payment to support a service, then most likely the service is not going to be a winner take-all business. It would basically sort of be supporting multiple players who would have a share of the market and each one of them would sort of have some kind of users and some share of subscriptions…

…When we look at e-commerce or advertising as a way to monetize, I think it’s also very early for even the U.S. players where the eCPM is actually much higher, right? The leading player has not been able to roll out very robust advertising model.

Tencent’s management sees Tencent as having many more flagship internal use cases for AI as compared to the hyperscalers in the USA

And so I think most big tech hyperscale companies with cloud businesses have one flagship internal use case where they’re allocating a large number of GPUs. We have multiple flagships. We have the foundation model. We have agentic developments within Weixin. We have — support. We have the AI deployment for advertising for games, now also for the WorkBuddy and CodeBuddy use cases. 

Tencent’s management thinks policy restrictions from the USA and limited manufacturing capacity in China are the reasons why there was a supply shortage of GPUs in China; the GPU supply shortage in China is now easing because there’s more capacity from China fabs and other foreign fabs to manufacturing China-designed AI chips; management does not see any supply shortage in China for CPUs and other networking chips; management is seeing that the suppliers of CPUs and networking chips are not raising prices indiscriminately over the short-term; management is seeing that the suppliers of CPUs and networking chips are negotiating long-term contracts with customers, and they are looking for a variety of customers 

The reason why there’s been a GPU bottleneck that’s been much more pronounced in China than elsewhere is a combination of policy restrictions on certain foreign design GPUs being brought into China and then the China design GPUs facing limited fab capacity within China. And as a result, the country has really been short of GPU or ASIC capacity. And that’s now being addressed because the China designed ASICs are seeing more supply from fabs within China as well as more supply from fabs in neighboring countries.

But by contrast, we haven’t faced those sort of artificial additional constraints CPU or networking chips. We’ve been a big buyer of CPU and networking chips for many years before GPUs became such a big presence in data centers. We have very long-term relationships with the companies that supply the CPUs and supply the networking chips. And on their side, while one might think that these suppliers would be sitting back and just selling at the highest possible price into the spot market, that’s not actually the reality. The smart suppliers are taking very conscious 3- to 5-year forward views and negotiating long-term agreements in order to give them certainty of their revenue outlook over the next 3 to 5 years. And when they’re deciding with whom to sign those long-term agreements, they’re looking to work with a number of partners, not just a single partner, and they’re looking to work with partners who have been there for many years already and will be there for many years to come and ideally with partners whose demand they believe will grow substantially over time. And happily, we fulfill all of those criteria. We’ve been a big customer for the Intel and AMD and so forth for many years. We’ve been progressively growing our volume with them for many years, and they believe it will continue to progressively grow our volume for many years to come.

Veeva Systems (NASDAQ: VEEV)

Veeva’s management sees the company changing from an industry-specific application provider to an industry-specific application and agent provider; management wants Veeva to support both human users and agentic users; management is seeing pharmas leaning into a new technical architecture called MAAP (models, agents, and applications); management sees pharmas wanting to see AI in Veeva’s applications; management is thinking of building very specific agents that would go the last-mile and automate standardised actions for pharmas, and management thinks Veeva can lead in this area

Veeva is moving from an industry-specific application company to an industry-specific application and agent company. In our first chapter, we became the leader in applications. In this next chapter, we intend to also lead in industry-specific agents. This includes agents that support human users, as well as agentic labor, which represents an entirely new market and type of application user…

…[Question] As we think about pharma appetite for AI applications more broadly, I’m curious what areas you think they lean into first

[Answer] It’s not that they’re thinking mainly about transition from applications into AI applications. What they’re really leaning into is this new technical architecture, we call it the MAAP architecture of Models, Agents and Applications. So the applications that they get from Veeva, they’re looking for them to be more efficient, to have AI in there and help the users. What they really want to get to be is an agentic biopharma so that agents can do a lot of the work. And so the humans can do the more higher value work…

…Let’s just say there’s 100 million documents collected from clinical research sites around the world every year having to do with clinical trials, they have to be checked for quality and they have to be sorted into the right places. That’s work that agents can do, it’s difficult, specific work, but we can make agents that are very specific on that. Agents that take in a bunch of free text via e-mail or other channels and have to sort it out to see, is this a product complaint? If so, how to handle that? And categorize that? Or no, this is an adverse event. This is the issue with a medicine-making somebody potentially ill, okay? Well, what is that illness? Is that a headache or a throbbing headache? How serious is that? Is that involved in the clinical trial? What drug is that involved with? We will make agents to do that and do those very standard things. And this is an area where I’m enthused because Veeva can lead. 

This is where — just like for cloud applications, you got the very specific industry-specific cloud applications could add tremendous value if you went to the last mile and solve the thing. In industry-specific agents, agentic labor, we may be able to go the last mile and make specific agents that just do the thing for life sciences because we’ll go to that last mile and make it work, we may make agents that are better safety case processors and more reliable than humans.

That’s a heck of a lot of work, but we have a structural advantage to do that because we’re deep in life sciences, we have a consulting in life sciences, and we have the applications that those agents can use, it’s the same reason why Claude is getting very good at Claude Code because they have the agent, the coding agent and they have the model, and they have 2 layers. We don’t have a model we use, but we have applications and the agents. So that is a structural advantage.

Veeva’s agentic products will have different pricing depending on the type of agent

Pricing and packaging also vary by agent. Some agents are charged by usage, while others are part of a fixed-price subscription license.

Veeva recently acquired Ostro, which provides conversational AI for brands to provide patients and doctors with immediate, compliant answers; management believes Ostro can be a significant revenue driver for Veeva; Ostro had no material impact on Veeva’s financial results in 2026 Q1 (FY2027 Q1), but accounted for 25% of headcount growth; the buyer of Ostro’s product is the biopharma company, but the user is a healthcare professional or patient; Ostro is a brand engagement platform; management thinks it’s really hard to do what Ostro is doing; management thinks Ostro will be a really significant acquisition for Veeva; management has organised Ostro smartly so it can retain the speed of a startup

In March, we acquired Ostro, the leader in conversational AI for brands to provide patients and doctors with immediate, compliant answers through an easy-to-use chat experience. Ostro operates as a startup within Veeva and is now an important part of our Commercial Cloud. Things are going well, revenue and pipeline are growing as anticipated, and we have an ambitious product roadmap. We believe Ostro can be a significant revenue driver for Veeva and transformative for the industry, fundamentally changing how patients and doctors get information…

…We also acquired Ostro in the quarter, which had an immaterial impact on Q1 financial results and accounted for about 25% of net headcount growth…

…The buyer of Ostro is the biopharma company, the user of Ostro is the health care professional or the patient. So it’s a brand engagement platform for biopharma companies to help HCPs and patients ask questions and get answers instantaneously and do that in a compliant way. That’s very, very hard to do. It’s hard to do that at scale. It’s hard to do it in a compliant way, and that’s exactly what Ostro does…

…It’s going to play a bigger and bigger role in Commercial Cloud over time, and we see it as a really significant acquisition and a potential long-term growth opportunity for us…

…In an operating model for Veeva, we have a notion of the start-up models in the core models. And in the core models, we’re organized functionally like the central sales team, engineering team, things like that. In the startup model, it’s all fully contained under CEO, and we use that either when the market is very different or when the product really needs to evolve. So Ostro is in the start-up model. Everybody who works on Ostro is fully reporting to the CEO of Ostro. There’s guidance and help from other functional areas of Veeva, but it’s — and they’re certainly inroads like, okay, Ostro doesn’t have to use their own master subscription agreement anymore and all that type of stuff. So it operates as a start-up, they can retain its speed, but it has a really smooth ramp up.

Veeva’s management will soon release standard agents and the ability to build custom agents for all Vault applications; management will soon release Veeva Falcon, an agentic platform and for clinical, regulatory, and safety; Veeva Falcon is on track to be released in November 2026; Veeva Falcon will be the first agentic solution for the industry; management recently talked about Veeva Falcon to Veeva’s customer base, and it was very well received; management envisions Veeva Falcon to be replacing jobs that humans used to do; the presence of Falcon means Veeva’s applications need to be headless; agents within Vault applications are meant for human users and to improve the productivity of human users; Veeva Falcon is not a platform for pharmas to build custom agents; the platform for pharmas to build custom agents would be Vault AI or other 3rd-party agentic platforms; nobody is asking for the kind of solution Veeva Falcon presents, but management believes it’s the way to go; management is very positive on Veeva Falcon; Veeva Falcon will be tackling the simplest and highest volume labour, specifically the processing of documentation related to clinical trials, and processing safety cases; management’s still unsure how Falcon will be priced, but they’re toying with the idea of pricing Falcon on a per document or per case basis; management sees Veeva Falcon as being completely accretive to Veeva; management expects small biopharmas to be among the first customers of Veeva Falcon because the small biopharmas are running all their processes on Veeva; Veeva Falcon reports directly to Veeva’s CEO; the kind of labour Veeva Falcon is designed to replace does not involve CROs (contract research organisations), and Veeva Falcon could in fact even benefit CROs

In August, our standard agents and the ability to develop custom agents will be generally available across all Vault applications. 

We also announced Veeva Falcon, our agentic platform and standard agents that provide agentic labor for clinical, regulatory, and safety. Many of the processes in these areas are ripe for automation. We are on track with our plan to release Falcon for early adopters in November. Delivering agentic labor in this area will be a first for the industry and the quality and control requirements will be significant. Falcon is a disruptive technology trying to solve a very hard and valuable industry-specific problem. It’s an outstanding fit for Veeva…

…Veeva Summits bring the industry together and are key to driving customer success and product excellence. It was a milestone event as we talked about Falcon to a broad audience for the first time. Falcon was very well received, and customers are excited about the potential to lower costs and increase speed in drug development…

…Falcon specifically is at the agent layer and that’s agentic labor. So fully replacing parts — jobs that people used to do. People who used to do these jobs using our applications, now will deliver the agentic labor to do that. So it’s a big new area for Veeva. It’s something we haven’t done before, and that’s why it’s disruptive. Those agents have to become users of our applications, which means our applications have to become very good in operating at a headless manner. Now at the same time, we have agents inside of the Vault applications. So that’s Vault AI inside of the applications. That’s where when people are actually using the application because there’s definitely things that people still need to do in our applications, that’s where the AI agents can help them do it more efficiently, much like you might use ChatGPT or Gemini at your work, okay, that helps you do it more efficiently…

…For Falcon, the actual effort there is taking the path less traveled. So that’s a platform for us to build and operate standard agents to actually solve the problem for the industry. So it’s not really a platform for customers to develop their custom agents. For custom agents that live inside of our applications, of course, they can use Vault AI for that. For custom agents that are outside the applications, there are many agent building tools, and they will dip into the Veeva applications operating in a headless manner…

…In 2012 for the first time we laid out our first visions for Development Cloud. 2014, they got sharper; in 2016 that really became apparent what we were doing. We’re trying to simplify and standardize and integrate the tech of the development area of life sciences. That’s not anything that anybody asked us for, right? That’s the vision that we have, and that’s not anything that anybody has tried to do before. Falcon is the same thing. It’s the same magnitude of disruptive innovation. It’s not giving tooling to people to design agents. This is to designing and operating the standard agents for the industry rather than the industry having to hire humans for those specific jobs…

…I think Falcon is just going to deliver value. It’s going to be great revenue for Veeva, but it’s going to deliver value far above and beyond that for the industry, and that’s going to allow the industry to grow. It’s a disruptive thing. It’s not an incremental thing or a tool…

…[Question] How are you deciding which labor roles to address or to attack with Falcon agents? 

[Answer] I think the most right there ones are actually the simplest and the highest volume. And actually, when you look inside of life sciences, those are the areas where they have a tendency, some of the companies to do some outsourcing today already. So that makes it also — they’re used to outsourcing. Of course, they would outsource that to humans. In Falcon, the first ones we’re looking at are processing of documentation involved with clinical trials, specifically the stuff that comes from clinical sites, the millions and millions, hundreds of million, tens of millions of documents that come from research sites. They need to be collected, inspected for quality, categorized, the metadata pulled out of them, filed in the TMF the right way. So that’s one, the intake and control of documents. Another one is the safety cases, the safety cases that come in, the triage and the categorization and the collection of the safety cases. So those are the 2 main ones, we’ll also take on regulatory health authority correspondences because that’s another high-value one, and there’ll be more…

…[Question] How are you pricing Falcon?

[Answer] You can imagine most likely that Falcon will be charged by the document, most likely. We haven’t fully decided that. You can imagine that safety will be most likely charged by the case. So that’s how that is…

…[Question] On the Veeva Falcon. You’re mentioning the displacement of potential roles at these larger firms. I’m just wondering, is there anything that you would consider timing-wise from an economics perspective. So let’s say, these roles were to move in another direction? Do you think it could potentially cannibalize some of the revenue that you get from those customers?

[Answer] Definitely all accretive because this is not a market we address today. We don’t play in that market today. This is not type of labor or work that we supply. So it’s definitely going to be accretive. And these agents, they need a system of record. You can’t operate them without a system of record. So it definitely doesn’t cannibalize the systems of record…

…Veeva Basics, small biotechs. We continue to win a lot of those that are going on Veeva Basics. And by the way, those will be some of the first consumers of things like Falcon and our other AI solutions…

…Basics are smaller companies, very nimble. Also, they’re running not only our products, but they’re running our processes. So they have an absolute standard configuration of Veeva, where they’re running our processes. So we don’t have to wonder how they have configured Vault or MAAP Vault or done this Vault or with that Vault. They’re running absolute — let’s say, we have over 100 Basics customers in the clinical area, their configuration is exactly the same. How they’re using product is exactly the same. And we operate those systems in a way for the customers. So that’s — if we have our agent working on for one Basics customers, it will work for them all. With the enterprises, the larger companies, our agents have to be a little more adaptive. They have to first go through a phase of, okay, understanding how that customer is using that Vault, testing it out. Okay, I’m going to classify these documents that they’ve previously classified. Do I get the same of what they got. And if so, that’s good. If not, what happened there? Basics is just going to be smoother, very, very smooth…

…Falcon, for example, reports directly to me. This is our first step into digital labor. You can’t — you have to operate that effectively, back when we were the CRM company, way back when before we went public, Vault was this tiny little thing that reported directly to me. Falcon is like that…

…In terms of where can agentic labor play and what can agents do. The best places to do are high-volume repetitive work that actually gets outsourced. So that type of work actually it’s not so much the CROs, it’s other specialized labor providers that do that. So I think this could actually be beneficial for the CROs because that — we can do that lower volume work, which is generally done by the pharma company or a specialized outsourcer. We can do that cheaper, faster, better. That will hopefully allow pharma companies to run more trials, and that’s where the higher margin work is for the CROs.

Veeva’s management believes AI will change the commercial model for pharmas, and Veeva is well-positioned to bring the right solutions; Veeva’s Agentic Call Report in Vault CRM and Ostro help biopharmas capture compliant Commercial Evidence for the first time at scale; there are currently 10 customers live with Vault AI for PromoMats’ Quick Check Agent; management will be focused on commercial content for AI investment to solve the MLR (medical, legal, and regulatory) review bottleneck; management thinks agents on the commercial side will not be a full replacement for field salespeople

While it is early days, AI will fundamentally change the commercial model. This represents a major transformation, and we believe Veeva is well-positioned to help the industry bring the right medicines to more patients through new and better ways of working with AI. With major innovations like the Agentic Call Report in Vault CRM and Ostro’s conversational AI on brand websites, biopharmas are now able to capture compliant Commercial Evidence for the first time at scale. It’s a real breakthrough that allows companies to gain insights and take actions that were simply not possible before AI…

… I am also excited about the progress of Vault AI for PromoMats. We have 10 customers live today for Quick Check Agent, across both small and large biopharma. Commercial content will be a key area of AI investment as we look to solve the MLR content review bottleneck for the industry…

…In commercial, that won’t be — agentic labor there will not be — you’re not going — you’re going to have helper agents that help the field teams do things, but I don’t think you’ll have — you will — you’re not going to replace a field person. That’s about managing relationships, things like that. There may be some things in commercial for example, there’s a medical legal regulatory process that is burdensome and expensive and occupies many parts of people’s time in Life Sciences. I think that can largely be automated, 70% or more with the right agents over time. But the actual field person, I think, it’s going to augment them. 

Veeva’s management expects immaterial AI revenue and margin-impact in 2026 (FY2027)

For this year, our overall expectation had been for AI to be fairly immaterial outside of Ostro. And we’re really focused on getting AI live in all of our customer areas, getting the product excellence, getting to customer success. It starts with that deep value creation for customers. So on the margin side, you also don’t see a material impact, Craig. And in Vault AI, where its usage based on tokens. I think we have a pretty good understanding of what that dynamic looks like, and it’s factored into our guidance. But I don’t expect there to be a material impact on margins driven by AI this year.

Veeva is using AI throughout the company, including general-purpose tools and specific tools; Veeva is using Claude Code from Anthropic and finding great efficiency, which has led to Veeva needing to hire less; management thinks the productivity from AI tools, and the need to hire less, outweighs the cost of tokens

We use AI throughout the company, we’ve got general-purpose tools and then also specific tools and major functional areas. Probably the most significant place for using it is around the product because that’s where we spend the most. And so you heard Peter mention earlier, in product engineering, we use Claude Code, and it’s come a long way. So we’re seeing great efficiency from that tool. And I think in general, that means we’ll hire a little less than we would have and accomplish more than we would have and go a little bit faster. But for us, it’s more about productivity and the combination of hiring a little less, accomplishing a little more, we think easily outweighs the token cost, and that’s all factored into our guidance.

Wix (NASDAQ: WIX)

Base44 has reached $150 million of ARR, or annualised recurring revenue (was $100 million in March 2026); Wix Harmony and Base44 can now be accessed within popular AI chatbots such as ChatGPT and Claude; management recently released Superagents inside Base44; Superagents allow users to build and deploy autonomous AI agents without coding; Superagents can run continuously in the background without any manual intervention; Base44 users can interact with their Superagents through popular messaging apps such as WhatsApp and Telegram; Base44 now has better app-design tools; Figma is now integrated with Base44; Base44 is currently incurring significant AI processing and compute costs as usage ramps, but management believes the costs are front-loaded as new Base44 users tend to consume more AI inference bandwidth during their initial build phase; Base44’s user behaviour and cohort quality look positive, with retention improving, and monetisation steadily increasing; management has been lowering inference costs in the core Wix business through optimising 3rd-party models, open source models, and building a proprietary LLM, and management expects to apply the same strategy to Base44’s AI costs; use-cases in Base44 remain wide, but management thinks specialisation will happen over time; some use-cases seen in Base44 are also applicable for business owners on Wix

Base44, which is now the leading AI-powered application creation platform in North America (per Similarweb data) with ~$150 million of ARR as of May…

…Both are now accessible within ChatGPT, Microsoft Copilot, and Anthropic’s Claude. Users can type “@Wix” or “@Base44” in these platforms, describe their idea, and a full website or application is created in conversation and managed there too, without any context switching…

…In March, we unveiled Superagents, a new experience inside Base44 that lets anyone build and deploy their own autonomous AI agent simply by describing what they want it to do. Base44 automatically builds the underlying workflows, connects the necessary tools, and deploys the agent. No coding, no configuration and no infrastructure to manage. Once deployed, Superagents run continuously in the background, responding to triggers, schedules, and real-time events, executing tasks without the need for any manual intervention. It can connect to third-party platforms and applications, remember preferences and priorities across conversations, and become more effective over time. Users can also interact with their agents directly through iMessage, WhatsApp and Telegram – wherever they are already messaging…

…Base44 now includes a fully rebuilt set of tools for shaping how an app looks and feels. Users can set colors, typography, and overall style across their entire app from one place, with any change carrying through automatically. Images, documents, and data files can be uploaded to an asset library or generated on the fly, and pulled into any app directly from the visual editor or chat…

…Design screens in Figma, paste the frame link, and Base44 builds a working app on top of it. The layout stays intact, and users go straight from design to a live app…

……Creative Subscriptions non-GAAP gross margin was 80% in Q1’26, down from 84% in Q1’25. Creative Subscriptions non-GAAP gross margin in our core Wix business was stable in the first quarter as AI costs remained minimal while we carefully controlled costs as we scale our platform, particularly Harmony… Creative Subscriptions non-GAAP gross margin was driven by accelerating contribution from Base44, which is incurring significant AI processing and compute costs as demand and usage continues to ramp. We believe these AI costs to be front-loaded as new Base44 users consume more AI inference bandwidth during their initial build phase…

…We also saw positive signs in the user behavior and cohort quality of Base44. Retention is improving as more users are choosing annual subscriptions, either through new purchases or renewals. Monetization is also steadily increasing, resulting in stable TROI even as marketing spend stepped up in the first quarter…

…We have been lowering inference cost of users by optimizing third-party AI model usage, leveraging open source models and most recently building our own LLM to power Harmony. As we apply this strategy to more of our products, particularly Base44, we believe that the large majority of these AI costs will be firmly in our control…

…About the Base44, I think we’re happy actually to say that we’re still using — we’re seeing a very wide variety of use cases. And it’s really — some of it is personal uses, some of it is solopreneurs, some of it is small businesses. And we think that there’s going to be more and more specialization that’s going to go and happen throughout the platform over time as we understand what is — where there is differentiation happen between those different use cases and where everyone can benefit from the generalized platform…

… I think there’s another opportunity that is very interesting, which we’re seeing is that some of the more small business-oriented use cases can also be relevant to applications needed by business owners that on Wix.    

Wix’s management thinks the differentiation for website builders is not in the AI models, but in the experiences built around the models, and Wix has the necessary knowhow

As powerful AI models become increasingly accessible across the industry, I believe differentiation will come not from the models themselves, but the experiences built around them. The real value lies in the capabilities layered on top of the models: the backend infrastructure, agent orchestration, tooling, integrations, and everything that comes after turning a prompt into a website or app. With our deep infrastructure, world-class distribution, product expertise and years of technological innovation and market intelligence and understanding, this is where I believe Wix is uniquely positioned to win in today’s AI world. 

Wix’s management recently built Wix’s first proprietary large language model (LLM) that’s designed to power Wix Harmony, the company’s first-of-its kind website builder blending visual editing with vibe coding; Wix’s proprietary LLM is faster and has fewer errors when building websites; having its own LLM means Wix can move faster, and operate with lower inference costs; Wix is currently experiencing only tiny benefits from using its own LLM, but management expects the company’s own LLM to drive the company’s profitability over the long run; the LLM is just the first in a broader portfolio of AI models that Wix will release; Wix can build websites with its own LLM at just 5% of the cost of 3rd-party alternatives; management thinks the AI advantage in website building belongs to the most specialised model; Wix Harmony is now in all of Wix-supported languages; Wix Harmony and Base44 can now be accessed within popular AI chatbots such as ChatGPT and Claude; AI has made creating a website easy, but the real work is done after the website is published; despite having its own LLM now, Wix still has the flexibility to use the best 3rd-party models when appropriate; Wix Harmony was rolled out to the company’s main geographies in late-January 2026; management thinks Wix’s own LLM could eventually be used for Base44, but there’s no exact time line; Wix spent only a small sum of money to train its own LLM, so ongoing training costs will also be reasonable 

We recently built our first proprietary LLM, purposefully designed to power Wix Harmony – a significant milestone in our innovation journey and a project I am personally very proud of. Thorough A/B testing is showing that our Wix-built model is faster while resulting in fewer errors and significantly better results when applied to building Wix Harmony websites. Having our own model means that we can accelerate the cycle of improvement, which we believe creates a continuous flywheel for our platform that general-purpose models can’t replicate with success.

Importantly, building and relying on our own LLM means significantly lower inference costs that sit completely within our control as we scale the Harmony platform. While the margin benefit is small today, we expect this model to drive profitability over the long term. We expect this to be just the first in a broader portfolio of proprietary AI models across a number of use cases as they become increasingly central to our product roadmap…

…Big LLMs optimize for broad scopes and with limited feedback; we’re optimizing for one thing, every day, with millions of real users building real websites. This gives us full control over our roadmap, reduces dependency on external vendors, and significantly accelerates our iteration cycle. The result is a model that’s faster and more accurate, and we will be able to create beautiful websites that are optimized specifically for our users’ needs at approximately 5% of the cost of third-party alternatives. We believe that the AI advantage won’t go to the biggest model; instead it will go to the most specialized one…

…Wix Harmony is now available in all Wix supported languages…

…Both are now accessible within ChatGPT, Microsoft Copilot, and Anthropic’s Claude. Users can type “@Wix” or “@Base44” in these platforms, describe their idea, and a full website or application is created in conversation and managed there too, without any context switching…

…AI has made building online simple and anyone can generate a simple good-looking website in minutes. But that’s as far as it goes. The real complexity begins the moment you hit publish. How does it drive engagement? How do you host it, get found on search engines, run your storefront, secure your customers’ data and actually operate a business day-to-day. These are the hard problems, and we’ve been solving them for 20 years through continuous product innovation and user feedback…

…Still, we also have the flexibility to continue to leverage the best third-party models for the right use cases. So we are never constrained…

…Harmony, which was rolled out in late January across our main geographic markets…

…Where can we expect to have the same thing on Base44. The answer is that I don’t have an exact time line. Obviously, it’s a bigger or more complex undertaking than the Harmony one just because it is much more generalized the — the Harmony use case. That being said, it is something we believe and our top engineers are the ones who are dealing with it…

…In terms of the spend on the Harmony LLM and again, we’re not breaking out the exact number, but it’s quite small, okay? These are not like massive research costs and GPU investments that you can consider when you think about big frontier models. This is something that we managed to do at a very reasonable cost, which also means that for us to continue training it and improving it, should not be something that puts any real weight on our expenses.

Wix websites are now optimised with agentic AI

Wix collaborated with Microsoft to enable users to connect their sites to NLWeb directly from their Wix Dashboards, making Wix sites agentic-optimized. Now available through the Wix SEO & GEO Dashboard, the integration allows structured, continuously updated site data to be queried by AI systems using the ASK protocol, delivering accurate, context-aware answers in real time. 

Wix has ramped up the use of AI in its customer-care organisation for the last 3-plus years, and this has led to a 40% decrease in headcount since 2022, while maintaining or improving service; management is shifting Wix’s R&D (research & development) to be more aligned with Base44’s

We have ramped the integration of AI over the past 3-plus years. This has allowed us to optimize headcount, which has decreased by more than 40% since 2022, while maintaining or even improving in some areas, our services to users…

…We are working to shift our Wix R&D structure to align more closely with that of Base 44, which has been a leader in leveraging AI to drive productivity since day one. We are learning from them and working to implement those same operating principles at Wix. As we execute on this strategy with good line of sight, we expect faster output will more than balance out the cost of AI usage across our organization.

Wix’s Partners are using other AI platforms as well as Wix; the Partners are generally happy with Wix Harmony, but are also pointing out specific areas for improvement; a decent amount of Wix’s Partners are also using Base44

I also think in terms of what they’re using, they are using some AI platforms. By the way, some of them are using Harmony and are very happy with it on one end. And also they’re pointing out to us specific holes, if you may, or missing capabilities that are obviously there because we build Harmony for self-creators and not in the view of partners, but it gives us great visibility into what kind of innovation, what do we need to do next on the partner side in order to make them more successful and happier…

…I’m not going to share percentages, but I can say that we are seeing like there is a decent amount of partners’ usage on Base44. So it’s not marginal.

Wix’s management has no current plans to change the pricing strategy for the core Wix product

I think on Wix at this stage, we think the current structure is the right one. Obviously, if at some point, we introduce something which is very intense on token consumption, then we’ll have to charge for that as well. But at least for now, that’s not the case.

---

Disclaimer: The Good Investors is the personal investing blog of two simple guys who are passionate about educating Singaporeans about stock market investing. By using this Site, you specifically agree that none of the information provided constitutes financial, investment, or other professional advice. It is only intended to provide education. Speak with a professional before making important decisions about your money, your professional life, or even your personal life. I have a vested interest in Alphabet (parent of Google), Amazon (parent of Amazon Web Services), Meta Platforms, Microsoft, MongoDB, Nu Holdings, Okta, Salesforce, Sea, Tencent, Veeva Systems, and Wix. Holdings are subject to change at any time.