Nvidia Transcript: Jensen Huang Unveils Agentic Computing Future, 50% Cash Return, and Vera Rubin CPU Architecture
GTC Taipei 2026 Financial Analyst Q&A, June 2026
Major Capital Return Announcement
Jensen Huang opened the session with a significant financial announcement. Nvidia plans to return 50% or more of free cash flow to shareholders this year, next year, and beyond. This follows the previously announced $80 billion share repurchase program and a 25x increase in the dividend. Huang emphasized that the company plans to increase stock repurchases and dividends over time, describing this as a substantial commitment to shareholders.
The Big Idea: Agentic Computing Pattern
Huang spent considerable time explaining what he called the big idea that he has been discussing for two years. The computing pattern of AI is agentic, and agents are the modern applications. These agents can reason, use tools, and access long-term memory. The memory can be structured or unstructured. Agents can use tools that could be on a PC, in the cloud, design tools, software programming tools, database retrieval, database processing, or chip design tools. This computing pattern will run everywhere, just as applications ran everywhere in the past. It will run in the cloud, in PCs, in workstations, in cars, and even in humanoid robots.
This computing approach is distributed and disagregated, meaning each part of the agent computing pattern runs in different parts of the data center. Hopper was built for pre-training. Grace Blackwell introduced inference in addition to pre-training and post-training. MVLink 72 made it possible to generate the lowest cost tokens in the world, not by 20% but by 20 times. Nvidia is now the lowest cost way to generate tokens. The goal is not to have a low-cost data center but to produce product with low cost. Nvidia's Grace Blackwell generates tokens at the lowest cost of anything in the world.
Vera Rubin Architecture
Vera Rubin was designed for pre-training, post-training, inference, and running agents. This computing pattern is disagregated and distributed. Different parts of that workload run on different parts of Vera Rubin. The heavy part that makes money for AI is token generation. That is why maximizing the number of GPUs is important because what companies sell is tokens. Companies will start by asking how many GPUs can fit in a one gigawatt data center because that maximizes revenues.
The middle is all thinking, which is very heavy compute. Thinking includes reading the context, reading all the documents, reasoning about it, coming up with a plan, acting, which is generating commands for tools. The tool feedback comes back, and the system evaluates whether the answer is right. This back and forth requires tool use to be very fast because the AI is waiting. That is why Nvidia is working with the entire software industry to accelerate their applications. Adobe has accelerated their application, completely rearchitecting Adobe Photoshop and Premiere for the first time in decades. Nvidia is accelerating Cadence, Synopsys, Ansys, and Siemens.
CPUs cannot just be easy to rent or cheap to rent. They have to be fast to respond, which is why single-threaded performance is important. Not multi-threaded, not multi-core, but one CPU doing one job for one AI has to be super fast. Because it is disaggregated, Nvidia thought about how to design the right CPU and where to put it. Vera Rubin is the world's best data processor for memory systems. An AI has to have long-term memory and short-term memory. Memory is data. Moving data around all over the data center requires a lot of bandwidth.
Vera has the most IO bandwidth. Vera also has the most CPU to CPU bandwidth because when doing data processing, the CPUs have to talk to each other. There is no chiplet tax because it is completely on one giant die. Nvidia could make it into four little dice or six little dice, but every time you cross the die, there is a chiplet tax. CPU to CPU bandwidth is three and a half times higher. The cross-sectional bandwidth inside the chip is absolutely the best in the world. The IO bandwidth is absolutely the best in the world, not by 15% but by X factors.
Custom CPU Core Design
Grace was the first CPU Nvidia decided to do this with. Vera was the second generation. The CPU core in Vera is completely custom because Nvidia wanted this CPU to have the world's highest instructions per clock. The CPU fetches 10, decodes 10, and executes 10 instructions at the same time all the way through the pipeline. No CPU in the world does that. Vera was not designed for humans. Vera was designed for agents, which are very impatient. Accelerated computing designed for agents. This entire system was designed not just for pre-training, not just for inference, but to run agents.
PC Reinvention with Microsoft
About three years ago, Huang talked to Satya Nadella and said that in the future, AI wants to run on a device too because people want to have assistance running with them all the time. Right now, if someone wants to talk to their laptop, they have to wait until they get back to their room. In the future, if someone needs their laptop to do something, they just text it with WhatsApp. The laptop becomes an AI, becomes an assistant all day long. People do not want to necessarily run everything in the cloud because if it can run locally, it is free, just like laptops, just like phones.
Huang and Nadella, Microsoft and Nvidia, decided to create a whole new line of computers. This entire line of computers is the first in the world that has tensor processing, parameter compression, and an operating system that allows for a sandbox that is secure because people want to put their agents in a sandbox and give them permission. The entire line includes workstations, desktops, and laptops. In the last three years, they have been working around the clock. Everything is now compatible. Every application that is important has been tested, performance benchmarked.
The entire PC industry, the entire computer industry, not one person left behind, everybody is going to join to reinvent the computer. This is the first reinvention of the PC literally in 40 years. The behavior of a PC is going to change. It will do everything that it used to do better and then it will also be an assistant.
Foundational Models and Ecosystem
Nvidia announced some foundational models, the world's frontier physical AI models. These are the frontier of physical AI models for robotic systems and for autonomous driving. Nvidia made those available to the ecosystem. The same idea is that once you have that model, you put an agentic workflow in there and then just run it everywhere. That is the future. Nvidia is reinventing computing in every single respect.
CPU Market Opportunity
When asked about the $20 billion target for CPU, Huang explained that before Vera, every CPU was built for humans. The characteristics of CPUs of the past and the characteristics of CPUs going forward are going to be very different. Vera is the first one with such extraordinary IPC, such extraordinary bandwidth per core, such extraordinary number of cores with bandwidth between them, and such extraordinary energy efficiency. The CPUs of the future for the agent world are very different than the CPU of the past.
Every data center that has Nvidia GPUs in it will likely just use Vera. Nvidia sells millions of GPUs. Divide that in half and that is the number of CPUs in the head node. Outside of the head node are CPUs to orchestrate the workload and also in the storage server. This storage server CPU is very high performance. In the world of Nvidia GPUs, the CPUs will likely be Nvidia in all three configurations. This effectively doubles the number of CPUs.
Nvidia's CPU share is likely to be higher than GPU share because Nvidia has 100% GPU share and will sell more CPUs outside of Nvidia GPUs. When Nvidia partners with somebody on MVLink Fusion, it sells them switches, NICs, and CPUs. Nvidia Vera will sell beyond Nvidia GPUs. For data processing, which is the number one workload in the world in the cloud, Nvidia is going to sell a lot of CPUs. The company will also sell a lot of CPUs for EDA and simulation because single-threaded performance is so important.
The CPUs of the past and the CPUs of the future have different design centers. Nvidia is going after a zero billion dollar market called agents. Agents is a zero billion dollar market because six months ago it did not exist. Today agents has made it possible to have useful AI and now it is driving enormous demand. Vera was built for that and its time has come.
Reasoning About the Future: CPU to GPU Ratio
Huang emphasized that the ability to predict the future is about reasoning, not about guessing or hoping. He explained that whatever power companies have, they are only making money on tokens. AI companies cannot rent a CPU core. They do not want to rent a CPU core. They want to sell tokens. The business model is tokens. Companies want two things: to increase the ASP of tokens, which means making them as smart as possible with large models, and to make the throughput as high as possible to produce as many tokens as possible.
This factory is only valued for tokens. The first thing Huang would advise customers is to maximize the amount of Vera Rubin MVLink 72s in the data center. Number two, put as many CPUs as necessary, as few as possible to support the GPUs. Companies want to maximize the number of Vera Rubins because it allows them to generate as much money as possible. If spending $50 billion or $60 billion on a data center, companies might as well make a lot of money. CPUs will not make that money back. They do not generate tokens. Why would companies want $30 billion of CPUs that do nothing?
Where are the agents running? Are they all running in the cloud today? Yes. But where are they going to run in the future? Everywhere. They all have CPUs. That is why Nvidia has great CPUs everywhere. Today, there is no choice but to run the agent in the cloud. At Nvidia, the company has agents, and they are all currently running in the cloud, but Nvidia is trying to bring them all back. They should run just on laptops and then they can call the AI models in the cloud or otherwise. So the CPUs would distribute. Nvidia is still going to sell a lot of CPUs.
The reason is very simple. There is 1 billion users of computers today. Tomorrow there are going to be tens of billions of agents using computers. By definition, this entire new population of intelligence needs computers to operate. They are going to need a lot of CPUs, laptops, workstations, and Vera Rubins to think. The CPU market is going to be much bigger, but it cannot reasonably come close to GPU in terms of value.
Enterprise Software Stack
When asked about the enterprise business, Huang explained that the enterprise stack is how Nvidia will help every enterprise software company become agentic companies. The company showed one example with Cadence. The computing pattern has four things: model harness, tools and skills, and runtime. These four things are the ingredients, the operating structure, the operating system of agents. Nvidia has been working with all of the SaaS companies. Everything is open except for the Nvidia AI Enterprise layer. That is the runtime layer for enterprise. Nvidia charges approximately $1,000 to $1,500 per GPU per year.
That software license is obviously growing. When companies run it, when SaaS companies run it in the cloud, that is the software license. Huang believes that is a fairly large business opportunity. It should be billions and billions.
PC Strategy and Value Proposition
Huang explained why Nvidia is entering the PC market. The company has been in the PC industry a long time. Nvidia is not clamoring to build another commodity device. That would be very outside their character. The company does not just build a CPU for CPU sake. Nvidia builds a CPU because the world has changed or the company wants to change the world. A long time ago when Huang came into the graphics industry, the ASP of a graphics card was $49. The high end was $100. Obviously, Nvidia now has $1,500 graphics cards, $2,500, $5,000, $8,000 graphics cards. The company reinvented what graphics means and what graphics does and therefore turned it from a graphics card into a GPU.
Nvidia is going to do the same with PCs. The company did not get in to build a PC. Nvidia is going to reshape what a PC is. A PC today is like a typewriter. It is a device that people type and click into. In the future, it is going to become an assistant that is running all the time. When that PC changes from a smart typewriter to an agentic system, a smart AI assistant running in the background all the time, always available, always at beck and call, oftentimes calling to let people know that something is finished, that way of thinking about an assistant changes the value proposition.
Maybe today people think that should be $1,500, but the idea of having a $10,000 assistant that is used every single day doing things is not illogical. Just as back in the old days, it was sensible to have a $99 phone, but now people are spending $2,000 on smartphones. It is a sensible thing, but the category has to be reinvented. What Satya and Huang are doing, what Microsoft and Nvidia are doing, is reinventing this category, taking along everything that people love about PCs, making it better, but reinventing the whole concept of a personal computer altogether. It is now a personal AI. That is why Nvidia is in this market, not to drive down the PC and compete it into commodity. Nvidia never does that.
Think about 10 years ago. Huang went into the car industry and the embedded controller of a car, the computer in the car was something like $29. He said the company is not trying to compete for those $29. What Nvidia wants to do is reinvent that car into a robotics car, into an autonomous driving car. The first thing the company has to do is make it software programmable and redefine the chassis, the architecture of it. Now Hyperion is everywhere. There is no such thing as a $29 Hyperion. Nvidia reinvented what a car can be. Each one of these categories, the company is just reinventing. That is the Nvidia way.
Optics and Copper Strategy
When asked about optics in the data center, Huang explained that people should use copper as long as they can, as much as they can. They should use optics whenever they must. Copper initially started with very short distances, but because of the SerDes Nvidia invented, the company can now go in the case of MVLink the longest running SerDes in history. Nvidia ran the entire backplane of a rack. Nobody thought that was possible. As a result, Nvidia took copper and turned copper into sexy copper. The company brought sexy back to copper.
Because of the copper, because Nvidia made copper sexy, the company also made the connectors, these micro connectors that they worked with Amphenol on, sexy as well. The company should keep copper and run copper for as long as possible because it is reliable and extremely cost effective. However, Nvidia should use optics wherever it must. One meter or so is kind of the limit. The company thinks it might be able to go a little further, but not 10 times further.
The data centers Nvidia is building are going from in the case of the first OpenAI system 18,000 GPUs. That was Ampere. The first OpenAI supercomputer was 18,000. In the era of Hopper was kind of 100,000. In the era of Blackwell is kind of at the limit about 250,000 or 200,000. But now in the era of Vera Rubin, it is at least half a million. Half a million GPUs is going to need some pretty fancy networking. That is what Spectrum 6 is designed for. Spectrum 6 is the world's first 800 gigabit CPO and it is designed to scale out for AI factories that are hundreds of thousands, a million giant systems. Copper has no chance of doing that.
Nvidia scales up with copper. The company might scale up further with silicon photonics, with optics. Then Nvidia scales out with optics and scales across with optics. The bottom line is the company is going to need a lot of copper, a lot of connectors, a lot of optics, which is why Nvidia partnered with and invested in Coherent, Lumentum, Corning. The partnership with Marvell is about preparing the world to be able to scale up with Nvidia, scale out with Nvidia. Nvidia supply chain is pretty cool.
Inference and Agent Adoption
When asked about inference, Huang explained that inference is going to come into PCs when RTX Spark gets to PCs. The reason for that is simple. What is agent? Agent equals useful AI. This useful AI is not designed to rewrite all of the software on the PC. The agent is going to use the operating system, DirectX, Adobe Photoshop, Autodesk. The agent is finally smart enough, that computing pattern is finally smart enough to use the tools on computers to help assist people in doing their jobs.
Most people, even the experts, know only a fraction of the actual features of Adobe Photoshop, Adobe Premiere, Lightroom. Now with agents, the agent will learn the skill files, the manuals, the operating manuals of all these tools just by reading it. These agents are now going to be experts at every tool. All people have to do is ask the agent how to help do something. They do not have to know the actual commands. The agent will help do it. All of a sudden, all PCs are going to be more useful. PCs are going to be agentic. All of that is inferencing. Every time the agent is thinking, it is inferencing. Inference equals thinking. In order to do, the agent has to think first. In order to do something, it has to come up with a plan. It has to inference. When agents come in, that is when inference takes off because when agents came in, useful AI arrived.
In terms of the data center, Huang would guess, though he does not know exactly because Nvidia systems today are used for training, but later when the company comes up with Vera Rubin, customers take all of the Grace Blackwell systems and use them for inference. That is one of the beauties of Nvidia's system. Researchers are X-factoring their performance for training every single year instead of using a data center that was just built for inference or just built for training and being stuck with it forever. The system is completely fungible. Nvidia built it to be fungible for good reasons.
When something is fungible, its utility goes up. When something is fungible, the utility goes up. When the utility goes up, the TCO comes down. When the utility goes up, the lifetime extends. When the lifetime extends, the TCO comes down. Huang is certain Nvidia's platforms are the lowest TCOs in the world. He can prove it. A100 has been written off for how many years now? Three or four. They are minting coins at $3 to $4 per hour. Here is a chip that is free making $3 an hour, 24 hours a day. That is better than when Huang was a bus boy.
Addressing AI Perception and Adoption
When asked about AI being more unpopular than nuclear power and how to win over the American public to support AI and infrastructure buildout, Huang said this was actually the most optimistic question. In Asia, AI is loved. In the United States, AI is hated. The reason for that is because many people use words that are intended to position their company, intended to cause regulatory capture. They have to be smart about that because it is harmful to the country.
If the United States or South America or Europe do not use AI and they compare it to nuclear bombs, which is completely ridiculous, everyone should have AI, no one should have a nuclear bomb. That comparison is ridiculous. It is nonsensical. It is hyperbolic for no good reason. It scares people. If the industry ends up scaring communities of not using AI, it has done the community and country a huge disservice. That is Huang's biggest concern.
Of course, AI has to be built safely. Of course, there need to be policies in the end markets to keep it from being used improperly. Of course, safety, security, functionality, all of these things have to be true. It is the responsibility of the technology industry, the product makers, the service makers to ensure that is true. The idea that somehow one company is the only company building safety is nonsense. It is like only one car company building safe cars and all the other cars are just randomly killing people. It is the entire industry's job. It is the entire industry's responsibility to make great products, safe products, secure products.
It is possible to be safe and to be concerned and optimistic. These are not conflicting ideas. The world has to build for safety, security, and an optimistic future at the same time. Both of those ideas are possible. It takes a lot of work. The industry has to be very serious about it. But one thing that should not be done is to scare the community into thinking that this technology is dangerous and therefore their children do not engage it. Huang tells both of his children to use AI. He does not want them to be left behind. His advice to his children is the ultimate test of how he feels about the technology.
Nobody is going to advise their children not to use AI. Do not let them be left behind. As analysts write things people read, everybody should have AI. But the industry must take it seriously. Build it properly. Build it safely. Build it securely.
Cost Per Gigawatt Economics
When asked about cost per gigawatt going from $50 billion to $90 billion, Huang asked what is better for a one gigawatt data center: to hold $50 billion worth of computers or $1 trillion worth of computers. One trillion is better. The reason for that is because it is staying one gigawatt. In order to support $1 trillion worth of computers, that $1 trillion worth of computers better be very productive and its energy efficiency must be off the charts. Huang was simply projecting goodness. The alternative is dumb.
This year, $50 billion worth of computers can be put in one gigawatt. If next year only $10 billion can be put in there, when somebody tells you that their dollars per gigawatt, their compute per gigawatt is low, you have to ask yourself is that good news. Huang is simply projecting out good engineering principles. That is what Nvidia is incredibly good at. The company is incredibly good at perf per watt. Perf per watt is the only thing that matters.
It is possible to have three times the perf per watt. Very possible. It is not possible to reduce cost by three times. The reason for that is very simple. Net of gross margins, there are still memories and cables and power generators and MLCCs. You cannot cost reduce everything away. What can be cost reduced is some percentage, but it is not unusual that Nvidia's perf per watt is three times, 10 times. Therefore, perf per watt is utterly vital.
Nvidia's extreme co-design capability, the fact that the company designs across the entire rack and all the software stack allows Nvidia to squeeze literally everything out of it, to smartly architect it away so that energy efficiency is incredible. Perf per watt is world class and Huang continues to believe that is going to be Nvidia's shining capability and the most important characteristic of an AI factory going forward.
Business Segment Disclosure Rationale
When asked about the motivation for resegmenting financials, Huang explained that the purpose of disclosure is to explain how the business works. When everything got lumped into one giant data center, it does not show how the business works at all. It is just one big number. The question is how does the business work.
The business works in three ways inside the hyperscalers alone. One, Nvidia brings customers to the hyperscalers. That is why the company is in every cloud. They are in a lot of ways distributing Nvidia compute because Nvidia brings them customers, big ones. They keep Nvidia in their cloud because the company brings them customers. The fact that the ecosystem is so rich allows Nvidia to bring them a lot of customers.
Number two, Nvidia runs their internal workloads: search, data processing, SQL queries, a ton of stuff, speech, transcription. Nvidia's transcription models are world's best. The company runs a ton of stuff: computer graphics, remote PCs. That is all Nvidia. So there is all of that internal workload, non-AI stuff, classical accelerated computing stuff.
And then there is a third that is AI. It is Anthropic. The company is really pleased it is growing with them. It is OpenAI. It is XAI. The business with CSPs is described in that texture. That alone is plenty for people to think about.
Then there is a second category which is OEMs like HP and Dell and Lenovo and they sell to industrials and OEMs or they sell to what are called NCPs, the NeoClouds or AI native clouds or AI clouds. CoreWeave relies on Nvidia. Lambda relies on Nvidia for many things including helping them stand up their entire data center. They do not want to buy in pieces. They want to buy the reference architecture. The reason for that is because that entire software stack is so complicated and they just do not have enough engineers nor do they want to have that many engineers. They want to move fast. Agility is their skill, their secret sauce.
They can secure land, power and shell. They find it and they are all over the world. They are in Australia, they are in Europe, they are all over the United States and they are super smart about finding land, power and shell because they are regional. Land is regional. Land is not in the cloud. Now they can find land, power and shell. They need Nvidia's computing reference architecture. They need the software stack. They need Nvidia to bring them to customers. And then after that, they need financing. Nvidia makes a small investment in them to secure for them their reputation, to anchor their investment. Then they can go raise 90% of it themselves. Then Nvidia brings customers to them and they tie that whole thing together.
That whole second category, they are not really designers of architectures. They are not architecting new ASICs. They do not want to do that. Their purpose is not to design and build a computer. Their purpose is to operate a computer for a service. They do not want to design a computer. They want to operate a computer. That entire second segment, as it turns out, is 50% of the business. It is growing 100% a year. Long term, it is likely to be even larger. The reason for that is not because the CSPs are going to decline. It is because there is too much stuff at the edge.
Every single factory is going to have a factory brain. That agentic workload is going to run inside every single factory. That computer cannot sit in the cloud. There are many companies that want to because of industrial reasons or telco reasons or just because it has to be regional or sovereign or data protection reasons, they need to build the data centers and control it. That marketplace is going to be quite large.
The third segment is robotics and edge. Do you believe in the future of robotics? Do you believe in physical AI? If so, Nvidia is not going to be in all of them, but the company is going to be in many of them. Nvidia would like to introduce a whole new line of edge systems. That is the three categories: cloud service providers, AI clouds and enterprise and industrial, and then robotics edge.
By disclosing it in that way, it could really help with the granularity of the business and people could decide for themselves how they want to project each one of these segments. This led Huang to the conclusion that Nvidia is gaining share and that is weird, not because the company is taking it from anybody, but because the future of AI is growing in all these different ways. When thinking about the world of AI, Nvidia started out what people think is a very large position. The company is now growing with Anthropic and all of the AI models, OpenAI and so forth. But there is a whole other segment of AI that people do not talk about and it is underserved. It was not until recently that Michael Dell just blew out their quarter and none of them were CSPs. The company completely blew it out.
Productivity Impact Evidence
When asked about recession risk, Huang provided evidence of AI's productivity impact. The single largest body of employees in dollars in the world is software engineers: $3 trillion to $4 trillion. Not including all the marketing people who code or supply chain people who code, just coders. This three or four trillion dollars a year of labor was contributing in 2023 300 million submits. You submit the software when done programming it, tested it, and commit it to production. Then 400 million submits, then 500 million submits in 2025. Last year was 500 million.
The way to think about that is $3 trillion of opex contributed 500 million submits of software. In the first few months of 2026, in the first few months, it went from 500 million per year to 1.4 billion per year. It tripled in size. So what just happened? The productivity of $3 trillion just tripled. The industry produced in excess of $6 trillion of productivity. It is an insane productivity generator.
Meanwhile, people keep talking about how software engineers are going to get laid off. If there is a software engineer who could use agents at $3 trillion of opex and can generate in excess of $6 trillion, companies are not going to lay that person off. They are going to hire more. It makes no sense to lay that person off. Companies are going to hire more so they could generate even more. The world has a lot of code that has to be generated. Code equals problem solving. Code equals GDP growth. Code equals innovation.
There is absolute evidence now that there are a lot of things you could save money on, but the one thing you do not want to save money on is software coding. That is what Nvidia does for a living. That is what the machine does. It generates code, generates tokens.
Supply Chain Support
Huang addressed supply chain questions directly. Nvidia has the support of the entire ecosystem and supply chain to provide for very robust growth. The company grew nearly 100% year over year from a base that was already very large. Nvidia has the ability with the support of the supply chain to grow very robustly. However, the company does not have enough supply and the reason for that is because the world supply chain is supply constrained. But Nvidia has the support of the ecosystem to have very robust growth well in support of whatever guidance has been provided.