Jensen Huang：Nvidia 的未来、Physical AI、Agent 崛起、推理爆发与 AI 公关危机

This is a special episode this week. We've preempted the weekly show, and there are only 3 people we preempt the show for: President Trump, Jesus, and Jensen. I'll let you pick which order we do that. What an amazing run you've had, and what a great event.

Every industry is here. Every tech company is here. Every AI company is here. Incredible.

Incredible. Extraordinary. One of the great announcements of the past year has been Groq. When you made the purchase of Groq, did you realize how insufferable Chamath would become?

I had an inkling that—

We're his friends. We have to deal with him every week.

I know it.

We had to deal with him for the 6-week close.

I know it.

2 weeks.

It's all coming back to me now. It's making me rather uncomfortable.

The thing is, many of our strategies are presented in broad daylight at GTC years in advance of when we do them. 2½ years ago, I introduced the operating system of the AI factory, and it's called Dynamo. Dynamo, as you know, is a piece of instrumentation—a machine created by Siemens to turn essentially water into electricity. Dynamo powered the factories of the last industrial revolution. So I thought it was the perfect name for the operating system of the next industrial revolution, the factory of that.

Inside Dynamo, the fundamental technology is disaggregated inference. Jason, I know you're super technical.

Absolutely. I'll let you take this one. Go ahead and define it for the audience. I don't want to step on you.

Yeah, thank you. I knew you wanted to jump in there for a second.

It's disaggregated inference, which means the pipeline—the processing pipeline of inference—is extremely complicated. In fact, it is the most complicated computing problem today. It's an incredible scale, with lots of mathematics of different shapes and sizes.

We came up with the idea that you would disaggregate parts of the processing, such that some of it can run on some GPUs and the rest of it can run on different GPUs. That led us to realize that maybe even disaggregated computing could make sense—that we could have different, heterogeneous types of computing. That same sensibility led us to Mellanox.

Yep.

Today, NVIDIA's computing is spread across GPUs, CPUs, switches, scale-up switches, scale-out switches, networking processors, and now we're going to add Groq to that. We're going to put the right workload on the right chips. We really evolved from a GPU company to an AI factory company.

I think that was probably the biggest takeaway that I had. You're seeing this fundamental disaggregation, where we've gone from a GPU to a composition of all these different options that will eventually exist. The thing that you said on stage was, "I would like the high-value inference people to take a listen to this," and you said 25% of your data center space should be allocated to this Groq LPU-GPU combination.

About 25% of the Vera Rubin systems in the data center.

Can you tell us how the industry looks at this idea of creating this next-generation form of disaggregated prefill and decode, and how do you think people will react to it?

Yeah. Take a step back. At the time that we added this, we went from large language model processing to agentic processing. When you're running an agent, you're accessing working memory, you're accessing long-term memory, and you're using tools. You're really beating up on storage very hard.

You have agents working with other agents. Some of the agents are very large models, some of them are smaller models, some of them are diffusion models, and some of them are autoregressive models. So there are all kinds of different types of models inside this data center. We created Vera Rubin to be able to run this extraordinarily diverse workload.

We added what used to be a 1-rack company; we now add 4 more racks. NVIDIA's TAM, if you will, increased from whatever it was to probably something like 33% to 50% higher. A lot of that is going to be storage processors—it's called BlueField. A lot of it, I'm hoping, will be Groq processors, and some of it will be CPUs. A lot of it is also going to be networking processors.

All of this is going to be running, basically, the computer of the AI revolution called agents—the operating system of modern industry.

What about embedded applications? My daughter's teddy bear at home wants to talk to her. What goes in there? Is it a custom ASIC, or does there end up being a much broader set of TAM, with developing tools that are maybe different for different use cases at the edge and in embedded applications?

I think that there are 3 computers in the problem. At the largest scale, when you take a step back, there's one computer that's really about training the AI model—developing and creating the AI. There's another computer for evaluating it.

Depending on the type of problem you're having—for example, you look around and there's all kinds of robots and cars and things like that—you have to evaluate these robots inside a virtual gym that represents the physical world. It has to be software that obeys the laws of physics. That's a second computer. We call that Omniverse.

The third computer is the computer at the edge, the robotics computer. That robotics computer could be a self-driving car, a robot, or a teddy bear—a little tiny one for a teddy bear. One of the most important ones is one that we're working on that basically turns telecommunications base stations into part of the AI infrastructure.

It's a $2 trillion industry. All of that, in time, will be transformed into an extension of the AI infrastructure. Radios will become edge devices—factories, warehouses, you name it. So there are these 3 basic computers, and all of them are going to be necessary.

Jensen, last year, I think you were ahead of the rest of the world in saying inference isn't going to a thousand—

Just last year.

Yes. Is it going to—

feelings.

Is it going to 1 million X? Is it going to 1 billion X?

Yeah.

I think people at the time thought it was pretty hyperbolic because the world was still focused on pre-scaling, on training. Here we are. Now inference has exploded. We're inference-constrained.

You announced an inference factory that I think is leading-edge, that's going to be 10X better in terms of throughput than the next factory. Yet, if I listen to the chatter out there, it's that your inference factory is going to cost $40 billion or $50 billion, and the alternatives—custom ASICs, AMD, and others—are going to cost $25 billion to $30 billion, and you're going to lose share.

Why don't you talk to us? What have you seen? How do you think about share? Does it make sense for all these folks to pay something that's a 2X premium to what others are marketing?

The big takeaway, the big idea, is that you should not equate the price of the factory with the price of the tokens, the cost of the tokens. It is very likely that the $50 billion factory—and in fact, I can prove it—will generate for you the lowest-cost tokens. The reason for that is because we produce these tokens at extraordinary efficiency—10 times.

You know, the difference between $50 billion—now, it turns out $20 billion is just land, power, and shell, right?

Right.

Then, on top of that, you have storage anyway, networking anyway, CPUs anyway, servers anyway, and cooling anyway. The difference between that GPU being 1X the price or ½X the price is not between $50 billion and $30 billion. Pick your favorite number, but let's say between $50 billion and $40 billion.

Yeah.

That is not a large percentage when the $50 billion data center is actually 10 times the throughput.

Right. Jensen, I want—

That's the reason why I said that even for most chips, if you can't keep up with the state of the technology and the pace that we're running, even when the chips are free, it's not cheap enough.

Yeah. Can I just ask a general strategy question?

Yeah.

You're running the most valuable company in the world. This thing is going to do $350 billion-plus of revenue next year and $200 billion of free cash flow. It's compounding at these crazy rates. How do you decide what to do? How do you actually get the information?

It's famous now, these sorts of emails that people are meant to send you, but how do you really get an intuition for how to shape the market, where to really double down, where to maybe pull back, and where to actually go into a greenfield? How does that information get to you? How do you decide these things?

In the final analysis, that's the job of the CEO. Our job is to define the vision and the strategy. We're informed, of course, by amazing computer scientists, amazing technologists, and great people all over the company, but we have to shape that future.

Part of it has to do with this: Is this something that's insanely hard to do? If it's not hard to do, we should back away from it. The reason for that is, if it's easy to do, obviously—

Lots of competitors.

Of competitors. Is this something that has never been done before, that's insanely hard to do, and that somehow taps into the special superpowers of our company? I have to find this confluence of things that meets the standard.

And in the end, we also know that a lot of pain and suffering is going to go into it.

Yeah.

There are no great things that are invented because they were just easy to do and, on the first try, here we are. If it's super hard to do, nobody's ever done it before, and it's very likely that you're going to have a lot of pain and suffering, you better enjoy it.

Can you look at maybe 3 or 4 of the more long-tail things you announced and talk about the long-term viability of whether it's the data center in space, what you're trying to do with ADAS in autos, or what you're trying to do on the biology side? Give us a sense of how you see some of these curves inflecting upward and some of these longer-tail businesses.

Excellent. Physical AI is a large category. We believe—and as I just mentioned—we have 3 computing systems, all the software platforms on top of them. Physical AI, as a large category, is the technology industry's first opportunity to address a $50 industry that has largely been void of technology until now.

And so, we need to invent all of the technology necessary to do that. I felt that that was a 10-year journey. We started 10 years ago. We're seeing it inflect now. It is a multibillion-dollar business for us. It's close to $10 billion a year now. And so, it's a big business, and it's growing exponentially. That's number 1.

In the case of digital biology, I think we are literally near the ChatGPT moment of digital biology. We're about to understand how to represent genes, proteins, and cells. We already know how to understand chemicals. And so, the ability for us to represent and understand the dynamics of the building blocks of biology—that's 2, 3, 5 years from now.

In 5 years' time, I completely believe that digital biology is going to inflect in the healthcare industry. These are a couple of the really great ones, and you can see they're all around us.

Agriculture.

Inflecting now.

No question. Yeah.

Jensen, I want to take you from the data center to the desktop. The company was built in large part on hobbyists, video gamers, and all those graphics cards in the beginning. You mentioned in front of, I think, 10,000 people here, just Claude, uh open Claude, Claude Code, and what a revolution agents have become.

Specifically, the hobbyists, who are really where a lot of energy and innovation come from, want desktops. You announced one here. I believe it's the Dell 6800. This is a very powerful workstation to run local models, with 750 GB of RAM. Obviously, the Mac Studio is sold out everywhere.

In my company, we're moving to open Claude everything. Freeburg just got Claude peeled. You got Claude peeled, I understand, and you're obsessed with these. What does this from-the-streets movement of creating open-source agents and using open source on the desktop mean to you?

So great.

Where is that going?

Yeah, so great. First of all, let's take a step back. In the last 2 years, we saw basically 3 inflection points. The 1st one was generative AI. ChatGPT brought AI to the common person, to everybody's awareness. But the fact of the matter is, the technology sat in plain sight months before ChatGPT. It wasn't until ChatGPT put a user interface around it and made it easy for us to use that generative AI took off.

Now, generative AI, as you know, generates tokens for internal consumption as well as external consumption. Internal consumption is thinking, which led to reasoning. o1 continued that wave of ChatGPT grounded information and made AI not only answer questions, but answer questions in a more grounded, useful way. We started seeing the revenues and the economic model of OpenAI start to inflect.

The 3rd one was something we saw only inside the industry: Claude Code, the 1st agentic system that was very useful. Really revolutionary stuff. But Claude Code was only available for enterprises. Most people outside never saw anything about Claude Code until OpenClaw. OpenClaw basically put into popular consciousness what an AI agent can do. That's the reason why OpenClaw is so important from a cultural perspective.

Yeah.

Now, the 2nd reason why it's so important is that OpenClaw is open, but it formulates, it structures a type of computing model that is basically reinventing the computer altogether. It has a memory system. It has a scratch file, a short-term memory file system. It has scales. Did you say skills or scales?

Skills.

Oh, skills.

Have scales, theoretically, yeah.

Yeah, scales. So, 1st of all, it has resources and manages resources. It does scheduling.

Yep.

Right? And cron jobs. It can spawn off agents, decompose a task, and solve problems. It has I/O subsystems. It can input, it has output, and it connects to WhatsApp. It also has an API that allows it to run multiple types of applications called skills.

Yeah.

These 4 elements fundamentally define a computer.

Yeah.

And therefore, what do we have? We have a personal artificial intelligence computer for the very 1st time.

Open source.

It's open source. It runs literally everywhere. And so, this is now the blueprint, the operating system of modern computing.

Yeah.

And it's going to run literally everywhere. Now, of course, one of the things that we have to help it do is—whenever you have agentic software, you have to make sure that agentic software has access to sensitive information, can execute code, and can communicate externally.

We have to make sure that all of it is governed, all of it is secure, and that we have policies that give these agents 2 of the 3 things, but not all 3 things at the same time.

Right.

And so, on the governance part of it, we contributed to Peter. Peter Steinberger was here, and so we've got a mountain of great engineers working with him to help secure and maintain that thing so that it can protect our privacy and protect our security.

Jensen, that paradigm shift makes some of the AI legislation that has passed around the country to regulate AI, and a lot of the proposed legislation, effectively moot, doesn't it? Can you comment for a second on how quickly the paradigm shift obviates a lot of the models for regulatory oversight of AI, which is becoming a very hot topic in politics right now?

Well, this is the part where, with policymakers, we need to always get in front of them, and Brad, you do a great job doing this. We have to get in front of them and inform them about the state of the technology—what it is and what it is not. It is not a biological being. It is not alien. It is not conscious. It is computer software.

Yeah, exactly.

And it is not something where we say, “We don't understand it at all.” It is not true that we don't understand it at all. We understand a lot of things about this technology. And so, I think, 1st, we have to make sure that we continue to inform policymakers and not allow doomerism and extremism to affect how policymakers think about and understand this technology.

However, we still have to recognize that this technology is moving really fast, and we don't want to get policy ahead of the technology too quickly. The risk that we run as a nation—our greatest source of national security concern with respect to AI—is that other countries adopt this technology while we are so angry at it, afraid of it, or somehow paranoid about it, so that our industries and our society don't take advantage of AI. And so, I'm just mostly worried about the diffusion of AI here in the United States.

Can you double-click on, if you were in the boardroom of Anthropic over that whole scuttlebutt with the Department of War? It sort of builds on this idea that people didn't know what to think. It's added to this layer of either resentment or fear or just general mistrust that people sometimes have at the software level of AI. What do you think you would have told Dario and that team to do differently to try to change some of this outcome and some of this perception?

The 1st thing that I would say about Anthropic is, first of all, the technology is incredible. We are a large consumer of Anthropic technology. I really admire their focus on security. I really admire their focus on safety. The culture by which they went about it, the technological excellence by which they went about it—really fantastic.

I would say that the desire to warn people about the capability of the technology is also really terrific. We just have to make sure that we understand that the world has a spectrum, and that warning is good; scaring is less good.

Right.

And because this technology is too important to us, I think that it is fine to predict the future, but we need to be a little bit more circumspect. We need to have a little bit more humility that, in fact, we can't completely predict the future. The ability to say things that are quite extreme, quite catastrophic, when there's no evidence of them happening, could be more damaging than people think.

And, of course, we are technology leaders. There was a time when nobody listened to us.

Yeah.

But now, because technology is so important in the social fabric, such an important industry, and so important to national security, our words do matter. And I think we have to be much more circumspect, more moderate, more balanced, and more thoughtful.

Well, I would nominate you. I think the industry has to get together. AI has 17% popularity in the United States. I mean, we see what happened to nuclear, right? We basically shut down the entire nuclear industry, and now we have 100 fission reactors being built in China and 0 in the United States.

We hear about moratoriums on data centers, so I think we have to be a lot more proactive about that. But I want to go back to this agentic explosion that you're seeing inside your company—the efficiencies and the productivity gains inside your company. There's a lot of debate about whether or not we're seeing ROI, right?

You and I, entering into this year, the big question was: Are the revenues going to show up? Are the revenues going to scale like intelligence? And then we had this kind of Oppenheimer moment of a $5–6 billion month by Anthropic in February. Do you think, as you look ahead—you announced $1 trillion of visibility into just Blackwell and Vera Rubin over the course of the next couple of years—when you see this happening at Anthropic and OpenAI, do you think we're on that curve now, where we're going to see revenues scale in the way that intelligence is scaling?

When you look around this audience, you will see that Anthropic and OpenAI are represented here, but in fact, 99% of everything that is here is all AI, and it's not Anthropic and OpenAI.

Right. Right.

The reason for that is because AI is very diverse. I would say that the second-most-popular model category is open models.

Number one is OpenAI, right? Open source—open weights, open source—is number two. Very distant third is Anthropic, and that tells you something about the scale of all the AI companies that are here.

And so it's important to recognize that. Let me come back and say a couple of things. One, when we went from generative to reasoning, the amount of computation we needed was about 100 times.

Right.

When we went from reasoning to agentic, the computation was probably another 100 times. Now we're looking at computation going up by a factor of 10,000× in just 2 years. Meanwhile, people pay for information, but people mostly pay for work.

Talking to a chatbot and getting an answer is super great.

Right.

Helping me do some research—unbelievable. But getting work done, I'll pay for.

Indeed.

And so that's where we are. Agentic systems get work done. They're helping our software engineers get work done. You take that, and you've got 10,000× more compute. You get probably, at this point, 100× more consumption now.

Yeah.

And we haven't even started scaling yet. We are absolutely at a million times.

Which is, I think, a great place to talk about the number of people you have. You have 20,000 or 30,000 at the company, something like that?

We have 43,000 employees. I would say 38,000 are engineers.

The conversation we've had on the pod a number of times is, "Oh my God, look at the token usage in our companies." It's growing massively, and some people are asking, "Hey, when I join a company, how many tokens do I get? Because I want to be an effective employee."

You postulated, I believe, during your 2.5-hour keynote—pretty long keynote, well done—that you were spending—

Well done; it would be shorter.

Yeah. You didn't have time to do a—

So you guys know, there is no practice, and so it's a grip and a rip. I just wanted to let you know I was writing the speech while I was giving the speech.

Yeah. Yeah. But does that mean, if we do back-of-the-envelope math, 75,000 tokens for each engineer, something like that? So are you spending $1 billion or $2 billion on tokens for your engineering team right now?

We're trying to. Let me give you a thought experiment. Let's say you have a software engineer or AI researcher and you pay them $500,000 a year. We do that all the time. This is happening all the time.

That $500,000 engineer, at the end of the year, I'm going to ask him, "How many tokens—how much did you spend in tokens?" If that person said $5,000, I will go ape. If that $500,000 engineer did not consume at least $250,000 worth of tokens, I am going to be deeply alarmed.

This is no different from one of our chip designers who says, "Guess what? I'm just going to use paper and pencil. I don't think I'm going to need any CAD tools."

This is a real paradigm shift in thinking about these all-star employees. It almost reminds me of what we learned in the NBA when LeBron James started spending $1 million a year just on the health of his body, maintaining it. Here he is, still playing.

It really is: If these are incredible knowledge workers, why wouldn't we give them superhuman abilities?

Where does that go if we extrapolate out 2 or 3 years from now? What is the efficiency of that all-star at NVIDIA and what they're able to accomplish?

Well, first of all, the thought "Wow, this is too hard" is gone. "This is going to take a long time"—that thought is gone. "We're going to need a lot of people"—that thought is gone.

This is no different from the last Industrial Revolution. Somebody goes, "Boy, that building really looks heavy." Nobody says that. Nobody says, "Wow, that mountain looks too big." Everything that's too big, too heavy, or takes too long—those thoughts, those ideas, are all gone.

What happens to creativity?

That's right.

What can you come up with?

Exactly.

Which means now the question is: How do you work with these agents?

Well, it's just a new way of doing computer programming. In the past, we coded; in the future, we're going to write ideas, architectures, and specifications. We're going to organize teams. We're going to help them define how to evaluate good versus bad. What does it look like when something is a great outcome? How do you iterate with them? How do you brainstorm?

That's really what you're looking for, and I think that every engineer is going to have 100 agents.

Back to the PR problem the industry has right now: You have executives like David Freiberg with Ohalo, who's looking at literally using technology—your technology and AI—to increase the number of calories produced and make high-quality calories. By what factor do you think you can bring the cost down, and what impact does this vision have for what you're doing?

Zero-shot genomic modeling, and it works.

Yeah.

Then you have that moment and you're like, "Holy shit." Honestly, that's after people are replacing entire enterprise software stacks in a night. I did something in 90 minutes—I was telling the guys about it. I replaced the whole software stack and a whole bunch of workload. In 90 minutes on Claude, I ran this agentic system, built the whole thing, and deployed it.

On a Sunday—

Night.

Sunday night. 10:00 p.m. I was done at 11:30, I went to bed.

As the CEO, you replaced the whole stack?

Yeah. Everyone on my management team had to do a similar exercise. Over the weekend, what we saw on Monday, I was like, "It's over."

But the technical stuff, the science stuff—we did something in 30 minutes using autoresearch, and I'd love your view on autoresearch and what that tells us about how far we still have to go in terms of efficiency.

Using autoresearch and a chunk of data, something was published internally that we said, "Oh my God." That would normally be a PhD thesis that would take 7 years. It would be one of the most celebrated PhD theses we've ever seen in this field, and it would be in the journal Science. It was done in 30 minutes on a desktop computer running autoresearch with all the data we had just ingested.

We got it on Friday, and we're like, "Hey, let's try it." We tried it, booted it up, went to GitHub, downloaded autoresearch, and ran it. You see everyone's face just go like—

The potential of what this is unlocking for us is the kind of thing that would take 7 years, and it happened in 30 minutes. We're experiencing it in genomics, and we're like, "This is unbelievable."

I think the acceleration is widening the aperture for everyone in a way that you didn't imagine a few years ago. But just going back to the autoresearch point, can you comment on what you think about the fact that this thing got published with 600 lines of code in a weekend, and the capacity that it has to run locally and achieve what it can achieve with all of these diverse data sets? What does that tell us about the early stages we're in, in terms of optimization on algorithms and hardware?

The fundamental reason why OpenClaw is so incredible, number one, is its confluence—its timing with the breakthroughs in large language models. Its timing was perfect. It was impeccable.

Now, in a lot of ways, Peter probably wouldn't have come up with it if not for the fact that Claude, GPT, and ChatGPT have reached a level that is really very good.

It is also a new capability that allows these models to use tools. The tools that we've created over time—web browsers, Excel spreadsheets, and, in the case of chip design, Synopsys, Cadence, Omniverse, Blender, Autodesk, and all of these tools—are going to continue to be used.

Some people say that the enterprise IT software industry is going to get destroyed. Let me give you the alternative view. The enterprise software industry is limited by butts in seats. It's about to get 100 times more agents banging on those tools. There are going to be agents banging on SQL, vector databases, Blender, and Photoshop.

The reason for that is that those tools, first of all, do a very good job. Second, those tools are the conduit between us. In the final analysis, when the work is done, it has to be represented back to me in a way that I can control.

Right.

I know how to control those tools. So, I need everything to be put back into Synopsys. I want everything put back into Cadence, because that's how I control it. That's how I ground-truth it.

So, we have these closed-source models; they're excellent. We have these open-weight models. Many of the Chinese models are incredible—absolutely incredible. Two days ago, you may not have seen this because you were busy onstage, but there was a training run that happened in this crypto project called Bittensor Subnet 3.

They managed to train a 4-billion-parameter Llama model totally distributed, with a bunch of people contributing excess compute, but they were able to do it statefully and manage a training run, which I thought was a pretty crazy technical accomplishment.

Yeah.

Because it's like random people, and each person gets a little share.

Our modern version of Folding@home.

Exactly. So, what do you think about the end state of open source? Do you see this decentralization of architecture as well, and decentralization of compute to support open weights and a totally open-source approach to making sure AI is broadly available to everyone?

I believe we fundamentally need models as a first-class product, a proprietary product, as well as models as open source. These 2 things are not A or B; they are A and B. There's no question about it. The reason for that is that a model is a technology, not a product. A model is a technology, not a service. For the vast majority of consumers, the horizontal layer, the general intelligence, I would really, really love not to fine-tune my own.

Right.

I would really love to keep using ChatGPT. I love using Claude. I love using Gemini. I love using X. They all have their own personalities, as you know, which just kind of depends on my mood and depends on what problem I'm trying to solve. I might do it on X, or I might do it on ChatGPT. That segment of the industry is thriving. It's going to be great.

However, all these industries—their domain expertise and their specialization—have to be channeled and captured in a way that they can control. That can only come from open models. The open model industry, we're contributing tremendously to. It is near the frontier. Quite frankly, even if it reaches the frontier, I think that world-class models-as-a-product are going to continue to thrive.

Every startup we're investing in now is open source first and then going to the proprietary model.

Yeah, and the beautiful thing is, because you have a great router, you connect the 2. On day 1, every single day, you're going to have access to the world's best model. And then it gives you time to cost-reduce, fine-tune, and specialize. So, you're going to have world-class capabilities out of the chute every single time.

Jensen, can I ask a question? Nobody wants the US to win the global AI race more than you, right? But a year ago, the Biden-era diffusion rule really was an anti-American diffusion of AI around the world. So, here we are, a year into the new administration. Give us a grade. Where are we in terms of global diffusion and the rate at which we're spreading US AI technology around the world? Are we an A? Are we a B? Are we a C? What's working? What's not working?

Well, first of all, President Trump wants American industry to lead. He wants the American technology industry to lead. He wants the American technology industry to win. He wants us to spread American technology around the world. He wants the United States to be the wealthiest country in the world. He wants all of that.

At the current moment, as we speak, NVIDIA gave up a 95% market share in the 2nd-largest market in the world, and we're at 0%. President Trump wants us to get back in there. The first thing is to get licenses for the companies that we're going to be able to sell to. We've got many companies that have requested licenses. We've applied for licenses for them, and we've got approved licenses from Secretary Lutnick. Now we've informed the Chinese companies, and many of them have given us purchase orders. So, we're in the process of cranking up our supply chain again to go ship.

I think, at the highest level, one of the things that we should acknowledge is this: Our national security is diminished when we don't have access to miniature motors and rare-earth minerals. It's diminished when we don't control our telecommunications networks. It's diminished when we can't provide sustainable energy for our country. It is fundamentally diminished. Every single one of these industries is an example of what I don't want the AI industry to be.

Right.

When we look forward in time and we say, "What do we want? What does it look like when the American technology industry, the American AI industry, leads the world?" we can all acknowledge that there is no way that the AI model race is won universally. We can all acknowledge that that is an outcome that makes no sense.

However, we can all imagine that the American tech stack, from chips to computing systems to the platforms, is used broadly by the world, where they build their own AI, they use public AI, they use private AI, whatever, and they can build their applications in their society. I would love it if the American tech stack were 90% of the world. Yes, I would love that. The alternative, if it looks like solar, rare earth, magnets, motors, and telecommunications, I consider that a very bad outcome for national security.

Agreed.

Yeah.

How much are you monitoring the situation with the conflicts around the world right now, and how much does it worry you, Jensen? So, China and Taiwan, and then helium availability coming out of the Middle East, I understand, can be a supply-chain risk to semiconductor manufacturing. How much do these situations worry you? How much are you spending on them?

Well, first of all, in the Middle East, we have 6,000 families there. We have a lot of Iranians at NVIDIA, and their families are still in Iran. And so, we have a lot of families there. The first thing is, they're quite anxious, they're quite concerned, and quite scared. We're monitoring them and keeping an eye on them all the time. They have 100% of our support.

I've been asked several times whether we're still considering being in Israel. We are 100% in Israel. We are 100% behind the families there. We are 100% in the Middle East. I was also asked, given what's happening in the Middle East, whether that's an area where we believe we can expand artificial intelligence to. I believe that there's a reason we went to war, and I believe that at the end of the war, the Middle East will be more stable than before. So, if we were considering it before, we should absolutely be considering it after. And so, I'm 100% in on that.

With respect to Taiwan, we have to do 3 things. 1, we have to make sure that we reindustrialize the United States as fast as we can.

Yeah.

Whether it's the chip-manufacturing plants, the computer-manufacturing plants, or the AI factories—

How are we doing on that?

We're doing excellent. By gaining the strategic support and the friendship of Taiwan's supply chain—by gaining their friendship and their support—we were able to build in Arizona, Texas, and California at incredible rates. They are genuinely a strategic partner. They deserve our support, they deserve our friendship, and they deserve our generosity. And they're doing everything they can to accelerate the manufacturing process for us. So, I think that's number 1.

Number 2, we ought to diversify the manufacturing supply chain. Whether it's South Korea, Japan, or Europe, we ought to diversify the supply chain and make it more resilient. And number 3, let's demonstrate restraint. While we're increasing our diversity and resilience, let's not push—

Unnecessarily.

So, we need to be patient.

It's thoughtful. Is helium a problem? A lot of reports out there.

I think helium could be a problem, but it's also the case that the supply chain probably has a lot of buffer in it. These kinds of things tend to have a lot of buffer.

Yeah. You've made massive progress in self-driving. You made a big announcement. You've added many more partners, including BYD. There was just a video of you driving around in a Mercedes. And a huge announcement with Uber that you're going to have a number of cars on the road from many different manufacturers. Your bet, I believe, is that there's going to be an Android-type open-source platform that you're going to play a major part in, with dozens of car providers.

And then maybe on the other side, there could be an iOS with Tesla or Waymo. What's your strategy thinking there, and how does that chessboard emerge? It feels like you have a pretty deep stack, and in some ways you're competing, while in other places you're collaborative.

Yeah. Taking a step back, we believe that everything that moves will be autonomous, completely or partly, someday. Number 1. Number 2, we don't want to build self-driving cars, but we want to enable every car company in the world to build self-driving cars.

And so we've built all 3 computers: the training computer, the simulation computer, the evaluation computer, as well as the car computer. We developed the world's safest driving operating system. We also created the world's first reasoning autonomous vehicle so that it could decompose complicated scenarios into simpler scenarios that it knows how to navigate through, just like us—reasoning systems.

And so that reasoning system, called Alpaca IO, has enabled us to achieve incredible results. We open-source this. We vertically optimize, we horizontally innovate, and we let everybody decide: Do you want to buy 1 computer from us? In the case of Tesla, they buy our training computers. Do they want to buy our training computer and our simulation computers? Or do they want to work with us to do all 4 and even put the car computer in their car?

Our attitude is, we want to solve the problem. We're not the solution provider, and we're delighted however you work with us.

To build on this question, because I think it's so fascinating: You actually do create this platform. A thousand flowers are blooming. But it's also true that some of those flowers now want to go back down in the stack and try to compete with you a little bit. Google has TPU; Amazon has Inferentia and Trainium. Everybody's sort of spinning up their own version of, "I think I can out-NVIDIA NVIDIA," even though they also tend to be huge customers. How do you navigate that? What do you think happens over time, and where do those things play in the complexion of this kind of ecosystem?

Yeah, really great. First of all, we're the only AI company. We're an AI company. We build foundation models. We're at the frontier of many different domains. We build every single layer, every single stack. We're the only AI company in the world that works with every AI company in the world. They never show me what they're building, and I always show them exactly what I'm building.

Right.

Yeah. And so the confidence comes from this. Number 1, we are delighted to compete on what is the best technology. To the extent that we can continue to run fast, I believe that buying from NVIDIA is still one of the most economic things they could do. And I see this incredible confidence there.

Number 2, we're the only architecture that can be in every cloud, and that gives us some fundamental advantages. We're the only architecture you can take from a cloud and put into on-prem, in the car, or in any region.

That's right—in space.

And so there's a whole part of our market—about 40% of our business. Most people don't realize this: 40% of our business. Unless you have the CUDA stack, unless you can build an entire AI factory, the customers don't know what to do with you. They're not trying to build chips; they're not trying to buy chips. They're trying to build AI infrastructure.

And so they want you to come in with a full stack, and we've got the whole stack. Surprisingly, NVIDIA's gaining market share. If you look at where we are today, we're gaining share.

What happens is these guys try it, and they realize, "Oh, my God, it's too much." And then they come back. Is that why their share grows?

Well, we're gaining share for several reasons. Number 1, our velocity has gone up. We help people realize it's not about building the chip; it's about building the system. And that system's really hard to build, so their business with us is increasing.

In the case of AWS, I think they just announced—I think it was yesterday—that they're going to buy 1 million chips in the next couple of years. I mean, that's a lot of chips from AWS, and that's on top of all the chips they've already bought. We're delighted to do that.

Number 1, we're gaining share this last couple of years because we now have Anthropic coming to NVIDIA. Meta SL is coming to NVIDIA. The growth of open models is incredible, and that's all on NVIDIA. We're growing in share because of the number of models.

We're also growing in share because all of these companies are outside the cloud, and they're growing regionally, in enterprise, in industries, and at the edge. That entire segment of growth is really hard to do if it's just building an ASIC.

Related to that, and not to get in the weeds on the numbers, but analysts don't seem to believe you. If you look at the consensus forecast, you said compute could go 1,000,000×, and yet they have you growing next year at 30%, the year after that at 20%, and in 2029, which is supposed to be a monster year, at 7%.

If you take your TAM and apply their growth numbers, it suggests that your share will plummet. Do you see anything in your future order book that would make that correct?

Yeah. First of all, they just don't understand the scale and the breadth of AI.

Yes. Yeah, I think that's true. Most people think that AI is in the top 5 hyperscalers. Right. That's right.

There's also an orthodoxy around these law-of-large-numbers arguments where they have to go back to their investment banking risk committee and show some model. They're not going to believe, in their minds, that $5 trillion goes to $15 trillion. They're like, "It can go to—it can go to—it can go to $7 trillion."

Or they need to have a $10 trillion company.

It's all just CYA stuff.

It has happened before, so you can't say it will.

And because you have to redefine what it is that you do. There was somebody who made an observation recently: "NVIDIA, Jensen, how can you be larger than Intel in servers?" The reason for that is because the CPU market of the entire data center was about $25 billion a year.

Right.

We do $25 billion a year, as you guys know. And so, obviously, obviously, that was a joke.

No, it's—but it's—

No, that was not guidance. But anyhow, the point is, how big you can be depends on what it is that you make.

Right.

NVIDIA is not making chips. Number 1, making chips does not help you solve the AI infrastructure problem anymore. It's too complicated. Number 3, most people think that AI is narrowly in the things that they talk about, hear, and see.

Right.

AI is much bigger. OpenAI is incredible. They're going to be enormous. Anthropic is incredible. They're going to be enormous. But AI is going to be much, much bigger than that.

Yeah.

And we address that segment.

Tell us about data centers in space for a second.

Yep. We're already in space.

How should the layman think about what that business is versus when you hear about these big data center build-outs happening on the ground?

Well, we should definitely work on the ground first because we're already here. Number 1. Number 2, we should prepare to be out in space, and obviously there's a lot of energy in space.

The challenge, of course, is cooling. You can't take advantage of conduction and convection, so you can only use radiation. Radiation requires very large surfaces. That's not an impossible thing to solve, and there's a lot of space in space. Nonetheless, the expense is still quite high.

We're going to go explore it. We're already there. We're already radiation-hardened. We have Kuda in satellites around the world. They're doing imaging, image processing, and AI imaging. That kind of stuff ought to be done in space instead of sending all the data back here and doing imaging down here. We ought to just do imaging out in space.

There are a lot of things that we ought to do in space. In the meantime, we're going to explore what the architecture of data centers looks like in space. It'll take years. It's okay. I've got plenty of time.

I wanted to double-click on health care. I know you've got a big effort there. We're all of a certain age where we're thinking about lifespan and health span. We all look great, I think—some better than others. I think some better than others. I don't know what your secret is, Jensen.

What are you taking? What's off the menu? You've got to talk to me when we're backstage. I want to know in the green room what you have going on.

Squats, push-ups, and sit-ups.

Perfect. Okay.

It works.

In terms of the build-out in health care, where is that going? What kind of progress are we making? I was just using Claude to do some analysis and asking, "Where are all these billing codes?" We spend twice as much money in the U.S., and we seem to get half as much. It seemed like 15% to 25% of the dollars spent were on these first GP visits.

I think we all know that ChatGPT, a large language model, does a better job more consistently today at a first visit. What has to happen there to break through all that regulation and have AI make a true impact on the health care system?

There are several areas that we're involved in in health care.

One is AI physics, or AI biology: using AI to understand, represent, and predict biological behavior. That's very important in drug discovery. The second is AI agents, which provide assistance with diagnosis and things like that. Open evidence is a really good example. Hippocratics is a really good example. I love working with those companies. I really think this is an area where agentic technology is going to revolutionize how we interact with doctors and how we interact for health care.

The third part that we're involved in is physical AI. The first one is AI physics, using AI to predict physics. The second one is physical AI—AI that understands the properties and laws of physics—and that's used for robotic surgery. There are huge amounts of activity there. Every single instrument, whether it's ultrasound, CT, or whatever instrument we interact with in a hospital in the future, will be agentic. A safe version of OpenClaw will be inside every single instrument. In a lot of ways, that instrument is going to be interacting with patients, nurses, and doctors in a very unique way.

There's so much investment in AI weapons. It would be wonderful to see some investment in AI EMTs and paramedics, saving lives, not just taking them.

Yeah.

Which I think is a great segue into robotics. You've got dozens of partners. We had this very weird—I don't know what to call it, a lost decade or 20 years—of Boston Dynamics. Google bought a bunch of companies, then wound up selling them and spinning them out, and people just thought robotics wasn't ready for prime time.

Now here we have the world's greatest entrepreneur at this time, tied with you—Musk, doing Optimus. Well, that was a good save, I hope. Optimus is pretty impressive. And then there are other companies in China.

How close is that to actually being in our lives, where we might see a robotic chef, a robotic nurse, or a robotic housekeeper—these humanoid form factors actually working in the real world? Knowing what you know about those partners and the fidelity, especially in China, where they seem to be doing as good a job as we're doing here, or maybe better?

We invented the industry, largely. America invented it. You could argue we got into it too soon. We got exhausted. We got tired about 5 years before the enabling technology appeared. We just got tired of it a little too soon. Okay, that's number 1.

But it's here now. The question is, how much longer from the point of a high-functioning existence proof to reasonable products? Technology never takes more than a couple of 2- or 3-year cycle iterations. A couple of 2- or 3-year cycle iterations would basically be somewhere around 3 to 5 years. That's it. In 3 to 5 years, we're going to have robots all over the place.

Yeah.

I think China is formidable. The reason for that is because their microelectronics, their motors, their rare earths, and their magnets, which are foundational to robotics, are the world's best. In a lot of ways, our robotics industry relies deeply on their ecosystem and supply chain. They're obviously moving very quickly. Our robotics industry will have to rely a lot on it. The world's robotics industry will have to rely a lot on it. I think we're going to need some fast movements here.

Ultimately, one for one, Elon seems to think we're going to have 1 robot for every human—7 billion for 7 billion, 8 billion for 8 billion.

Well, I'm hoping for more. I'm hoping for more. First of all, there are a whole bunch of robots that are going to be in factories working around the clock. There are going to be a whole bunch of factory robots that don't move. They move a little bit. Almost everything will be robotic.

What does the world look like?

Sorry, let me say this: I think robotics is one of the pieces that unlocks economic mobility opportunities for every individual. When everyone got a car, they could go and do a lot of different jobs. When everyone gets a robot, their robot could do a lot of work for them. They can stand up an Etsy store or a Shopify store. They can create anything they want with their robot. They could do things that they independently cannot do.

I think the robot is going to end up being the greatest unlock for prosperity for more people on Earth than we've ever seen with any technology before.

Yeah, no doubt. The simple math at the moment is that we're millions of people short of labor today.

Right.

Yeah.

Right. We're really desperately in need of robotics, and all of these companies could grow more if they had more labor. Number 1, some of the things that you mentioned are super fun.

Because of robots, we'll have virtual presence. I'll be able to go into the robot in my house and virtually operate it while I'm on a business trip.

Right. Tell it to walk around the house.

And walk the dog.

Yeah, walk the dog.

Rake the leaves.

Yeah, exactly.

Rake up the dog.

Maybe not quite that, but just wander around and see what's going on in the house, chat with the dogs, chat with the kids.

Yeah.

Yeah.

That's an—

Time travel is also—we're going to be able to travel at the speed of light, you know, and so, you know, clearly want to send our robots ahead of us.

Yeah.

I'm not going to send myself. I'm going to send a robot.

Right.

Check it out.

Yeah, yeah.

And then I'm going to upload my AI.

Well, it's inevitable. It unlocks the moon and it unlocks Mars as targets for colonization, which gives us infinite resources. Getting back from the moon is effectively zero energy cost to move material back because you can use solar and accelerate. You could have factories that make everything the world needs on the moon, and the robots are going to be the unlock for enabling that.

This distance no longer matters.

Distance doesn't matter.

Yep.

Yeah.

The more revenue we get out of models and agents, the more we can invest in building the infrastructure, which then unlocks more capabilities in models and agents.

Dario, on Dwarkesh's podcast, recently said, "By 2027–2028, we'll have hundreds of billions of dollars of revenue out of the model companies and the agent companies." And he forecasts $1 trillion by 2030, right? This is non-infrastructure AI revenue.

I think he's being very conservative. I believe Dario and Anthropic are going to do way better than that.

Wow.

Way better than that.

Wow.

Yeah, and then $30 billion to $1 trillion.

Yep. And the reason for that is the one part that he hasn't considered: I believe every single enterprise software company will also be a reseller—a value-added reseller—of Anthropic code, Anthropic's tokens, and a value-added reseller of OpenAI.

That's right.

Get this: logarithmic expansion.

Yes.

Yeah.

Their go-to-market is going to expand tremendously.

What do you think, in that world, is the moat? What's left over? You have some moats that are, frankly, I think, as this scales, almost insurmountable. The best one that nobody talks about is probably CUDA, which is just an incredible strategic advantage.

But in the future, if a model can be used to create something incredible, then the next spin of a model can maybe be used to disrupt it. In your mind, what do you think, for these companies that are building at that application layer, their moat is? How do they differentiate themselves?

Deep specialization. Deep specialization. I believe that these models are going to have general models connected into the software company's agentic system. Many of those models are cloud models and proprietary models, but many of those models are specialized subagents that they've trained on their own.

Right. So the call to arms for entrepreneurs is: know your vertical. Know it as deeply and as well as everybody else.

That's right.

Know it as deeply and better than everybody else.

That's right.

And then wait for these tools, because they're catching up to you, and now you can imbue them with your knowledge.

That's right. And the sooner you connect your agent with customers, the sooner that flywheel is going to cause your agent to get—

It very much is an inversion of what we do today, because today we build a piece of software and say, "What generalizes?" Then we try to sell it as broadly as possible and sell the customization around it.

And we trap— In fact, exactly right. We create a horizontal, but notice there are all these GSIs and consultants who are specialists who then take your horizontal platform and specialize it into—

Exactly. And that's arguably a 5- or 6-times-bigger industry: the customization.

It is. Absolutely.

Yeah, the whole—

That very much is.

That's right.

Yeah, domain expert.

I just want to give you your flowers. I think it was 3 years ago you said, "You're not going to lose your job to AI. You're going to lose your job to somebody using AI."

And here we are. The entire conversation has revolved around this concept of agents making people superhuman, and the business opportunity and entrepreneurship expanding. You actually saw it pretty clearly. Have you changed your view?

Well, I guess this is the doomer, doomer, doomer, doomer, doomer. I'm not a doomer.

No, you can hold space for, I think, 2 ideas. One is that there are going to be—

That's the spiral J-Cal, we call it.

No, no. There, you can see—

Well, that's just because he doesn't hang out with me enough.

Well, we talk a little bit.

He will show up at your breakfast table, and he'll follow you around.

I'm not asking for it. I'm just saying.

He'll follow you around.

I'm not asking for it.

You can come with me and Tucker. We ski in Japan every January. We love it.

Me and Tucker

Road trip.

Okay.

There is going to be job displacement, and then the question becomes: Do those people have the fortitude, the resolve, to then go embrace these technologies? We're going to see 100% of driving done by humans go away. That's a beautiful thing, but we have to recognize that's 15 million people in the United States—10 to 15 million—who are employed in that way. And so that is going to happen, yes?

I think that jobs will change. For example, there are many chauffeurs today who drive the car. I believe that many of those chauffeurs will actually be in the car, sitting behind the steering wheel while the car is driving by itself. The reason for that is, remember what a chauffeur does: In the end, these chauffeurs are helping you. They're your assistants. They're helping you with your luggage, and they're helping you with a lot of things.

I would be surprised, actually, if the chauffeurs of the future become your mobility assistants and help you do a whole bunch of other stuff. The car's driving by itself.

The autopilot in planes created a lot more pilots and didn't take any of the pilots out of the cockpit, even though the autopilot is flying the plane 90% of the time.

And by the way, while that car is driving itself, that chauffeur is going to be doing a bunch of other work on his phone, and he's going to be making money doing other things.

Doing other stuff, coordinating a lot of things for you, getting—

The whole pie just grows in a way that—

So one of the things is that, yes, every job will be transformed. Some jobs will be eliminated. However, we also know that many, many jobs will be created. The one thing that I will say to young people who are coming out of school, who are concerned and anxious about AI, is: Be the expert at using AI.

Yes. Look, we all want our employees to be experts at using AI, and it's not trivial. Not trivial. Knowing how to specify, not to overprescribe, leaving enough room for the AI to innovate and create while we guide it to the outcome we want—all of that requires artistry.

You had this great advice when you were at Stanford, I think it was: “I wish you pain and suffering.” Do you remember that?

Yeah.

Fantastic. What's your advice to young people around what they should be studying if they're about to leave high school? Those are the kids who are really AI-native. They haven't made a decision about college, what to study, or whether to go to college at all. How do you guide those kids? What would you tell them?

I still believe in deep science, deep math, and language skills. As you know, language is the programming language of AI now.

Programming language.

As it turns out, it could be that the English major could be the most successful. And so, I think I would just advise that whatever education you get, make sure that you're deeply, deeply expert in using AI.

One of the things that I wanted to say with respect to jobs, and I want everybody to hear it, is that, in fact, at the beginning of the deep learning revolution, one of the finest computer scientists in the world, whom I deeply, deeply respect, predicted that computer vision would completely eliminate radiologists. The one field he advised everybody not to go into was radiology.

Ten years later, his prediction was 100% right. Computer vision has been integrated into 100% of the radiology technologies and radiology platforms in the world. The surprising outcome is that the number of radiologists actually went up, and the demand for radiologists is skyrocketing.

The reason for that is because everybody's job has a purpose and a task. The task that you do is studying the scans, but your purpose is to diagnose—to help the doctors help the patient diagnose disease. What's surprising is that, because the scans are now being done so quickly, they can do more scans, improving health care.

Yes.

Doing more scans more quickly allows patients to be onboarded a lot more quickly and treated a lot more quickly. And as it turns out, hospitals enjoy making money too.

Yeah.

Right.

They're doing more scans.

Reading more—

They're treating more customers and more and more patients. The revenue is going up, and guess what? They need more radiologists.

In a country that grows faster and where productivity increases, a wealthier country can put more teachers in the classroom, not fewer teachers in the classroom. You just give every one of those teachers a personalized curriculum for every student in the room. It makes them all bionic and leads to a lot more.

Every single student will be assisted by AI, but every single student will need great teachers.

Yeah. Amazing. Jensen, congratulations on all your success. This is an incredibly positive, uplifting discussion. We really appreciate you taking the time for us.

He is the steward we need.

You are the steward we need.

I think you need to be more vocal about it.

I'm being very, very vocal about the positive side of it. I think there's too much doomerism.

But I also think it takes humility to have this level of success and be humble about the fact that we're making software, guys. And I think that's actually really healthy for people to hear. We have done this before. We have invented categories and industries before. We don't need to go to this scaremongering place. It does nothing.

And we get to choose, right? We have autonomy and agency. We get to pick how to deploy this.

We sure do.

Okay, everybody. We'll see you next time on the All-In interview. Okay. Well done, brother.

Thanks, man.

Good job.

Thank you, sir. That was awesome.

Good. Good.

You guys are awesome.

Jason, thank you.

Look at this. Look at this big crowd behind you guys.

Man, I think they're here for you.

I'm going all in.