저자: JensenHuang · 외국인
요약 보기
엔비디아는 GPU 회사에서 AI 팩토리 회사로 전환 중이며, Groq LPU 추가로 추론 워크로드를 분산 처리한다. 에이전트 시대로 추론 연산량이 2년간 약 1만 배 늘었고 AI 인프라 TAM이 33~50% 확대된다고 주장. 피지컬 AI는 약 10억 달러 규모(연 100억 달러 수준)로 성장 중이며, 디지털 바이오는 5년 내 돌파 기대. 발언에 특정 종목의 가격 방향 예측은 없음.
다듬은 글
This week's special edition. We've put the weekly show on hold, and there are only three people we're doing it for. President Trump, Jesus, and Jensen. And, uh, I'll let you choose the order in which we do it . But what an amazing success you had, and a wonderful event. Uh-huh. Every industry is represented here.
Every technology company is here. Every artificial intelligence company is here. Incredible. Extremely. And one of the loudest announcements last year was Groq. When you bought Groq, did you realize how unbearable Chamat would become? I had, I had a feeling that, that, that ... We are his friends. We have to deal with him every week.
I know. You had to deal with him for six weeks of closing the deal. I know, it's like two weeks. Fortnight. All this comes to mind now. This, this makes me feel quite awkward. The fact is that many of our strategies are presented publicly at GTC years before we implement them. Two and a half years ago, I introduced the operating system of the AI factory, and it's called Dynamo.
The Dynamo, as you know, is a tool, a machine, created by Siemens to actually convert water into electricity. And Dynamo, uh , powered the factories of the last industrial revolution. So I thought this was the perfect name for the operating system of the next industrial revolution, for its factory. And inside Dynamo, the fundamental technology is disaggregated inference.
Jason, I, I know that you, you're a super technical expert. Certainly. I know. I'll let you take care of it. Come on, define it for the audience. I don't want to interrupt you. Yes, thank you. I, I, I know you wanted to insert a word, but this, this is disaggregated inference, which means that the pipeline, the inference processing pipeline is extremely complex .
In fact, it is the most difficult computational problem today. Incredible scale, lots of mathematics of all shapes and sizes. And we came up with, came up with an idea that you would change, you would, you would disaggregate parts of the processing so that part of it could run on some GPUs , and the rest on others.
And that led us to realize that maybe even disaggregated calculations make sense; that we could have different heterogeneous nature of computation. That same understanding led us to Mellanox. Yes. You know, today Nvidia's compute power is distributed across GPUs, CPUs, switches, scale-out switches, network processors, and now we're going to add Groq to that and direct the right workloads to the right chips.
You know, we just went from being a GPU company to being a company that builds AI factories. I mean, that was probably the most important takeaway for me. You see this fundamental disaggregation where we've gone from just GPUs to a whole suite of different options that will emerge over time. What you said on stage, or you said on stage: I would like the people who are doing high-value inference to listen to this, and 25% of your data center space, as you said, should be dedicated to this combination of Groq LPUs and GPUs.
Grock for about 25%Vera Rubin in the data center. So can you tell us how the industry is looking at this idea of creating a next generation disaggregated pre- fill and decoding process? So, let's take a step back: at the time we added this, we were moving from processing large language models to agent-based processing.
Now, when you run an agent, you access working memory and long-term memory. You use tools. You are putting a lot of strain on the data warehouse. You have agents working together with other agents. Some of the agents are very large models. Some of them are smaller models. Some of them are diffusion models.
Some of them are autoregressive models. So inside this data center there are all kinds of models. We created Vera Rubin to be able to handle this extremely diverse workload. My feeling is that we have added to what was previously a one-stop company. Now we've added four more racks, right? So Nvidia's TAM (total market size), if you will, has increased from what it was to something like, say, 33-50% more.
Part of that 33-50%, a significant portion of it, will be for processors for data storage. It's called BlueField. Some of it, I hope, a significant portion, will be Groq processors, some of it—CPUs , and all of that, and a large portion— network processors. So all of this will work, essentially, like an AI revolution computer called "agents," right?
The operating system of a modern, contemporary industry. What about built-in apps? You know, my daughter's teddy bear at home wants to talk to her . What's inside? Is this a specialized chip, or will a broader market eventually emerge with development tools that can be different for different use cases in the peripheral and embedded applications?
We think that at the highest level, if you take a step back and look at the problem, there are three types of computers. One computer is designed for training AI models, developing and creating AI, and the other is designed for evaluating it. Depending on the type of task you're facing, for example, if you look around, there are all kinds of robots, cars, and the like.
You must evaluate these robots in a virtual environment that replicates the physical world. So, it must be software that obeys the laws of physics. And this is the second computer. We call it the Omniverse. The third computer is a peripheral computer, a robotic computer. Such a robotic computer could be a car with an autopilot.
Another could be a robot. Another one might be a teddy bear. A small device like this for a teddy bear. One of the most important projects we are working on is turning telecommunications base stations into part of the AI infrastructure. This is a $2 trillion industry. Over time, all of this will turn into an expansion of AI infrastructure.
Radio stations will become peripheral devices, as will factories, warehouses— whatever. So, there are these three basic computers . All of them, you know, will be necessary. Jensen, last year I think you were ahead of the rest of the world in saying that inference wouldn't be limited to a thousand. Only last year.
Yes. Will it reach a million, or perhaps a billion Xs? Yes. That's right. And I think people thought it was an exaggeration back then, because the world was focused on scaling and learning. Here we are . Inference exploded. We are limited by our inference capabilities. You announced an "inference factory," which I think is cutting-edge and will be 10 times more productive than the next one.
But if you listen to the conversations around you, your inference factory will cost 40-50 billion, and the alternatives— specialized ASICs, AMD, and others— will cost 25-30 billion, and you will lose market share. So tell us: what are you seeing, what do you think about market share, and does it make sense for all these people to pay twice the price of what others are offering?
The main conclusion, the main idea is that the price of the factory should not be equated with the price of tokens, with the value of tokens. It is quite likely, and in fact I can prove, that a $50 billion factory will provide you with the lowest token value. And the reason is that we create these tokens with extreme efficiency—10 times higher.
You know, the difference between 50 billion: it turns out 20 billion—it's just land, energy, and shell, right? And besides, you have storage, networking, processors, servers, cooling anyway—the difference between whether this GPU costs full price or half price is not the difference between 50 billion and 30 billion.
Pick your favorite number, but let's say between 50 and 40 billion. This is a small percentage when a $ 50 billion data center actually has 10 times the bandwidth. Yes, Jess? That's why I said that even for most chips, if you can't keep up with the level of technology and the pace at which we work, even if the chips are free, it's still not cheap enough.
Yes. Can I, can I just ask a general strategic question? Yes. You run the most valuable company in the world. Next year, it will receive more than 350 billion in revenue. 200 billion of free cash flow. It is growing at such a crazy pace. How do you decide what to do? For example, how do you actually get information ?
I mean, you've become famous for these emails that people are supposed to send you, but how do you actually decide how to get an intuitive understanding of how the market is shaping up, where to double down, where to maybe back off, and where to take on a completely new area? How, how does this information reach you?
How do you decide these things? Ultimately, that's the CEO's job. Yes. And our job is to define strategy, define vision, define strategy. Of course, we're informed by great computer engineers, great technologists, great people throughout the company, but we have to shape this future. Well, part of it has to do with whether it's something incredibly difficult to do?
If it is not difficult to do, we should give it up . And the reason for that is that if it's easy to do, then of course, um... There are a lot of competitors, a lot of competitors. Is it something that has never been done before, that is incredibly difficult to do, and that in some way utilizes our company's special superpowers?
So I have to find a combination of factors that meets the standard, and ultimately we know that it will require a lot of pain and suffering. Yes. There are no great inventions that came about simply because it was easy to do, and on the first try—here we are. So if it's extremely difficult, no one has done it before , it's likely that you're in for pain and suffering, so it's better to enjoy it .
So can you just go over three or four of the longer-term things that you announced? and just talk about long-term viability, whether it's data centers in space , or what you're trying to do with ADAS in cars, or , you know, what you're doing in biology. just give us an idea of how you see some of these curves heading up in some of these long-term directions.
Perfectly. Uh, physical AI is a big category, as we think, and I just mentioned that we have three computing systems and all the software platforms based on them. Physical AI as a broad category. This is the first opportunity for the tech industry to embrace a $50 trillion industry that has been largely devoid of technology until now.
So we need to invent all the technologies necessary for this. I felt like it was a 10- year journey. We started 10 years ago. We see the breakdown happening now. This is a multi-billion dollar business for us. Now it's about $10 billion a year. So it's big business, and it's growing exponentially. So, that's number one.
I think in the case of digital biology, we are literally on the cusp of a Chat GPT moment for digital biology. We are about to understand how to represent genes, proteins, cells. We already know how to understand chemicals. So our ability to represent and understand the dynamics of the building blocks of biology is a two-, three-, five-year perspective.
In five years, I'm pretty sure that the healthcare industry is where digital biology will make a breakthrough. So these are some really great examples, and you see they're all around us. Agriculture, farming is currently showing signs of a breakdown. Without a doubt. Yes. Jensen, I want to take you from the data center to your desktop.
The company was largely built on enthusiasts, video gamers, and all those graphics cards in the beginning. And you mentioned in front of, I think, 10,000 people here, about Open Claude, about the Claude code, and about what a revolution agents have become; and especially enthusiasts, who are a source of great energy, we see a lot of innovation, breakthroughs, they want desktops.
You announced one here, I think it's a Dell 6800, it's a very powerful workstation for running local models, 750 gigabytes of RAM. Apparently, Mac Studio is sold out everywhere in my company. We are switching to Open Claude for everything. Friberg just became a Claude user. You also switched to Claude, I understand.
And you are obsessed with them. What does this "street" movement of creating open source agents and using open source on the desktop mean to you? Where is this all leading? Yes. Very good. First of all, let's take a step back. Over the past two years, we have seen essentially three tipping points. The first was the generative ChatGPT.
He opened artificial intelligence to the public, to our consciousness. But the fact is that the technology was in sight months before GPT. It was only when ChatGPT built a user interface around it and made it easy to use that generative AI took off. Now, generative AI, as you know, generates tokens for both internal and external consumption.
Internal consumption is the thinking that led to the ability to reason. o1 and o3. We continued the ChatGPT wave, grounded the information, forced AI not just to answer questions, but to answer questions in a more reasoned and useful way. We started to see OpenAI's revenue and economic model start to change.
Then the third point was that only within the industry did we see Claude Code—the first agent system that was very useful, really revolutionary , but Claude Code was only available to enterprises. Most people on the outside had never seen anything about Claude Code until Open Claude came along. Open Claude actually introduced into the mass consciousness the understanding of what an AI agent can do .
Mm. This is the reason why Open Claude is so important from a cultural perspective. And the second reason why this is so important is because Open Claude is open, but it formulates, it structures a type of computational model that actually reinvents computing as a whole. It has a memory system. It's like the short-term memory of a file system.
He has, he has, he has scale. Did you say skills or scale? Skills. Oh, skills. Theoretically, they have scale. Yes. Yes. Skills. So, first of all, he , you know, has the resources. He manages resources. He does the planning. Aha. True. And it is a cron job. He can create agents. It can, you know, decompose tasks and solve problems, just like a planner.
It has input/output subsystems. He can, you know, take input. He has a conclusion. It connects to WhatsApp. And it also has an API that allows you to run many types of applications called skills. Yes. These four elements fundamentally define a computer. Yes. So what do we have? For the first time, we have a personal computer with artificial intelligence.
Open source. This is open source. It works literally everywhere. So this is now, this is essentially the blueprint, the operating system of modern computing. Yes. And it will work literally everywhere. Of course, one of the things we had to help with is that when you have agent software, you have to make sure that it has access to sensitive information.
It executes the code. It can communicate with the outside. We have to make sure that all of this is regulated. All of this needs to be secure, and we need to have policies that give these agents two of the three options, but not all three at the same time. So on the governance side, we did our part, Peter Steinberger was here, and we have a whole mountain of great engineers working with him to help protect and maintain this so that it can protect our privacy and security.
Jensen, this paradigm shift makes some of the AI legislation passed across the country to regulate AI, and many of the proposed laws, effectively meaningless, right? Can you briefly comment on how this paradigm shift is quickly eroding many models of regulatory oversight for AI, which is becoming a very hot topic in politics right now?
Well, this is the part where we have to always be ahead of the politicians, and Brad, you're doing a great job of it. We need to be one step ahead and inform them about the state of the technology, what it is and what it is not . This is not a biological being. This is not an alien. It is not a conscious being.
Um, it's computer software. Yes. That's right. And it's not that we say we do n't understand it at all. This is not true. We do n't understand at all. We understand a lot about this technology. So, I think that, first of all, we need to continue to inform politicians and not let pessimism and extremism influence their vision and understanding of this technology.
However, we still have to admit that technology is developing very quickly, and we should not get ahead of it with political decisions too quickly. The risk we take as a nation, our biggest national security concern with AI, is that other countries adopt this technology while we are so angry, scared, or paranoid about it that our industries and our society do not take advantage of AI.
So what I'm most worried about is the spread of AI here in the United States. Can you elaborate more if you were in the Anthropic boardroom during those Department of Defense rumors ? This somewhat reinforces the idea that people didn't know what to think. It seemed to add another layer of resentment, fear, or just general distrust that people sometimes feel at the AI software level.
What do you think you would advise Dario and that team to do differently to change some of that outcome and perception? The first thing I would say about Anthropic is that their technology is incredible. We are a big user of Anthropic technologies. I really admire their attention to safety. I really admire their attention to safety.
Um, the culture with which they approached this. The technological excellence with which they approached this is truly fantastic. Um, I would say that the desire to warn people about the possibilities of technology is also really great. We just have to make sure we understand that the world has a spectrum, and that warning is good and intimidation is less good, right?
Um, and because this technology is too important to us, right? And I think it's okay to predict the future, but we have to be a little more careful. We need to have a little more humility, because we can't completely predict the future , and saying things that are quite extreme or catastrophic, for which there is no evidence, can be more harmful than people think.
And, of course, we are leaders in technology. There was a time when no one listened to us. But now, because technology is so important to the social fabric, is such an important industry, is so important to national security , our words matter , and I think we have to be much more careful. We should be more moderate.
We must be more balanced. We must be more sensible. Well, you know, I would nominate you. I think the industry needs to unite. 17% of AI popularity in the United States. So we see what happened to nuclear energy, right? We have effectively shut down the entire nuclear industry, and now China has 100 nuclear reactors under construction , while the United States has none.
We hear about moratoriums on data centers. Therefore, I believe that we need to be much more proactive in this matter. But I want to go back to this explosion of agent technology that you're seeing in your company, the efficiency and productivity gains in your company. There's a lot of debate about whether we're seeing a return on investment, right?
Both you and I, at the beginning of this year, the main question was: will there be revenues? Will income scale with intelligence? And then we had a kind of " Oppenheimer moment": $5-6 billion in a month from Anthropic in February. Do you think, looking into the future, because you announced a trillion dollars—you see a trillion dollar prospect just from Blackwell and Vera Rubin over the next few years.
When you see what's happening at Anthropic and OpenAI, do you think we're now on that curve where we'll see revenue scale the same way intelligence scales? When you look around, when you...I'll answer that in a couple of ways. When you look at this audience, you'll see that Anthropic and OpenAI are represented here.
But in reality, 99% of everything here is AI, and it's not Anthropic or OpenAI. Yes. Yes. And the reason for this is that AI is very diverse. I would say that the second most popular model as a category is open models. Number one is, yes, open source, open paths, open source code. OpenAI is number one.
Open source software—number two. Very far. Third—Anthropic. And that tells you something about the scale of all the AI companies that are here . So it's important to realize this. Let me go back and say a few things. First, when we moved from generative AI to reasoning, the amount of computation we needed increased by about a hundredfold .
Isn't that right? When we moved from reasoning to agent systems, the computations increased probably another 100 times. Now we see that in just two years, computation has increased 10,000 times. Meanwhile, people pay for information, but mostly they pay for work. Yes. Talking to a chatbot and getting answers is great, right?
helps me with my research. Incredible. But I'm really willing to pay for the work to be done . So, this is where we are now. Agent systems do the work. They help our software engineers perform tasks. And now, if you take this, you have 10,000 times more computation. You probably have 100 times more consumption now.
Yes. Aha. And we haven't even started to scale yet. We are absolutely at the million- fold mark. This, I think, is a great moment to talk about the number—the company has 20-30 thousand people, something like that. We have 43,000 employees. You know, I would say that 38,000 are engineers. A conversation we've had repeatedly on the podcast: "Oh my god, look at the use of tokens in our companies.
It's growing rapidly ." Some people ask, "Hey, when I join a company, how many tokens do I get because I want to be an efficient employee?" And you suggested, I think, during your 2.5-hour speech, a fairly long speech. well done you spent. If it had been done well, it would have been shorter. I just want you to have time to write in an hour and forty-five.
So you guys know there's no rehearsal, you just take it and do it. You take it and do it. Yes. Yes. So I just want you to know: I was writing the speech while I was giving it. Okay . So, you never know. But does that mean if we make approximate apologies. Rough estimates are $75,000 in tokens for each engineer or something like that .
So are you at Nvidia spending a billion or two on tokens for your engineering team right now? We're trying to... let me offer you a thought experiment. Let's say you have a software engineer or AI researcher who you pay $500,000 a year. We do this all the time. Okay, this happens all the time. This engineer with a salary of $500,000—at the end of the year I'll ask him how much he spent on tokens, and if he says $5,000, I'll go crazy or something.
So it's true, if that engineer with a salary of $500,000 hasn't used up at least $250,000 worth of tokens , I'll be deeply concerned, okay? It's no different than our chip designer saying, you know what, I'll just take a paper and a pencil, I don't think I'll need CAD systems. It's a real paradigm shift to start thinking about these star employees in this way .
It almost reminds me of what we learned in the NBA when LeBron James started spending a million dollars a year just on the health of his body and its maintenance. That's right. Here he is, at 41 years old, still playing. Really, hey, if these are incredible intellectual workers, why do n't we give them superhuman abilities?
What exactly will this lead to? If we project the situation two or three years ahead, what will be the performance of that star at Nvidia and what will they be able to achieve? What will they look like? Well, first of all, things like "wow, this is too complicated." That thought disappeared. "Oh, that's going to take a long time." That thought disappeared.
"We're going to need a lot of people." That thought disappeared. This is no different from the past industrial revolution. Someone says, "Wow, this building looks really heavy." Nobody says that. No one. "Wow, this mountain looks too big." Nobody says that. Right. Anything that is too big, too heavy , takes too much time, those thoughts, those ideas—all of that is a thing of the past.
It all comes down to creativity. That's right . What can you think of? Exactly. And this means that now the question is how to work with these agents? Well, it's just a new way of programming. In the past, we wrote code. In the future, we will write ideas, architectures, specifications. We will organize teams.
We will help them determine how to judge the difference between good and bad . What does a great result look like? How to iterate with them, how to brainstorm. This is what you are really looking for. And I think every engineer will have a hundred agents. Getting back to the PR problem the industry has right now.
You have leaders like David Friberg at Oho, who is trying to literally use technology, your technology and AI, to increase the number of calories produced and make them quality; What do you think the cost reduction ratio is, and how does this vision influence what you do? Genomic modeling with zero training, and it works.
And there comes a moment when you say: "Oh my God. " Honestly, this is what happens after people replace entire software stacks overnight. I did something in 90 minutes, which I was telling the guys about: I replaced the entire software stack and a bunch of workloads in 90 minutes in the cloud, I got this agent system up and running, I built everything, I deployed it, and that was on a Sunday night .
Sunday night, at 10pm, I finished at 11:30. I went to bed. As CEO, you replaced ... Yes. And everyone on my management team had to complete a similar exercise over the weekend. When we saw it on Monday, I said, " This is the end." But the technical stuff, the scientific stuff—we did something in 30 minutes with auto research, and I'd like to hear your thoughts on auto research and what that tells us about how much more we need to do in terms of efficiency.
Using auto research and some of the data, we published something like this internally, and then we said, "Oh my god." Normally, this would be a PhD thesis, which would take seven years to complete. This would be one of the most prominent papers we've seen in this field, it would be in the journal Science, and it was done in 30 minutes on a desktop computer, running auto research with all the data we just downloaded.
We got this on Friday and thought, " Hey, let's give it a try." We tried to run , went to GitHub, downloaded Auto Research and ran it. And you see everyone's faces changing, and the potential of what this opens up for us is something that would have taken seven years, but happened in 30 minutes, and we feel it in genomics, and we're like, " This is incredible." So I think this acceleration is expanding opportunities for everyone in ways you could n't have imagined a few years ago.
But going back to the auto research point, can you comment on the fact that it was released with 600 lines of code over the weekend, and its ability to run locally and get results with all these different datasets, and what that tells us about the early stages we're at in terms of optimizing the algorithms and the hardware?
The main reason why Open Claw is so incredible is, first, its coincidence, its synchronicity with breakthroughs in major language models. Yes, his timing was perfect. It was flawless. In many ways, Peter probably would n't have come up with this if Claude, GPT, and ChatGPT hadn't reached a level that's really, really good, right?
It's also a new capability that allows these models to use the tools that we've built over time: web browsers, Excel spreadsheets, and, you know, in the case of chip design—Synopsis, Cadence, Omniverse, Blender, Autodesk, and all of those tools will continue to be used. Some say the enterprise IT software industry will be destroyed.
Let me offer an alternative view. The enterprise software industry is limited by the number of users. Soon there will be a hundred times more agents working with these tools. These will be agents working with SQL. There will be agents working with vector databases, agents in Blender, agents in Photoshop.
The reason is that these tools, first of all , do a very good job. Secondly, these tools are the connecting link between us in the end. When the work is done, it should be presented to me in a way that I can manage. Is n't that right? And I know how to operate these tools. So I need everything to be returned to Synopsis.
I want everything back in Cadence because that's how I control it. This is my foundation of truth. Let me ask you about open source software. So we have these closed-source models. They are wonderful. We have these models with open scales. Many Chinese models are incredible. Absolutely incredible . Two days ago, you might not have seen it because you were busy on stage, but there was a training run on this crypto project called Bit Tensor Subnet 3.
They managed to train a 4 billion Llama model in a fully distributed way with the help of many people who provided excess computing power, but they were able to do it with state support and managed training, which I think is a pretty crazy technical achievement. Yes. Because it's like random people, and each person gets a small share.
Our modern version of Folding at Home. Exactly. So what do you think about the end state of open source? Do you see this decentralization of architecture, as well as decentralization of computation, to support open scales and a fully open approach to ensure broad availability of AI for all? I believe that we fundamentally need both first-class proprietary product models and open source models.
It's not a question of choosing "A" or "B." These are "A" and "B". There is no doubt about it. And the reason is that models are a technology, not a product. The model is a technology, not a service. For the vast majority of consumers, for the horizontal level, for the general intelligence, I would really, really like not to engage in independent customization.
I would really like to continue using ChatGPT. I like using Claude. I like using Gemini. I like using X. And they all have their own personalities, as you know, which depends on my mood and what problem I'm trying to solve. You know, I can do it in X or in ChatGPT, so that segment of the industry is thriving.
This will be great. However, all of these industries, their industry expertise, their specialization, need to be directed and captured in a way they can control, and this is only possible through open models. The open model industry, to which we make a huge contribution, is on the verge of progress.
And honestly, even if it reaches that point, I think products as a service, premium models as a product, will continue to thrive. Every startup we invest in now is open source-focused at first and then moves to proprietary models. Yes. And the best part is, thanks to the great router you connect to from day one, you'll have access to the best model in the world every day.
This gives you time to optimize costs, fine-tune, and specialize, so you ’ll have first-class capabilities for any task. Jay, can I have it? Of course. No one wants the US to win the global AI race more than you do, right? But a year ago, the Biden-era rules of proliferation were effectively anti-American AI proliferation around the world .
So, we've been under a new administration for a year now. Rate us. Where are we in terms of global deployment and the pace at which we are spreading American AI technologies around the world? Are we on "A "? Are we on "B"? Are we on "C"? What works? What's not working? Well, first of all, President Trump wants American industry to be a leader.
He wants the American technology industry to be a leader. He wants the American tech industry to win . He wants us to spread American technology around the world. He wants the United States to be the richest country in the world. He wants all of this . At this point, as we speak, Nvidia has given up 95% market share in the second largest market in the world, and we have 0%.
President Trump, it's true. President Trump wants us to go back there. And the first thing we need to do is get licenses for the companies we can sell to. We have many companies that have submitted requests for licenses. We applied for licenses for them and received approved licenses from Secretary Lutnik.
We have now notified Chinese companies and many of them have sent us orders, so we are now in the process of restoring our supply chain to begin shipping. I think, at the highest level, Brad, I think one of the things we have to recognize is this. Our national security is weakened when we don't have access to miniature engines , rare earth minerals.
It weakens when we do not control our telecommunications networks. It weakens when we cannot provide our country with sustainable energy. It is fundamentally weakening. Each of these industries is an example of what I don't want the AI industry to become. Is n't that right? When we look into the future and ask ourselves, what do we want?
What does it look like when the American tech industry, the American AI industry, leads the world? We can all agree that it is impossible for AI models to be one size fits all. We can all admit that this is a result that makes no sense. But we can all imagine that the American technology stack— from chips to computing systems to platforms—is widely used in a world where they build their own AI, use public AI, private AI, whatever, and can create their own applications for their society.
I would like American tech stack to cover 90% of the world. Yes, I would really like that. The alternative, if it's like solar energy, rare earths, magnets, engines, telecommunications—I think that's a very bad outcome for national security. Perfectly. Yes. How closely do you monitor the conflict situation around the world right now?
And how much does that bother you, Jensen? So, China and Taiwan, as well as the availability of helium from the Middle East, I understand, could pose a risk to the supply chain in semiconductor manufacturing. How much do these situations bother you? How much do you spend on solving them? Well, first of all, I think we have 6,000 families in the Middle East.
Yes. We have many Iranians at NVIDIA whose families are still in Iran. So, we have a lot of families there. First of all, they are very worried. They are very worried, very scared. We think about them all the time. We are constantly monitoring the situation and do not let them out of our sight. They have our full support.
I have been asked several times if we are still considering a presence in Israel? We are 100% in Israel. We 100% support the families there. We are 100% in the Middle East. I was also asked if we consider the Middle East as a region where we can expand the scope of artificial intelligence, given the events there?
I believe that the war has reasons, and I believe that after it ends, the Middle East will be more stable than before. So if we were there and considered it before, we should certainly consider it after . I 100% support this opinion. Regarding Taiwan, we must do three things. First, we must ensure the rapid reindustrialization of the United States.
Whether it's chip factories, computers, or artificial intelligence factories, how are we doing with this ? We have had great success thanks to the strategic support and friendship with Taiwan's supply chain. Thanks to their friendship and support, we were able to build factories in Arizona, Texas, and California incredibly quickly.
They are truly a strategic partner. They truly deserve our support. They deserve our friendship. They deserve our generosity, and they are doing everything they can to speed up the production process for us. I think this is point number one. Point number two: we must diversify the production supply chain.
Whether it's South Korea, or Japan, or Europe—we need to diversify the supply chain, make it more sustainable. And point number three: let's exercise restraint . As we increase our diversity and resilience, let's not push, you know, unnecessarily, we need to be patient. Is helium a problem? Many reports say that I think helium could be a problem, but at the same time, there is probably a significant margin of safety in the supply chain.
Such things usually have a considerable stock. Uh, but-but, um, you know, yeah. You have made huge progress in autonomous driving. You made a loud statement. You have added many more partners, including BYD. There was just a video of you driving a Mercedes and a big announcement from Uber that you will have many cars on the roads from different manufacturers.
Your bet, as I understand it, is that there will be an open platform like Android, where you will play a major role with dozens of automakers, and on the other side there could be iOS with Tesla or Waymo. What is your strategy on this and what does that chessboard look like, because it seems like you have a pretty deep stack and you're competing in some ways and collaborating in some ways.
Yes. Um, if we take a step back. We believe that everything that moves will one day become fully or partially autonomous. Firstly. Secondly, we don't want to build autonomous cars, we want to enable every car company in the world to build their own autonomous cars. Therefore, we created all three computers: training, simulation, evaluation, and also a computer for the car itself.
We are developing the world's most secure operating system for management. We also created the world's first autonomous car capable of logical thinking, so it can break down complex scenarios into simpler ones that it knows how to navigate, similar to our reasoning systems. So this reasoning system called Alpommyo allowed us to achieve incredible results.
We open it up, we vertically optimize. We innovate horizontally and let everyone decide. Do you want to buy one computer from us ? In the case of Elon and Tesla, they buy our training computers. Um, do they want to buy our training and simulation computers, or do they want to work with us to use all three, and even put a computer in your car.
So our approach is this: we want to solve the problem. We are not the only solution providers. And we are happy to cooperate with us in any format. Let me continue this question, because it seems so fascinating to me. You are really building this platform. A thousand flowers bloom. But it's also true that some of those flowers now want to move down the stack and try to compete with you a little bit .
Google has TPU, Amazon has Inferentia and Trainium. You know, everyone is trying to create their own version of "I think I can beat Nvidia." even though they are also usually huge customers. How do you deal with this ? And so, what do you think will happen over time and what place does this have in the overall picture of such a vision?
Yes, that's great. You know , first of all, we are the only AI company, we are an AI company. We create basic models. We are at the forefront in many areas. We build every single level, every single stack. We are the only AI company in the world that collaborates 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.
True. Yes. So, confidence stems from that. We're excited to compete for the title of best technology , and as we continue to move forward rapidly, I believe buying from Nvidia is still one of the most cost-effective options for them, and that's just incredible confidence. First of all . Second, we are the only architecture that can be in every cloud, and this gives us some fundamental advantages.
We are the only architecture that can be taken from the cloud and deployed locally, in a car, in any region in space. That's right. In space. And so there's a whole segment of our market, about 40% of our business, most people don't realize this, 40% of our business—if you don't have a CUDA stack, if you can't build a whole AI factory, customers don't know what to do with you.
They are not trying to create chips. They are not trying to buy chips. They are trying to build AI infrastructure. So they want you to come in with a full stack. And we have the whole stack. And so, not surprisingly, Nvidia is increasing market share. If we look at the current situation, we are increasing our share.
You think they're trying, and then they realize, "Oh my god , this is too hard." And then they come back? Is that why the share is increasing? Well, we are increasing the share for several reasons. First, our speed has increased. We help people realize that it's not about building a chip, it's about building a system.
And this system is very difficult to build. And that's why their business with us is growing. In the case of AWS, I think they just announced, it seems like yesterday, that they're going to buy a million chips over the next few years. I mean, that's a lot of chips for AWS. And that's on top of all the chips they've already bought.
So we're happy to do it. First, we've been growing our share over the last few years, because now Anthropic is coming to Nvidia. Meta Llama is coming to Nvidia. And the growth of open models is simply incredible. And all this on Nvidia. So we are increasing our share through the number of models. We're also growing because all of these companies are outside the cloud, they're growing regionally in the enterprise sector, in the industrial sector, in the periphery, and this growth segment is very difficult to capture if you just create as Brad said in this regard, and without going into the numbers, analysts don't seem to believe it.
So if you look at the consensus forecast, you said that computing could grow 1 million times, but they're predicting you a 30% growth next year, and a 20% growth the year after that. And in 2029, which should be incredibly successful, by 7%. Isn't that right? So, if you take your TAM and apply their growth rates, your share will drop dramatically.
Do you see anything in your order book that would confirm this? Yes. First, they simply don't understand the scale and reach of AI. Yes. Yes. I think that's true. Most people think that AI is only about the top 5 hyperscalers, right? That's right. There is also a certain orthodoxy around the law of large numbers, where they have to go back to their investment banking risk committee and show some kind of pattern.
They just can't believe that 5 trillion will turn into 15 trillion. They say, well, it could go up to seven, or they could have a 10 trillion company. These are all just theories that I believe have never happened before. So you can't claim that it will happen, and because of that you have to rethink what exactly you're doing.
Someone recently remarked: Nvidia, Jensen, how can you be bigger than Intel in the server segment? And the reason is that the processor market for the entire data center was about $25 billion a year, right? We're making $25 billion a year, as you know, while we're sitting here. And so, of course, of course , it was a joke.
No, but this is the "All-In" podcast. Do n't worry. Everything in this show is freeform. Don't worry about it. Everything is here. In any case , it was not a prediction. But anyway, the bottom line is how big you can become depends on what you produce, right? Nvidia doesn't just make chips. First, building chips no longer solves the problem of AI infrastructure.
This is too complicated. Third, most people believe that AI is limited only by what they say, hear, and see. AI is much broader, OpenAI is incredible. They will become huge. Anthropic—incredible. They will become huge. But AI will be much, much bigger than that. Tell us and we will review this segment.
Tell us about data centers in space for a minute. Yes. Ahem. We are already in space. How should an ordinary person think about this business compared to those large data centers being built on Earth? Well, we should definitely work on Earth first, because we're already here, that's the first thing.
Secondly, we have to prepare for space travel, and obviously there is a lot of energy in space . Um, the challenge, of course, is cooling: you can't use conduction and convection, so you can only use radiation, and radiation requires very large surfaces, and that's not impossible to solve, and there's a lot of space in space .
Um, but despite everything, the costs still exist, and we're going to investigate that. We are already there. We are already resistant to radiation. We have CUDA in satellites all over the world. They do imaging, image processing, AI visualization, and that should be done in space, instead of sending all the data here and processing it here.
We should just do image processing in space. So there are many things we have to do in space. In the meantime, we will explore what data center architecture looks like in space. And it will take years. This is normal. We have a lot of time. I would like to dwell in more detail on healthcare. I know you have a lot of experience there.
We are all at that age where we start thinking about life expectancy and health. I mean, we all look great. I think some are better than others. I think some are better than others. I don't know what your secret is, Jensen. You look pretty good, these... I mean, what are you accepting, what did you give up?
You have to tell me behind the scenes. I want to know what you're doing in the dressing room. Squats, push-ups and abdominal exercises. Perfectly. Good. Hmm, but it works. What do you know about the development of the healthcare sector , where everything is heading and what progress are we making? I was just using Claude for analysis, trying to figure out where all these account codes are coming from ?
We in the US spend twice as much money. And we seem to get half as much. It seems that 15-25% of the money spent is on these initial visits to the therapist. And I think we all know that ChatGPT and the big language models today are better and more stable at handling initial review. So what needs to happen to overcome all these regulations and allow AI to truly impact the healthcare system ?
We are engaged in several areas of healthcare. One of them is AI. This is AI physics or AI biology— using technology to understand, represent, and predict biological behavior, which is very important for drug development. The second is AI agents that help with diagnostics and other similar tasks. Open Evidence is a great example of this.
Hypocratic is also a very good example. I like working with these companies. I truly believe this is an area where agent-based technologies will revolutionize the way we communicate with doctors and receive healthcare services. The third area we are working on is physical AI. The first is AI in physics, which uses models to predict physical phenomena.
The second is physical AI. AI that understands the properties of the laws of physics, used, for example, in robotic surgery, where a huge number of operations are performed. Every tool, whether it's an ultrasound machine, a CT scan, or any other device that we interact with in a hospital, will become an agent in the future.
Yes. A secure version of a model similar to Claude will be built into every tool. And in many ways, this tool will interact with patients, nurses, and doctors in a completely unique way. So much investment in AI weapons . It would be great to see investment in AI for ambulances and paramedics to save lives rather than take them, which I think is a great transition for robotics.
You have dozens of partners. We have this very strange situation. I don't know whether to call this a lost decade or 20 years of Boston Dynamics. Google bought a bunch of companies. Then they sold them off and spun them off into separate businesses when everyone thought robotics was simply not ready for mass adoption.
And now we have the most outstanding entrepreneur of our time . Together with you, Elon Musk is achieving success, I hope this was an apt remark. Optimus, quite impressive. And also other companies in China. How close is this to being implemented in our lives, where we see a cook, a robot chef, a robot nurse, a robot housekeeper, you know, this humanoid factor that actually works in the real world, knowing what you know about those partners and the accuracy, especially in China, where they seem to be doing just as well as we are, or even better?
Um, we mostly invented this industry. America invented. We, one might argue, started too early. Yes. And we were exhausted. We got tired about five years before the relevant technologies appeared. Brain. Yes. Yes. And we just got tired of it a little too quickly. OK. This is the first thing. But now it's here.
Now the question is, how much more time? From the moment of proving the existence of a highly efficient technology to the creation of smart products, it never takes more than two or three cycles. So, two or three cycles—that's about three to five years. That's all. In 3-5 years, jobs will be everywhere.
I think China is formidable, and the reason is that their microelectronics, their motors, their rare earth metals, their magnets, which are the basis of robotics—they are the best in the world. And so in many ways, our robotics industry is deeply dependent on their ecosystem and supply chain. And they, of course, move very quickly.
Our robotics industry will have to rely heavily on this. The global robotics industry will have to rely heavily on this. So I think you'll see rapid shifts here. Ultimately, one to one. Elon seems to think we'll have one robot for every human. 7 billion by 7 billion, 8 billion by 8 billion. Well, I hope for more.
Yes, I hope for more. Yes. First of all, there will be a whole bunch of robots working in factories 24/7. There will be a whole bunch of those that don't move. They move only a little. Almost everything will be robotic. What does this world look like? Excuse me, I think that robotics for me is one of those elements that, in my opinion, opens up opportunities for economic mobility for every person.
Just as once everyone got a car and was able to do a lot of different jobs, so it will be now. When everyone has a robot, it will be able to perform many tasks for its owner. They will be able to open a store on Etsy or Shopify. With the help of a robot, they will be able to create anything. They will be able to do what they are unable to do on their own.
I think robots will be the biggest boost to prosperity for more people on Earth than any technology before. Yes, without a doubt. Simple math shows that we are currently short millions of workers. Yes. Yes. Right. We actually desperately need robotics, because all these companies could grow faster if they had more workers.
I mean, we're in first place. Some of the things you mentioned are very interesting. Through robots, we will have a virtual presence. I will be able to connect to the robot in my home and control it virtually. For example, I'm on a business trip, right? Walk around the house. Yes. Walk the dog. Raking leaves.
Yes. That's right. To scare a dog. Maybe not exactly , but just walking around and seeing what's going on in the house. Talk to dogs, talk to children. Yes. Yes. And time travel: we will be able to travel at the speed of light, so we will certainly send our robots ahead of us. Yes. I will not send myself .
I will send a robot. Check everything. Yes. Yes. And then I'll upload my AI. Well , it's inevitable. This opens up the Moon and Mars as targets for colonization, which would give us limitless resources. Returning from the Moon requires virtually no energy expenditure to transport materials, as solar energy can be used for acceleration.
So it is possible to create factories on the Moon that will produce everything the world needs, and robots will be the key to this. That's right. Distance no longer matters. Distance doesn't matter. Yes. The more revenue we generate from models and agents, the more we can invest in infrastructure, which in turn opens up new opportunities for models and agents.
Dario on the Dwaresh podcast recently said that by 2027–2028, we will have hundreds of billions of dollars in revenue from modeling companies and agents. And he predicts a trillion dollars by 2030, right? This is AI revenue, not infrastructure-related. Um , I think he's, I think he's very conservative.
I believe Dario and Anthropic will do much better. Oho. Much better than that. Oho. So, from 30 billion to a trillion. Yes. And the reason for that is because he missed one thing: I believe that every enterprise software company will also become a value-added intermediary for Anthropic, their code, their tokens, or for OpenAI, that's right, and they'll get this logarithmic expansion.
Yes. Their market penetration will expand significantly this year. What do you think will be the " moat" (protective barrier) that will remain in such a world? I mean, there are some barriers that, frankly , I think, become almost insurmountable with scaling; The best one that no one 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 version of the model can be used to, perhaps, destroy it. What do you think can be said about companies that build at the application level? What is their barrier? How do they distinguish themselves? Deep specialization.
Deep specialization. I believe these models, they will have generic models that are connected to the software company's agent system, right? Many of these models are cloud-based and proprietary models, but many of them are specialized subagents that they have trained themselves. Yes. OK. So, a call to entrepreneurs: listen, know your niche.
That's right. Know her deeper and better than anyone else. That's right. And then wait for these tools, because they catch up with you, and now you can empower them with your knowledge. That's right. The faster you connect your agent, the faster you connect your agent to customers, that flywheel will make your agent become this is very much like the inversion of what we do today, because today we build software and say it generalizes and then try to sell it as broadly as possible and then sell customization around that.
And we, actually, that's exactly what we do , we create a horizontal solution, but notice, there are all these GSIs and all these consultants who are specialists, who then take your horizontal platform and specialize it under and this is probably a five or six times bigger industry—this is customization.
That's absolutely right , so I think these platform companies have the opportunity to become those specialists, to become that vertical, right? yes, experts in the field You know, I just want to express my respect to you. I think three years ago you said, " You won't lose your job because of AI; you'll lose it because of the person using AI." And here we are: the entire discussion has centered on the concept of agents making people superhuman, expanding opportunities for business and entrepreneurship.
You actually saw it quite clearly. You changed your mind. I was, I'm not a pessimist, but I have my share of pessimism. No, you can combine, I think, two ideas. One is that there will be a lot of this spiral, Jake, that's what we call it. No, no, you can, but that's only because he doesn't spend enough time with me .
Well, we, I mean , we communicate a little bit . We're not talking about that. He will show you your breakfast. He will follow you everywhere. I'm not asking for this. I'm not following you. I'm not asking for this. You can come with me and Tucker. We go skiing in Japan every January. Perfectly. Tucker and I are going on a road trip.
Hmm, there will be job displacement . And then the question becomes, you know, do these people have the resilience, the determination, to then go and adopt these, well, technologies? We will see human driving disappear 100%. This, well, this is great in terms of the lives saved, but we have to realize that this is 15 million people in the United States, 10 to 15 million of whom are employed in this field, and so this is going to happen.
Yes, I think, I think jobs will change. For example, there are many drivers who drive cars today. I believe that many of these drivers will actually stay in the car, sitting behind the wheel, while the car drives itself. And the reason for this is because remember what the driver ultimately does . These drivers, they help you, they are your assistants.
They help you with your luggage. They help you, I mean, they help you with a lot of things, and so I wouldn't be surprised if the drivers of the future become your mobility assistants, helping you do a bunch of other things. to the hotel. Yes. And the car drives by itself . Autopilot in airplanes has created many more pilots and has not removed any pilots from the cockpit, even though the autopilot controls the plane 90% of the time.
And by the way, while the car is driving itself, the driver will be doing a bunch of other things on the phone, and he will be negotiating, like coordinating a bunch of things for you, getting, you know, that's all, the pie just grows in a way that one of the things that yes, every job will be transformed.
Um, some professions will disappear. However, we also know that many, many jobs will be created. The only thing I would say to young people who are graduating, who are worried, who are anxious about AI: become experts in using AI. as much as, listen, we all want our employees to be experts in using AI, and it's not trivial, it's not trivial, and so it's important to know how to set tasks, not overload with instructions, leaving room for AI innovation and creativity while we direct it to the desired outcome.
it. All this requires skill. You gave some great advice when you were at Stanford, I think, which was, "I wish you pain and suffering." Do you remember this? Yes. Fantasy. What is your advice to young people about what they should study? So, if they're going to graduate from school, because these are the kids who are in such a native state right now, they haven't made a decision yet about college, what to study, or whether to go to college at all.
How do you guide these children? What would you tell them? I still believe that deep science, deep mathematics, language skills, you know, language—it's the programming language for AI, the ultimate program. And, as it turns out, it may happen that philologists will become the most successful. Yes. And so I think what I would just advise is, whatever education you get, just make sure you have a deep, deep understanding of AI.
One of the things I wanted to say about jobs, and I want everyone to hear this, is that actually at the beginning of the deep learning revolution, one of the best computer scientists in the world, who I respect deeply, very deeply, predicted that computer vision would completely replace radiologists, and that the one field he advised no one to go into was radiology.
After 10 years, his prediction came true 100%. Computer vision has been integrated into all radiology technologies and platforms in the world at 100%. The surprising result is that the number of radiologists has actually increased, and the demand for them has skyrocketed. The reason for this is that everyone's work has a purpose and a goal.
The task you perform is to study the images, but your goal is to help doctors, to help the patient diagnose the disease. And what's amazing is that because the scans are now so fast, they can take more pictures, improving healthcare. Yes. But doing more scans faster allows patients to be admitted much faster, to receive treatment much faster.
And as it turns out, it's because hospitals like to make money too. Aha. Yes. They take more pictures. They treat more clients. Incomes are growing. And you know what? Ideally. And a country that develops faster increases its productivity. A richer country can get more teachers into classrooms, not fewer.
This is true. You simply provide each of these teachers with a personalized curriculum for each student in the classroom. This makes them " bionic" and leads to much more. Every student will be supported by AI, but every student will need great teachers. Yes. Yes. Surprisingly. Jensen, congratulations.
I know about your success, and this is truly an incredibly positive , inspiring discussion. We really appreciate you taking the time to talk to us. He is the leader we need. You have to, you have to be more public. I'm very open about the positive side of this. I think there's too much pessimism, but I also think you have to have humility to achieve this level of success and be humble, saying: we're just building software, guys.
Yes. And I think that's actually really helpful for people to hear. We've done this before. We've invented categories and industries before. We don't need to panic. This does nothing. And we can choose, right? We have autonomy and freedom of action. We can decide for ourselves how we can definitely use it.
Okay, everyone. See you next time for the All-In interview. OK. Great job, brother. Thanks, man. Good job. Thank you, sir. That was cool. Good. Good. I appreciate you. You guys are incredible. Look at this. Look at that big crowd behind you, man. I think they are here for you.
Comments 0