Transcription
Good morning everyone. Um, my name is Constantine Buer and I'm a partner at Sequoia Capital focused on AI investing.
Both Nvidia and Citadel Securities actually have a lot in common. Uh, they're both exceptional businesses. They're really well-run, really brilliant leaders. Um, they are exceptionally well-run. They were both powered by the computing revolution and both are leaders in their respective industries with technology. They also have another lesser-known fact: in both cases, their first outside investor was Sequoia Capital. So when $1 million, they, they risked $1 million in Nvidia in 1993. You were worth it. One solid million dollars went way out on the limb. It was a little more into Citadel. Um, so when we were asked to speak about AI at this conference, it was incredibly clear who the best person in the world to speak would be. It's the man who has built the entire infrastructure for the AI revolution upon which all of this AI rests and the man who built the most valuable company in the world. Please join me in welcoming Jensen Huang.
Oh, thanks. It's nice to wake up this way.
Well, you've been working for hours.
Yes, I have. Um, so, uh, Jensen, we have a room full of institutional investors who are some of the best in the world. They manage many trillions of AUM and they are constantly looking for an edge. Uh, you are someone who always has an edge, and in every one of our conversations, you have compelling insights about what the future is going to look like. In the next 60 minutes, we have an ambitious agenda, uh, to cover stories of the edge from the very beginning of NVIDIA all the way through its rise to the center of the AI revolution. And then we'll spend the majority of the time on what's next for Nvidia and AI.
Okay. So, let's start at the very beginning. It's 1993. You're 30 years old. What's the insight that gave you the edge to start Nvidia?
We were going through, uh, uh, basically the PC revolution and the revolution of CPUs. It was the era of Moore's Law. It was the, the time when, uh, integrated microprocessors, Intel, uh, Moore's Law, uh, the scaling, the scaling laws of transistors, uh, that was the buzz, and that was nearly all of the investment dollars in Silicon Valley in the computer industry. And, and, uh, we observed, uh, something a little different. We said, uh, there are many problems that, one of the benefits of the CPU is general purpose, but the, the fundamental problem of general purpose technologies is that they tend not to be very good, extremely good at very hard problems. And so we conjectured two things. One, we observed that that, um, there are problems that we could solve with, uh, an accelerator, uh, that is more domain-specific, more targeted, and those problems could be interesting to solve. And we observed that that general purpose technologies, the shrinking of these transistors would eventually reach the limits. The idea that you could keep reducing the size of transistors and scaling it using this, this technique. It's, it's a, it's a set of heuristics called Dennard scaling, and Dennard scaling, and, uh, me and Conway came up with what are really the fundamental principles behind Moore's Law. And if you go back to those, you'll, you'll discover that that there will be a limit to how far you can shrink transistors, and at some day, uh, you'll get, you'll get diminishing returns. And there are large computing problems. We believe that the computing problems that we could solve are nearly infinite in scale. And so, one of these days, a new type of computing approach would emerge. And we, you know, we, uh, focused our company on, on, um, on augmenting, supplementing, uh, general purpose computing with this technology called accelerated computing. And so that was really the, the observation.
And, and you said something earlier about about, um, how, how Nvidia is always ahead of the curve. Mhm. Um, often times, often times if you reason about things from first principles, what's, what's working today incredibly well, uh, if you could reason about it from first principles and ask yourself, on what foundation is that first principle built on top, and how would that change over time, it allows you to, to, you know, hopefully see around corners.
So when you built the graphics accelerator, yeah, uh, you were early to the party, but then hundreds of other competitors sprung up. Yeah. You eventually won in that market. In the early 2000s, you said, "Hey, this technology might be able to generalize itself." You're talking about the generalization of a CPU. Perhaps the GPU could also be generalized for more processing. Let's talk about CUDA. Yeah. How did that come about? Where did you get that insight?
The story goes that it's from researchers. How did you read their work and conclude that the GPU could be a general computing device?
Well, the, um, so first of all, the, the reason why Nvidia was hard to build was because we had to invent a new technology and invent a market. M. And at the time in 1993, uh, in order to create a new computing platform, you need a large market. And Silicon Graphics, which was doing 3D graphics at the time, the markets were too small to enable a new computing platform. And, and so if we wanted to create a new computing architecture, we needed a new, a large market. And that large market didn't exist because the architecture didn't exist. You got the chicken and the egg problem. And so, some, what, what Nvidia became good at, and, and the modern, modern 3D graphics video game market, we, we contributed tremendously to. And so Sequoia Capital's big issue at the time with Nvidia's, uh, funding principle is that we had to go invent the technology and the market simultaneously. And the odds of that happening is approximately 0%. And I still remember when I pitched the story, and I said, and, and Don Valentine at the time, he says, "Well, Jensen, where's your, where's your app? Where's the killer app?" And I said, "I was, I..." Oh, yeah. "There's this company called Electronic Arts." And I didn't realize, uh, Don had just invested in Electronic Arts. And I said, "Electronic Arts? They're going to create, we're going to help them create 3D graphics games, and, and we're going to create this market." And he goes, "You know, Jensen, I want you to know that we invested in Electronic Arts, and their CTO is 14 years old and it's driven to work. And you're telling me that's your killer app?" And so, anyhow, we, we, uh, we created the modern 3D graphics, uh, gaming ecosystem, and as you know, it's one of the largest entertainment industries in the world. Um, the, the fundamental problem of 3D graphics is basically simulating reality. If you go back to first principles, what it's doing is trying to recreate reality. And the, the, uh, the fundamental, the mathematics of, of, um, uh, reproducing, uh, photorealistic images and dynamic worlds is, you know, fundamentally physics simulation. And so, uh, linear algebra is obviously very important to it. And, and, um, uh, we, we realized that concept. And so the question is, how do you bring something general purpose into something very, very specialized? And that's the great, that's the great invention of our company. We invented the technology, we invented the market, and we invented the pathways for us to systematically grow from a very vertically focused industry to eventually become more and more general purpose. And so that, that hardly ever happens. And that pathway was hard to do. But I don't want to take up the rest of the time explaining it. But I, I think the, the, um, CUDA's invention is part invention of the technology, which is observation, um, of how we can generalize our GPUs, uh, but it's a lot of it is about the invention of new products, how to take it to market, invention of new strategies, how to get the market to adopt, and, and invent, and inventing essentially ecosystems that ultimately create the flywheel that makes a computing platform happen. So we invented all of those things. They're all brand new. And, and if you go back, you take a, take a step back, and you ask yourself, aside from ARM and aside from x86, what is another computing platform that exists in the world that almost everybody uses? Doesn't exist. And so inventing a new computing platform rarely happens. And, uh, in our case, it took us almost 30 years.
So, you were able to take this very specialized, extremely high-performant acceleration device and generalize it so that researchers and academics around the world would be able to run their processing much faster. You know, the Moore's Law limitations that they were up against, all of a sudden were relaxed dramatically. Yeah. Now let's jump forward to the early 2010s. Yeah. At the time, deep learning was kind of an academic backwater. Uh, the idea of neural networks had gone through a winter phase. Then in 2012, there was a breakthrough with AlexNet and computer vision. And that was all accelerated on Nvidia GPUs. Was that the moment that you realized this AI revolution was becoming real? And if so, how did you capitalize on it? What was the edge to make Nvidia the center of this revolution?
Yeah. Two, two, uh, two serendipitous, two serendipitous moments, and then one which is just a great, again, first-principled observation about deep learning. Um, the serendipity, um, started with, I was trying to solve computer vision. And, um, uh, we wanted to solve computer vision for a lot of different reasons, whether it, you know, well, anyways, we want to solve computer vision. And, um, uh, computer vision, computer vision was, was, uh, really brittle, really hard to generalize, um, a collection of a whole bunch of tricks. And I really, really hated, um, uh, how, how the industry was evolving and, and, and really quite frustrated with the progress. Uh, meanwhile, one of our major strategies of, of, um, uh, democratizing the architecture, uh, is to go to higher, to get scientists in higher education to use our, our platform, use CUDA. And, and so I started with seismic processing, molecular dynamics, particle physics, you know, quantum chemistry. I went, you know, I took, I took Nvidia CUDA everywhere. And actually, there was a, there was a strategy at the company called CUDA Everywhere, that that meant Jensen schlepping CUDA all over the world. And so I went to universities everywhere, and we meet with researchers. And, and, um, that initiative of getting CUDA into higher education and researchers everywhere caused, uh, some researchers to reach out to us in 2012, 2011. And, uh, Jeff Hinton was, uh, trying to solve, uh, computer vision, and Andrew Ng was trying to solve computer vision, and Yann LeCun was trying to solve computer vision, because there was this, there's a contest coming up called ImageNet that Fei-Fei is in charge of. And I was trying to solve computer vision. And so when you're naturally trying to solve a problem, and then all these interesting people are solving similar problems, they attract your attention. So that's serendipity. Um, the, the, the thing that, that was a great observation is that, that, um, uh, we could create a, a new type of solver for them that's called cuDNN, kind of like the sequel of in-storage computing. We invented cuDNN, which is, which is in-network computing, if you will. And that, that way of doing computation, this library called cuDNN, uh, made it possible for all of them to use CUDA successfully. Um, but the thing that, that was, I saw the same results as everybody else, and, you know, everybody saw the big jump in, in, uh, computer vision effectiveness. But, but where, where we took it further was we reasoned about, so, so this is so good at computer vision, and why? And what else could it be good at? And the ability for, for, um, deep neural networks, uh, to be extremely deep, meaning because each layer is trained independently of the others, and you could backpropagate, um, from a loss function all the way back to its input, uh, you could learn almost any function. And we came to the conclusion, this is a universal function approximator. And if we can then add to it state, which is, you know, CNN was was a, a, a kind of a, two-dimensional, uh, multi-dimensional, uh, pattern recognizer. And then RNNs gave you a state machine within it. And LSTM gives you an even better state machine. And then Transformers give you the ultimate state machine. And so, so the, the idea that that we would have a universal function approximator that can learn almost any function. Well, the question is, is what problem can it solve? Now you invert the pro, invert the question. And we came to the conclusion, most of the problems we wanted to solve, uh, could have a deep learning component to it. And so we decided, uh, you know, how would we reason about, uh, where deep learning could be 10 years from now, 20 years from now? We broke down the computation problem, and we came to the conclusion that every single chip, every single system, every software, every, every single layer of the computing stack could be reinvented. And that, that decision to go after it was probably, you know, one of the better decisions in history.
I, I was doing AI research at the time at Stanford, and the major constraint was always the compute. Yeah. You know, we had limited clusters in order to run these algorithms, and Nvidia came in and not only relaxed that compute but made it possible with with the CUDA infrastructure. Yeah. That is largely your history. You make more and more compute possible. In 2016, you very famously created the world's first AI factory, the DGX-1. Yeah. Uh, you actually hand-delivered it to Elon Musk at OpenAI. At...
Well, I built this brand new computer, and it doesn't look, it doesn't look like anything the world's ever seen before. It doesn't work like anything the world's ever seen before. And, and I remember announcing it at GTC, and literally the audience was just like this. Nobody knew what I was talking about. And that was a joke. And so, so with the same amount of applause, and so, so I announced this thing, everybody go, uh-huh. And literally on that GTC, I was on, I invited Elon to talk about self, the two of us were working on self-driving cars. And so, so he came on stage and he says, "Jensen, what's that computer?" I said, "DGX-1." "Let me..." And I built it for this reason. He goes, he goes, "I could use one." And I finally got a PO. And, and then he goes, he goes, "Uh, he goes, yeah, I have this nonprofit." I said, "Oh, oh, you know, when, when you build something brand new, the last thing you want to hear is your first customer is a nonprofit." And so, anyhow, anyhow, I, I delivered, I delivered, uh, I was the, the DoorDash computer guy. And I, I, uh, DoorDashed this this computer up to San Francisco. And, and, uh, the company was OpenAI. It is a very profitable nonprofit, or revenue-scale nonprofit. Oh, we've been working together for a long time. Every, every model has been built on Nvidia since. Yeah.
And, and this thing is huge physically. When Jensen's talking about a computer, we're talking about a massive device. Nvidia's GPUs. When they say our GPUs, people imagine little GPUs. Our GPU is one GPU is now rack-scale. It's two tons, uh, 120,000 watts, about $3 million. That's a GPU. We also sell smaller GPUs, the ones that Jeff Hinton used, that's like $1,000, $500 that plugs into your PC, and you could use it for video games or AI and things like that. Um, but we also have bigger GPUs. And then, and then a one gigawatt AI factory GPU is about, you know, $50 billion. So, so tell us about these AI factories because you might have the small one, might be the AI blender. Uh, but then you have the really big one, these AI factories that you went all in in 2016 and started to say the world is going to need AI factories. How did you get that edge, that conviction, and then you got to reason about it. Exactly. You got to reason about it.
So, so we built the first one, DGX-1. It was the most expensive computer the world's ever seen, $300,000 per node. And, and, uh, it wasn't that successful. Uh, and so I came to the conclusion, uh, we didn't make it big enough. And so we made a bigger one, and the second one became super successful. And, and now the question then becomes, uh, how, how large do you make it, and how, how hard do you drive computation? The reason why things are moving so fast is Nvidia's product cycles and the way we innovate, the way we design. We, we're not designing a chip, we're designing an entire infrastructure all at one time. We're the only company in the world today that you can give a building, some power, and a blank sheet of paper, and we can create everything within it. All of the networking, all the switches, all the CPUs, all the GPUs, you know, all of the technology within that entire factory, we can build. And we can, and, and it all runs the, the same software stack from Nvidia. And because we can integrate like that, we can also move extremely fast. So I could redesign the next year, then redesign the next year, and every single year, they're all software compatible. The benefit of software compatibility is velocity. The reason why the PC was able to move so fast was because they were all Windows compatible. And, and, uh, and therefore, by definition, if you're compliant with the stack, you could build chips as fast as you like. And so we're now building AI factories as fast as we like, at the limits of what's physically possible. And so, um, and because we're innovating at such incredible scale and we're co-designing, meaning we're changing algorithms, we're changing software, we're changing networking and CPUs and GPUs all at the same time, we break out of Moore's Law's limits, which is, as you know, slowing down. And so generationally, we introduce performance levels by about 10 times. I mean, it's an incredible level of performance that we give to the market every single, every single year. The reason why we do that is we believe that just around the horizon is a problem that is so large, you need, um, a larger computer, faster computer. On the one hand. On the other hand, when we increase performance at the same power, we're decreasing your cost. M. And so our, we're driving cost down incredibly fast, which allows customers to do bigger things, which allows them to generate more revenues, uh, from the same factory. And so Nvidia's, you know, adoption today is because we are both the highest performance, we're the highest scale, and so if you want giant systems, you could do so. And we're the lowest cost. Our performance is so high. You know, for example, if your, if your data center is one gigawatt, you're not going to get more than that. You're one gigawatt. And so if our per, per watt, our energy performance per unit of energy used is three times, your company can generate three times more revenues in that factory. That's why I call it a factory. It's not a data center. It's a factory. They're making money from it. And, and so these AI factories want to keep driving the scale up. They want to keep driving the revenues up. They want to keep driving the throughput up. And so that's the reason why we're innovating so fast, and it's, it's hard to keep up with us. And, and, um, well, it also explains why we're successful.
Jensen, you have shifted from a component to a whole platform. That's the AI factory concept. For an investor audience, can you break down what goes into the platform and then also start to talk about what's next for what the platform looks like?
Well, the, um, uh, you know, there are CPUs, GPUs, uh, network processors. There are three types of switches. There's a scale-up switch that turns one rack into a whole computer. We invented rack-scale computing. It's called scale-up. You scale it out by taking a whole lot of these racks and connecting them together. That switch and that networking has a bunch of software on it. Software on, on top of all this stuff. And then you take, you, you create one giant system, the size of this building. And this, this building would probably be about a hundred megawatts. A gigawatt is, is, um, a few thousand acres. And then you, um, connect all these data centers together with even networking so that all of the, all the data centers can think together. And so that's what we, that's what we built together. That's what we built today. Um, there are several reasons why infrastructure is being built so fast. And, and, um, there's some, some questions that, that, uh, uh, floating around about about the bubble and comparing it to the year 2000. And so, just, just to compare it, during the time of 2000, internet companies, don't, there were hospital.com, there was pets.com. Uh, most of the internet companies were not profitable. And the size of the whole internet industry was about 20, 30 billion dollars, if you recall. And, uh, today, uh, the, the first thing you need to observe is that AI isn't just about the brand new companies, OpenAI and Anthropic and others. AI is transforming the way that hyperscalers do work. Like, for example, search is now powered by AI. Recommender systems, how you see ads and news and, and, uh, stories are now movies generated by, uh, recommenders by AI. User-generated content. So basically Google's business, Amazon's business, Meta's business, hundreds of billions of dollars of revenues are all powered by AI. Now, even in the absence of OpenAI and Anthropic, this entire hyperscale industry is being powered by AI. And so the first thing to observe is that whole thing needs to go from classical CPUs with classical machine learning to now deep learning with AI. So that transition alone is hundreds of billions of dollars. Does that make sense?
Absolutely. And so, so that's one. The second thing is that we now have this new market. This new market is called, you know, AI. And this, it's got a new industry, and they produce AI. And so OpenAI is the, Anthropic, the XAIs, the, uh, Gemini, uh, from Google, of course. And Meta is going to be an AI maker. And so this entire layer of AI model makers is also building AI factories. And these AIs are going to power the next generation of new opportunities. And this is where the Harveys, the, uh, OpenEvidence, uh, um, the Cursors. I mean, you're right, you see all of these AI-native companies, and they're going to be connected to AI models, and they're going to, they're going to go after, for the very first time in history, an industry that never was addressable, and that's the labor industry. And it's digital labor, digital cogni, called agentic AI, is going to supplement and augment the enterprise market. So, for example, Nvidia already today, we use 100% of our software engineers, 100% of our chip designers. Every single engineer in today is augmented by Cursor. We use Cursor largely inside our company. And so, so we now have AIs for all of our engineers. Productivity gains, uh, the work that we do, and so much better. Uh, you also see that there's a new industry showing up. It's called physical AI. So you have enterprise AI, you got physical AI are, um, are augmenting labor. And so, for example, a robo-taxi is essentially a digital chauffeur, right? And, uh, we're now going to have AIs that are going to be embodying, um, going to embed into anything that moves. And so, in the case of a robo-taxi, it's a steering wheel and wheels. But you're going to pick an, pick, you know, uh, pick and pick and place arms. You're going to have one arm, two arms, you got, you know, three legs, all kinds of different embodiments. And so, these two industries represent about a hundred trillion dollars of the world's economy. And for the very first time, we have technology that's going to be able to augment that. And so, that's the reason why, you know, people are so excited about, about the next, next wave of AI.
Let's talk for a moment on the previous wave because you mentioned how AI has already been offering an ROI. And for the investor audience, I think the Meta example is a great case study because in Q4 2022, Apple basically removed attribution data from Meta. And you all saw hundreds of billions of dollars of market cap decline. And the Meta team said, "How are we going to fix this?" They fixed that with AI powered by Nvidia GPUs. That's right. Yeah. And they got their attribution back up to where it was. And that has recovered many hundreds of billions. It's over a trillion higher than it was at its low. And that is all ROI that was powered really by your GPUs.
What, what Meta was classically, not just Meta, but, but, um, is one of the most complicated systems, software systems, is called a recommender system. And there's a couple of basic technologies. One of them is called collaborative filtering, which is, which is based on what I'm doing, um, and looking at what everybody else is doing. If we have similar patterns, it would recommend maybe the same movie to me, the same next item in your grocery list, you know, a book to me, a video to me, so on so forth. And then the other thing is called content filtering, just based on who I am and my preferences, and, uh, based on on what that book actually is, uh, you might be able to recommend that book to me. And so that recommender system is the largest software ecosystem in the world. And that ecosystem is moving very significantly, very quickly to AI. And so you're going to need a, a mountain of GPUs. And those systems were made famous by the Netflix challenge a couple decades ago. Now, Netflix, their recommendations are all powered by AI. And Amazon, as you said, when you go and purchase something, a significant number is by a recommendation system by AI. Search to AI, moving search to AI, all of this is being powered now. TikTok to AI, right? Google Shorts AI. I mean, without it, and now, now, um, uh, all of the personalized ads are going to AI. Yeah. So, just the amount of AI is just incredible. And that has nothing to do, notice I've just described a whole bunch of classical use cases. Well, quantitative trading is going to move to AI. What used to be human-engineered feature extraction is going to move towards AI. And I think that's actually an area that Citadel Securities has pioneered for the past 20-some years. So that's the classical AI. Citadel, Citadel is a great customer. Thank you. Uh, so that is a classical example. And for the investor audience, talking about AI ROI, it's already there in the form of trillions of market cap. Uh, let's talk about what's next for. Spend, so 2025 estimates can be as high as $500 billion of AI investment in the ground. Where do we go from here? Does this become a multi-trillion dollar a year investment category?
Yeah. So the manufacturing, the foundry part of AI, if you will, is the model makers. They're kind of like, think of them like wafer makers. The applications of that, and, and one way of thinking about AI is, is the large language models. That's the operating system, if you will, of the modern computer. And, and you build applications on top of these AI models. Not just one AI model, but a system of AI models. Okay? Okay. And so applications have a, you know, it's going to have a collection of different AIs that it connects together. And, and so the question is, what's the application space on top? The, the most sensible way of thinking about the application space on top, aside from all, all of the, whatever applications we have are going to be improved by AI that we've been talking about. Uh, a simple metaphor is just digital humans. And so, uh, a digital software engineer, right? AI coding. Mhm. It, it's going to be a couple of trillion dollar market opportunity, probably. Um, you know, an AI digital nurse, AI accountant, AI, uh, lawyer, AI, right? So there's AI marketer. So that we call all of that agentic AI. And that technology is, is, uh, evolving very nicely. And so for the very first time, technology is no longer just a tool used by accountants, tools used by software engineers. We're going to become digital software engineers. And I wouldn't be surprised if you, you license some and you hire some. And so depending on the quality and depending on the deep expertise, uh, and so future workforces in enterprise will be a combination of, uh, humans and, and digital humans. And some of them will be OpenAI-based, and some of it would be, uh, Harvey-based or, you know, OpenEvidence or Cursor or you, Replit or you, Lovable, or some of it will be third-party, and some of them you'll homegrow. And so we homegrow a lot of our own AIs because we have a lot of proprietary knowledge and data that we want to protect. And, and we have, we have skills in developing those AIs. Over time, more and more people will be able to cultivate, uh, their own digital AIs because it'll just be easier, easier to do so. And so enterprise agentic AI, you know, obviously, uh, augmenting the labor force is trillions of dollars of opportunity. And what's unique about about AI also versus previous software is that AI needs to, uh, think. Meaning, you can't pre-compile it, put it into a binary, download it, and use it. It's got to process all the time. And the reason why it has to process is it has to take your context. It has to think about what you want it to do and then produce an output. And so it's thinking and thinking and generating. It needs a machine. It needs computers to do that. And that's the reason why AI factories exist. And so these AI factories will be in the cloud. They might be on-prem. They'll, they'll be all over the world. And, and I, you know, it's part of the, the AI infrastructure, if you will. But there's, there's going to be a whole lot of thinking, uh, to produce these, these, we call them tokens, but basically intelligence. And so, so that's the cognitive AI, the digital workforce, if you will. And then the second one is robotics. You know, for the very first time. So let me give you a thought experiment. You know, why robotics is so close. Um, as you know, as you know, you can now prompt an AI and it could generate, uh, you know, prompt, Jensen picking up a bottle, opening it, and taking a sip. Okay? And it would generate the video of me right, opening up a bottle, taking a sip. Well, if it can generate all that, why can't it maneuver a robot to do that? And so your thought experiment would suggest that, you know, that's probably very likely. Now, if you could design a digital chauffeur that could drive a car, why can't you have a robot, a physical robot drive a car? And so, so if a physical robot, if you can embody a physical robot to even drive a car, why can't you embody a pick and, you know, pick and place arm, or any type of robotics? And so, notice we have the ability to embody almost anything. We could pick up, pick up, uh, knives and forks, and it becomes an extension of our body, and somehow we articulate it. We could pick up a baseball bat and use it as an extension of our body. Um, and so we embody these physical extensions. Future, future AIs will be able to embody, you know, and manipulate a car, um, robotic arms, a human or robot, a surgical robot, um, you know, so on so forth. And so, so I think these two, these two markets are, are within reach of, um, uh, of AI. And then lastly, if I just give you one example, you know, whenever you see the observation of one thing, the rest of it is just engineering, right? And so, and so we now, we've now seen the evidence of one excellent thing, which is a robo, a digital, an AI software coder, which is the reason why we use it so much. If you can have an AI software coder, why can't you have that AI software coder also, uh, write software to be a marketing campaign, or write software to, you know, help you solve any accounting, you know, whatever, whatever you want to do. And so, so almost the, the existence of that says the rest of it is engineering. Mhm. And then, and then we now have robo-taxis. You know, it's an embodied robot that controls a steering wheel and, and wheels. And, um, why, if that exists, why can't you generalize that? And so the rest of it is just engineering. And so I, I think it's, that's a good way to reason from first principles, uh, how likely it is we're going to be able to have this technology proliferate across, uh, industries and society. And then the next thing that you have to reason about is, okay, so how do you scale this out? How do you deliver this intelligence to all of these different applications? Well, you need AI factories. And so, so let's talk a little more about robotics. Uh, you have an exceptional robotics team. One of your executives who runs robotics here today. Uh, in a previous conversation, you shared some insight about how robotics might play out. You know, is it going to be a single humanoid project? Is it going to be open-source projects? How are those open-source projects going to tie back? How do you think robotics will actually manifest in the physical world and on what timeline?
Um, well, robo-taxis are here now. Yep. And their, their ability to generalize from city to city to city is really, really getting fast. And, and the reason for that is because the same fundamental technology. We went through the same journey. And for all the, the, the quant, quantitative trading, the algorithmic trading, uh, people in the room, you went from human-engineered features, machine learning, to, uh, using more and more deep learning. And, um, uh, you know, embedding certain modalities and, uh, multi-modality models to, to now, uh, largely end-to-end. And the reason why, and, and it's multimodal, um, in this journey, we, we became more and more generalizable. And, um, the, the AI model that you use for a self-driving car and the AI model that you use for a human or robot is highly similar. It's just in two different embodiments. And the reason why I know that for sure is because I can drive a car and I can manipulate my body. It's the same intelligence. And so, and I could pick up a fork and knife and somehow I, you know, pretend like I'm a surgeon, you know, and doing surgery on a, on a steak, you know, and so, so you could notice it's the same AI in different embodiments. Mhm. And so that's, that's where AI is going. Robotics is going towards a general, more and more generalizable AIs that are that are multi-embodiment. It's multi-modality. It's multi-embodiment. And in order, in order to create this future, you need three things. You need the AI factory I was talking about, where you have to train the models. And you need a place where the AI could learn how to be an AI, um, without, without having to come into the world right away. So it could try trillions of different iterations inside a virtual world. Well, that virtual world resembles a video game. And so the AI is basically playing a game inside a virtual world like a video game character. And it obeys the laws of physics. And, uh, when it's done learning how to be a great video game player, because the sim-to-real gap is extremely low, because the simulator is really, really good. We call it Omniverse. That Omniverse computer, then the robot can come out of that virtual world, and this world becomes one more version of the virtual worlds it's played in. And it comes into the physical world. When it comes into the physical world, it needs a computer as well. So you need three computers. You need the AI computer, training computer. You need the simulation, the lab, the, the virtual world computer. And then you need a computer where the robot actually operates, the brain. And so Nvidia offers all three of those computers. And we work with just about every robotics company, self-driving car company, you know, robotics of different embodiments. And this is likely going to be one of the largest markets of all.
So Nvidia touches just about everything in technology now. And as you've said in the past, you start with zero billion dollar markets and have helped turn them into trillion dollar markets. Robotics is one of the next frontier markets. Are there any other next frontier markets that you're particularly excited about? You mentioned healthcare a moment ago. Is that one you're passionate about? Are there others that the investors in the room should be on the lookout for?
Well, the technology, the technology needed for healthcare is really complicated. Um, and we're making fast progress. If you can understand, uh, the meaning of words, sequences of characters, you might, you might, um, and, and you could understand the meaning of structures, like the virtual world. Okay? Like when you, when you look at the reason why we're able to generate video is because we understand the virtual world to generate an image, a representation of the virtual, of the world. And so if you can generate video, it must be because you understand the world. If you can generate, if you can understand worlds, is it possible that you understand proteins and chemicals that have structure? And the answer is yes. And, and so we're increasingly, uh, getting closer to, closer to understanding the meaning of proteins. AlphaFold and others, uh, we're able to understand the meaning of cells. And we recently, uh, uh, did a, a partnership with ARC, and EVO2 comes, very is one of the first, uh, examples of a large language model that a foundation model for cell representation. So you, you can now talk to it and say, um, "I want you to generate other cells of these properties." Or, uh, you could talk to a cell, you know, "What are, what are your, your, um, properties, and what can you bind to, and what, what can you, um, your metabolism, what can you activate with?" And so you could talk to a cell like you could talk to, talk to a chatbot. And so, understanding the meaning of, of, uh, proteins, um, you know, so anyways, there's a, a lot of progress there. I mean, the list goes on. I mean, the, um, I'm excited about, about, uh, uh, the work that we're doing to, uh, bring AI into telecommunications. 5G and 6G will be revolutionized by AI. Uh, I'm excited about, about, uh, the collaboration we have with, uh, uh, quantum computers, so that we can, we can pull in the quantum computer schedule by about a decade by creating quantum GPU hybrid computing systems, where we do the error correction, we control the quantum computer, we do the post-processing, and, uh, so, so we have a new, new architecture called CUDA-Q, which extends CUDA to quantum. And that, that's getting incredible adoption. And so there, yeah, there's, there's a whole bunch of problems we can now solve that were hard, hard to solve before.
Let's talk a little bit about sovereign AI. We just had Mario Draghi on the stage. He was talking about the importance of new investments in technology for the European Union, including obviously AI at a large scale. This revolution is materially different in that governments are highly involved, both in potentially regulating but also in purchasing AI factories. Can you tell us, what do you think is the way forward both for sovereign AI, how countries have their own AI systems, and also for import/exports, how we as the United States should be interfacing with the rest of the world with AI?
Well, um, no country can afford to outsource all of their, their nation's data so that and import your own intelligence back to yourself. And I, I just think on first principles, that's not sensible. And, um, however, no, nobody needs to, only build everything themselves. You could, you could buy, you could import, um, but you shouldn't give up on the production of your own national, uh, intelligence. And, and so I, I think the, today the technology is rather hard, um, but it's getting easier and easier very, very quickly. And there's an enormous amount of open-source capability. And so, so I would, I wouldn't give up on, on, um, building your own sovereign AI. I wouldn't give up on, on, um, uh, taking the data that you have and creating your own national intelligence from it. And, and, and now countries all over the world are, are, um, are doing so. And so I think sovereign AI is likely, every country is likely to, um, import some, um, buy some, and also build some. And, and, uh, there's a lot of capabilities to doing that. And, and so we're, we're seeing just a lot of momentum around sovereign AI. Uh, the UK is doing it, you know, I was, I was in France. We, uh, support a company called Mistral. In the UK, there's a company called N-Scale. Uh, there's a company called Nebbius. In, in, uh, uh, in Italy, there's a, there's several companies. In Spain, there's several companies. Germany, there's several companies. In, and so there's companies all over the world. In, in, um, in, in Japan, there's companies, you know, in Korea, there's companies. And so there, sovereign AI is cropping up all over the world. Yeah.
So, one country that's come up a lot is China. Uh, what's the right thing for the United States in terms of exports to China of AI factories?
Well, AI is a new technology, and we have to think about, before we, you know, we have to be thoughtful about, ultimately how to regulate it. Um, the United States, of course, wants, wants to win the AI race. And I think, I think, um, the policymakers, all want to do the right thing, and they want America to to win. Um, however, it's important to be mindful that what is what harms China, could often times also harm America, and even worse. Mhm. And so before, before we leap towards, um, policies that are hurtful to other people, uh, take a step back and, and, and maybe reflect on what are the policies that are helpful to America. And, um, it, it probably is the case, uh, that you have to go back to first principles again. In the case of AI, what's most important about AI, and any computing, any software industry, the developers are vitally important, as you know. And so winning developers is what creates the future platform. And we want the world to be built on American technology, you know, and, and Nvidia is a proud American company. And, and, um, we want, we want, of course, we, we hope that we could create American technology that the world's built upon. Well, a lot of the AI researchers are in China. You know, China has about 50% of the world's AI researchers. Incredible schools, incredible focus in AI, lots of passion around AI. And I think it's a mistake to not have those researchers build AI on American technology, on first principles. I think that's a mistake. And so the question is, how, how do you balance winning, staying ahead, on the other hand, ensuring that the world builds on American tech stack? That's the balance. And in order to balance, you have to have nuance. And it, it's probably not, you know, all or nothing. And so nuance, a nuance strategy that, that, um, uh, changes, that, that, um, uh, you know, is, uh, uh, changing over time, has a, you know, that allows the United States to stay ahead while we continue to win researchers around the world, is probably the right balance. And, and that, that's what I would advocate. At the moment, uh, we are 100% out of China. So China is 0% of, we went from 95% market share to 0%. And so I can't imagine any policymaker thinking that that's a good idea. That whatever policy we implemented, caused one of, caused America to lose one of the largest markets in the world to 0%. But anyhow, in all of our forecasts, if there are any shareholders out there, all of our forecasts, we're, we're assuming zero, zero for China. If anything happens in China, which I hope it will, it'll be a bonus. But it's a large market. Uh, China is the second largest computer market in the world. Uh, it is a vibrant ecosystem. I think it's a mistake for the United States to not participate. And, and, um, and so, hope, hopefully, we'll, we'll continue to, uh, to, uh, explain and, and inform and, and, uh, uh, hold out hope for, for, for a change in policy.
Jensen, you were at our offices recently, uh, for an AI conference we were holding, and you had some really brilliant insights into the future of AI security and the importance of it. It's somewhat related. Uh, there are nation-state actors that might interfere with AI. There are individual users that might use AI.
incorrectly. What do you think the future of AI security looks like?
Well, AI security in the future is going to look a little bit like cyber security. It's going, it's going to require that, that, uh, we all work as a community. Um, you probably know this, all of your cyber security, your, your chief security officers, uh, we're all one large community and when, when somebody finds a, a, um, you know, some intrusion, uh, we share it with everybody. Whenever we find a vulnerability, we share with everybody. And so, uh, it's very likely that the future of AI security will be like cyber security.
Second, if the marginal cost of intelligence, the marginal cost of AI goes to zero, if the marginal cost of AI goes to zero, then why wouldn't the marginal cost of security-focused AI go also to zero? And so that's, that's very clear. It's very likely that every AI will be surrounded by a whole bunch of cyber security AIs watching it. And so we're going to have lots and lots of AI protectors, thousands of them, millions of them inside the company, outside the company. And so that's, that's, uh, kind of the future of, uh, the idea that, that an AI has to itself be good is good, but I don't think we should rely on it. Mhm. And so just like the idea that a piece of software should be properly functioning, we, we like, but the idea that it could have bugs or, you know, it could be a virus or whatever it is, could have, it could be an intruder, we have to assume. And so we're going to, we're going to make AI advance as safely as possible. Um, but we're all going to also surround AI with a lot of security AIs.
You shared that really the dynamics of the physical world are decoupled in this digital world where in the physical world you might have one security person to a hundred normal people. It could be inverted in an AI world. Yeah. You also shared this idea that I found my, for example, like cyber security. Yeah. Yeah. We have a lot more cyber security agents than we have people working in the company on cyber security.
You also shared this idea that in the future we're not just going to have rendered computation, but everything is going to be generated. Can you unpack what that prediction is and what that means for Nvidia?
Well, the, the great, the best example of that, um, a couple of examples. Let's, uh, Perplexity. Everything that you see on Perplexity when you ask it a question is completely generated, 100%. 100% of everything you see is generated. And yet in the past, before Perplexity, you would type in something and it would give you an a list and you would go and click on it and all of that content was, um, written by somebody or gen-created by somebody a priori. So search is storage-based computing. It's retrieval-based computing. It's retrieving information for you to consume yourself. Perplexity or AI is generating. It goes and studies it, goes and reads all the content, and it generates it for you. Okay. So Perplexity is a great example of the classical computer approach. We go retrieve a file and read it to a generative approach, Perplexity, which is AI-based.
Another one, which is, um, look at the videos that we see today. You know, Sora is of course, uh, nano, banana, of course, you know, all of those pixels are generated. It's conditioned and prompted by you. You know, you might, you might give it an initial seed of something and say, uh, you know, I would like, I would like you to generate a video of Constantine and and Jensen having a fireside chat and and then you would to prompt it and say, this fireside chat's going to talk. They're going to talk about, you know, crazy stuff. For those online, this is real, actually. And and then, and then Sora would generate it. And so every single pixel, every single motion, every single word is gener-generated. So, so the, the, the way of computation in the future will likely be generative.
And let me just give you one final idea. A 100% of what you and I just went through is generated. Every question you asked me, I didn't run back to my office and retrieve something and bring it to you. Is this what you meant, Constantine? And then you read it aloud for everybody to hear. That's yesterday's computer. Today's computer is we just, we're just interacting. And so we are generating everything in real time based on the context that is happening right here, based on the audience, based on what's happening around the world. And so we're generating everything real-time. That's the future computer. You know, your future computer is is a CEO in front of you, or it's an artist, it's a, you know, it's a poet, it's a storyteller, and you collaborate with it to create unique content for yourself. And so the future of computation is 100% generative and behind it, you need an AI factory, which is the reason why I'm 100% certain we're at the beginning of this journey and and, um, you know, we're a few hundred billion dollars, a few extremely small. We're only a few hundred billion dollars of infrastructure built for what likely will be trillions of dollars of infrastructure built each year. And so that's the easiest way to think through it. And that computing paradigm is so much more like the human mind. Yeah. It's thinking, you know, it's thinking.
So if you're up for it, how about we generate a few lightning round answers? Okay. Okay. Okay. Uh, just in the last few minutes together. I'm sure fried chicken is the answer. I don't know what the question is for that one. Um, so let's jump in. Uh, what's one KPI that Wall Street underweights in the future of AI factories?
Your throughput per unit energy governs the revenues of your customers. It's not just about selecting a better chip. It's about deciding what your revenues are going to be. And in fact, if you go back and look at all the CSPs, the ones that chose right saw revenue growth and the ones that slow down subsequently chose right. And so you, you could, you could see it playing out and people are starting to understand it. Per your throughput token, it's called tokens, token generation rate per unit energy of your factory is your revenues.
The most underrated piece of Nvidia's platform? Um, most people talk about CUDA and CUDA is very important, but there's a suite of libraries that sit on top of CUDA. And I mentioned one earlier today, it's called cuDNN. And it is probably one of the most important libraries ever created in the history of humanity. Uh, the past, the previous one was called SQL, SQL, and this one, cuDNN, there's a few others, cuDF, litho, which is going to be used for semiconductor manufacturing lithography. We have about 350 of these libraries and these libraries, that is Nvidia's treasure trove.
What's one technology that you think is wildly undervalued and one that you think might be overvalued? Um, undervalued. Undervalued. Undervalued. Wow. Um, I, I think that that, um, a, the virtual world for physical AI to learn to be, to learn to be a good physical AI. We call it Omniverse. Is is hard to understand, but it is, it is deeply undervalued. And and not, not because people use it and don't, they just don't know they need it yet. And, uh, but now Omniverse is sweeping across the robotics industry and, uh, everybody now gets it. Um, once you start building robots, you'll start to realize, you know, how visionary it was that we started working on Omniverse almost a decade ago. And, um, and so Omniverse is is really important.
Uh, what's the book that most shaped your business and leadership philosophy? Uh, um, one of my favorite books was was, uh, the your, you know, everybody's first calculus book. That's when you realize that math was was, um, in motion. Um, that was a good book. Uh, all of, all of Clay's books, Christensen's books, and he's passed, but, but good friend. Uh, all of his books were great. Um, uh, Al, Al Ries's positioning book, really good book, if you haven't had a chance. Of course, you know, Sapiens is always good. Um, but, but those are good ones. You know, Geoffrey's book on Crossing the Chasm. That's a good book. But all of Christensen's books, read them all.
Favorite comfort food. There you go. Fried chicken. There you go. Okay, we got it in. All right. And then last question. If you were a CIO in the audience with $10 billion to allocate toward AI in the coming years, what would you invest it into?
I would right away, um, experiment with building your own, uh, AI. I mean, I just, you know, the fact of the matter is, we take pride in onboarding employees and how the method by which you do so, the culture by which you bring them into the philosophies of your company, uh, the operating methods, the practices that, that, that makes your company what it is. Um, the, the, um, the collection of data and knowledge that you've embodied over time that you make accessible to them. And so, so that, that is what defines a company in the past. A company of the future includes that, of course, but you need to do that for AI. You need to onboard digital, you need to onboard AI employees. There's methodology for onboarding AI employees for, uh, we call them fine-tuning, but basically teaching them, you know, the, the, the, uh, the, the, the culture of the knowledge of the skills of, um, evaluation methods and, and so the entire flywheel of your agentic employee is something that you need to go and learn how to do. Uh, I tell my CIO, our company's IT department, they're going to be the HR department of agentic AI in the future. They're going to be the HR department of of, uh, uh, digital employees of the future, and those digital employees are going to work with our, of course, biological ones, and, and that's going to be the shape of our company in the future. And so if you get a chance to do that, I would do that right away.
Thank you, Jensen. Well, we heard an incredible story. Really, the story of Nvidia is one of exceptional generalization. From an accelerated graphics processor to the technology that powers all of AI in the world today, from a component and the world's first GPU to all of the components in a platform and the world's AI factory. We talked about how services are the baseline for this new revolution and how robotics are in all of our future. We covered foreign policy. We even touched on fried chicken. You did it all, Jensen. Thank you so much. Thank you. Good job. Thank you. [Applause]