Jensen Huang recounts the "NVIDIA entrepreneurial history": insights from 1993, breakthroughs in 2012, and the future of AI

On October 6th, at the Citadel Securities Future of Global Markets 2025 conference held at Casa Cipriani in New York, NVIDIA founder and CEO Jensen Huang spoke with Sequoia Capital partner Konstantine Buhler about artificial intelligence and the next frontier of growth.

Huang stated that when NVIDIA was founded in 1993, it foresaw the limitations of general computing (CPU) and the end of Moore's Law, thus determining the strategic direction of "accelerated computing." NVIDIA simultaneously invented new technologies and the massive market of modern 3D gaming, resolving the dilemma of "which came first, the chicken or the egg."

Entering the AI era, NVIDIA promoted CUDA to the research community through the "CUDA Everywhere" strategy. From 2011 to 2012, in collaboration with researchers like Geoffrey Hinton and Andrew Ng, and by providing enabling technologies such as cuDNN, NVIDIA accelerated breakthroughs in competitions like ImageNet. Based on the insight that deep learning is a "universal function approximator," NVIDIA made a radical decision to reshape the computing stack, integrating AI into all chips, systems, and software, establishing its core position in the AI revolution.

In 2016, NVIDIA launched its first AI factory, DGX-1 (with OpenAI as its first customer), entering the realm of hyperscale computing. The core secret to its success lies in "full-stack collaborative design": simultaneously designing and integrating the entire infrastructure (network, CPU, GPU) and running a unified software stack. This high level of integration allowed it to break through the limitations of Moore's Law, achieving approximately 10 times the performance leap between generations, providing customers with extremely high energy efficiency and significantly increasing the revenue output of AI factories.

Huang rebutted the AI bubble theory, emphasizing that AI has already achieved hundreds of billions of dollars in actual ROI in hyperscale data centers (such as search and recommendation systems). He predicted that AI will create two trillion-dollar new markets:

1. Digital Workforce (Agentic AI): Creating "digital humans" such as AI software engineers and AI lawyers.

2. Physical AI (Robotics): General AI-driven multi "embodied" robots (such as autonomous vehicles and humanoid robots).

He concluded that the essence of future computing is 100% generative, with all content being intelligently generated in real-time. To support robots (which require training, simulation, and operational computing), AI factories are an indispensable infrastructure, and their market demand is at the beginning of a multi-trillion-dollar explosion.

The following is a summary of key points:

• "We believe that the scale of computational problems we can solve is almost infinite, and therefore, one day, a new type of computing method will emerge. Our company focuses on enhancing and complementing general computing with a technology called 'accelerated computing.' This was our initial observation."

• "I think the invention of CUDA is partly a technical invention, in that we observed how to generalize GPUs; but it is also largely about the invention of new products, how to bring them to market; the invention of new strategies, how to get the market to accept them; and ultimately inventing an ecosystem that can create a flywheel effect, leading to the birth of a computing platform."

• "We concluded that this is a universal function approximator. ... So the question becomes: what problems can it solve? Now you turn that question around, and we conclude that most of the problems we want to solve can include a deep learning component. So we decided to think about how deep learning will evolve in 10 or 20 years."

• "We are not designing a chip; we are designing an entire infrastructure at once. We are 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 the networks, switches, CPUs, GPUs, all the technology in that entire factory, we can build. And they all run the same software stack from NVIDIA."

• "This is why AI-native companies like Harvey, Open Evidence, and Cursor are emerging. They will connect to AI models and, for the first time in history, venture into an industry that has never been touched by technology—the labor industry."

• "Your future computer will be like a CEO in front of you, or an artist, a poet, a storyteller, with whom you collaborate to create unique content for yourself. So, future computing is 100% generative. Behind it needs to be an AI factory, which is why I am 100% sure we are at the beginning of this journey."

The following is the original interview text:

Konstantine Buhler:

Good morning, everyone. My name is Konstantine Buhler, and I am a partner at Sequoia Capital focused on AI investments. NVIDIA and Citadel Securities actually have a lot in common; they are both outstanding companies.

Jensen Huang: The operation is going very well, exceptionally.

Konstantine Buhler:

Yes, they are all operating exceptionally well, driven by the computing revolution, and they are all technology-leading leaders in their respective industries. They also have a lesser-known fact: the first external investors in both companies were Sequoia Capital.

Jensen Huang:

In 1993, they took the risk to invest $1 million in NVIDIA.

Konstantine Buhler:

You deserved that investment.

Jensen Huang:

A full $1 million, sir, they took a huge risk.

Konstantine Buhler:

So, when we were invited to talk about artificial intelligence at this conference, who is the most suitable person in the world? The answer is obvious. He built the entire infrastructure for the AI revolution, all AI is built upon it; he created the most valuable company in the world. Please join me in welcoming Mr. Jensen Huang.

Jensen Huang:

Welcome.

The Origin of NVIDIA: Insights from 1993

Konstantine Buhler:

Jensen, this room is filled with the best institutional investors in the world. They manage trillions of assets and are constantly looking for an edge. You are someone who always seems to have an edge. In every conversation we have, you have compelling insights about the future. In the next 60 minutes, we have a grand agenda: covering the story from NVIDIA's founding to its rise as the center of the AI revolution, and then we will spend most of our time discussing the future of NVIDIA and AI. Alright, let's start from the beginning. Let's go back to 1993 when you were 30 years old. What insights gave you the edge to found NVIDIA?

Jensen Huang:

At that time, we were experiencing the personal computer (PC) revolution and the central processing unit (CPU) revolution; it was the era of Moore's Law. The hot topics were integrated microprocessors, Intel, Moore's Law, and the scaling law of transistors, and almost all investments in Silicon Valley's computer industry were focused on this. However, we observed something different. We believed that one advantage of the CPU was its versatility, but the fundamental problem with general-purpose technology is that they often do not perform well when dealing with very difficult problems.

Therefore, we inferred two things: First, we observed that some problems could be solved with a more domain-specific, targeted accelerator, and these problems might be worth solving. Second, we observed that general-purpose technology, namely the continuous miniaturization of transistors, would eventually reach its limits. The idea that one could continuously shrink transistor sizes and scale up using the so-called "Dennard Scaling"… In fact, the fundamental principles behind Moore's Law were proposed by Mead and Conway. If you trace back these principles, you will find that there will be a limit to how much transistors can be miniaturized, and one day you will experience diminishing returns We believe that the scale of computational problems we can solve is almost limitless, and therefore, one day, a new type of computing method will emerge. Our company focuses on enhancing and complementing general computing with a technology called "accelerated computing." This is our initial observation. You just mentioned how NVIDIA is always one step ahead, which is often because if you reason from first principles, think about what works very well today, and ask yourself: what is our foundation based on first principles? How will that foundation change over time? This can give you insights into the future.

From Graphics Accelerators to CUDA: Market and Technological Invention

Konstantine Buhler:

So, when you manufactured graphics accelerators, you entered the market early, but then hundreds of competitors emerged. Ultimately, you prevailed in that market. In the early 2000s, you realized that this technology could perhaps be generalized. You just talked about the generality of CPUs; perhaps GPUs could also be generalized for more processing. Let's talk about CUDA. Where did this idea come from? Where did you gain this insight? There are rumors that it came from researchers; how did you conclude that GPUs could become general-purpose computers by reading their research?

Jensen Huang:

First of all, the difficulty in creating NVIDIA was that we had to invent a new technology while simultaneously creating a new market. In 1993, to create a new computing platform, you needed a huge market. At that time, the Silicon Graphics company, which was doing 3D graphics, had a market that was too small to support a new computing platform. Therefore, if we wanted to create a new computing architecture, we needed a huge market, but that market did not yet exist because the architecture did not exist, which led to the "chicken or egg" problem. What NVIDIA excels at, and the significant contribution we made to the modern 3D gaming market, lies here.

At that time, Sequoia Capital's main concern about investing in NVIDIA was that we had to invent both the technology and the market simultaneously, and the probability of both happening at the same time was about 0%. I still remember when I pitched to Don Valentine, he asked, "Jensen, where's your killer app?" I said, "There's a company called Electronic Arts." I didn't know at the time that Don had just invested in Electronic Arts. I continued, "We will help them create 3D graphics games and create this market." He replied, "Jensen, I want you to know that we invested in Electronic Arts, and their CTO is only 14 years old and is driven to work every day. You're telling me this is your killer app?" In summary, we ultimately created the modern 3D graphics gaming ecosystem, which has now become one of the largest entertainment industries in the world.

The fundamental problem of 3D graphics is simulating reality. If you go back to first principles, what it does is attempt to recreate reality. The mathematical foundation for reproducing realistic images and dynamic worlds is essentially physical simulation Therefore, linear algebra is obviously crucial, and we have recognized this.

The question is, how do we introduce something general into a very specialized field? This is precisely the great invention of our company. We invented technology, created markets, and paved the way for us to systematically grow from a very vertical industry into an increasingly general platform. Such situations are rare. This path is difficult to walk, but I don't want to take up the remaining time explaining it. I believe the invention of CUDA is, in part, a technical invention, meaning we observed how to generalize GPUs; but it is also largely about the invention of new products, how to bring them to market; the invention of new strategies, how to get the market to accept them; and ultimately inventing an ecosystem that can create a flywheel effect, thus facilitating the birth of a computing platform.

We invented all of these things, and they are all new. If you take a step back and ask yourself, besides ARM and x86, what other computing platform is almost universally used in the world? The answer is none. Therefore, inventing a new computing platform is extremely rare. For us, it took nearly 30 years.

Starting Point of the AI Revolution: The 2012 ImageNet Breakthrough

Konstantine Buhler:

So you successfully generalized this highly specialized, high-performance acceleration device, enabling researchers and scholars around the world to run their processing tasks faster. The limitations of Moore's Law they faced before were suddenly greatly relaxed. Now let's fast forward to the early 2010s. At that time, deep learning was still a niche area in academia, and the concept of neural networks went through a "winter period." Then in 2012, AlexNet achieved a breakthrough in the field of computer vision, and all of this was accelerated on NVIDIA's GPUs. Was that the moment you realized the AI revolution was becoming a reality? If so, how did you seize this opportunity? What was the key advantage that made NVIDIA the center of this revolution?

Jensen Huang:

There were two serendipitous moments and a great observation about deep learning based on first principles. I was trying to solve the problem of computer vision. There were many different reasons we wanted to solve the problem of computer vision. At that time, computer vision technology was very fragile, difficult to generalize, and was just a collection of tricks. I was quite frustrated with the way the industry was developing.

At the same time, one of our main strategies for popularizing architecture was to get scientists in higher education to use our platform—CUDA. I started promoting it in fields like seismic processing, molecular dynamics, particle physics, and quantum chemistry. I took NVIDIA CUDA around the world. The company indeed had a strategy called "CUDA Everywhere," meaning I (Jensen) was running around the world. I visited universities and met with researchers globally. This initiative to introduce CUDA into higher education and research prompted some researchers to contact us around 2011 and 2012 At that time, Geoffrey Hinton, Andrew Ng, and Yann LeCun were all trying to solve computer vision problems because a competition called ImageNet, led by Fei-Fei Li, was about to take place. I was also trying to solve computer vision problems. So when you naturally try to solve a problem, all these interesting people are also solving similar problems, and they attract your attention. That's serendipity.

And that great observation was that we could create a new type of solver for them called cuDNN. It's like the technology we invented for in-network computing, analogous to the subsequent development of in-storage computing. This computing method, the library called cuDNN, enabled all of them to successfully use CUDA.

I saw the same results as everyone else; everyone saw a tremendous leap in computer vision efficiency. But we went further to think: why is this technology so outstanding in computer vision? In what other areas can it excel? The reason deep neural networks can be very "deep" is that each layer is trained independently of the others and can backpropagate from a loss function all the way to its inputs. You can use it to learn almost any function.

We concluded that this is a universal function approximator. If we add another state—convolutional neural networks (CNNs) are a two-dimensional, multi-dimensional pattern recognizer, recurrent neural networks (RNNs) provide a state machine, long short-term memory networks (LSTMs) provide a better state machine, and Transformers provide the ultimate state machine. Thus, we have a universal function approximator that can learn almost any function. The question then becomes: what problems can it solve? Now you turn that question around, and we conclude that most of the problems we want to solve can include a deep learning component. So we decided to think about what deep learning would evolve into in 10 or 20 years.

We broke down the computing problem and concluded that every chip, every system, every piece of software, and every layer of the computing stack could be completely reshaped. The decision we made to implement this might be one of the best decisions in history.

AI Factory: From DGX-1 to Full-Stack Platform

Konstantine Buhler:

At that time, I was doing artificial intelligence research at Stanford University, and the main limitation was always computing power. We only had a limited cluster to run these algorithms. The emergence of NVIDIA not only alleviated the limitations of computing power but also made it possible through the CUDA infrastructure. This is largely your history: making more and more computing possible. In 2016, you famously created the world's first artificial intelligence factory—DGX-1 You personally delivered it to Elon Musk at OpenAI.

Jensen Huang:

I created this brand new computer, which looks and works in ways never seen before. I remember announcing it at GTC, and the audience had no idea what I was talking about; the applause was as sparse as if I were telling a joke.

At that GTC, I invited Elon to talk about the autonomous vehicles we were both working on. When he got on stage, he said, "Jensen, what is that computer?" I said, "It's the DGX-1, built for this purpose." He said, "I could use one." I thought, oh, finally got a purchase order (PO). Then he said, "I have a nonprofit organization..."

When you create an entirely new product, the last thing you want to hear is that your first customer is a nonprofit organization. But anyway, I delivered it. I felt like a "Doordash" guy delivering computers, "delivering" this computer to San Francisco, and that company was OpenAI.

Konstantine Buhler:

It is now a very profitable nonprofit organization.

Jensen Huang:

We have been collaborating for a long time. Every model they have built since then has been on NVIDIA's platform.

Konstantine Buhler:

And the physical entity of this thing is enormous. When Jensen talks about a computer, we are talking about a massive device.

Jensen Huang:

When people hear about our GPUs, they might imagine a small GPU. But one of our GPUs is now rack-sized, weighing two tons, with a power of 120,000 watts, worth about $3 million. That is a GPU. Of course, we also sell smaller GPUs, like the ones Geoffrey Hinton uses, which cost around $500 to $1,000 and can be plugged into your PC for gaming or AI. But we also have larger GPUs; a 1 megawatt AI factory GPU is worth about $50 billion.

Konstantine Buhler:

Tell us about these AI factories. You might have small "AI mixers," but you also have these truly massive "AI factories." You went all in on this in 2016 and asserted that the world would need AI factories. How did you gain this insight and belief?

Jensen Huang:

You just need to reason it out. We built the first DGX-1, which was the most expensive computer in the world at the time, with each node costing $300,000, but it wasn't successful. So I concluded that we didn't make it big enough. Then we built a larger one, and the second one became super successful.

The question then became, how big should you make it, and how high should you push the limits of computing power? The reason things are developing so rapidly is due to NVIDIA's product cycle and our innovation and design approach. We are not designing a chip; we are designing an entire infrastructure all at once We are 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 the networks, switches, CPUs, GPUs, all the technology in that entire factory, we can build. And they all run on the same software stack from NVIDIA. Because we can achieve such integration, we can also move forward at an incredibly fast pace.

So I can redesign next year's products, then redesign the year after that, and every product released each year maintains software compatibility. The benefit of software compatibility is speed. The reason PCs have developed so quickly is that they are all compatible with Windows. Therefore, as long as you adhere to the standards of this technology stack, you can quickly manufacture chips at will. So we are now building AI factories at the fastest speed physically possible, at will.

Because we are innovating and co-designing at an incredible scale—changing algorithms, software, networks, CPUs, and GPUs simultaneously—we are breaking through the limitations of Moore's Law, which is slowing down. Therefore, each generation of our products increases performance levels by about ten times. This is the incredible level of performance we bring to the market every year. The reason we do this is that we believe there is always a problem so large on the horizon that you need a bigger, faster computer.

On the other hand, when we increase performance at the same power consumption, we are actually lowering your costs. We are rapidly reducing costs, allowing customers to do bigger things and enabling them to generate more revenue with the same factory. NVIDIA can be widely adopted today because we have both the highest performance and the largest scale. So if you want giant systems, you can achieve that. At the same time, we are also the lowest cost because our performance is very high.

For example, if your data center is 1 megawatt, you cannot get more power than that. If our performance per watt, that is, performance per unit of energy consumed, is three times that of others, your company can generate three times the revenue from that factory. This is why I refer to them as "factories" rather than "data centers"; they are making money with it. These AI factories want to continuously scale up, continuously increase revenue, and continuously improve throughput. This is why we innovate so quickly. It is difficult to keep up with us. This also explains why we are successful.

AI Investment and ROI: Already Evident

Konstantine Buhler:

Jensen, you have transformed from a component supplier to a complete platform provider, which is what investors refer to as the "AI factory" concept. Can you elaborate on what this platform includes? Additionally, what will the next development of this platform look like?

Jensen Huang:

The platform includes CPUs, GPUs, and network processors. There are three types of switches: one is a scale-up switch, which can turn a rack into a complete computer; we pioneered rack-level computing, known as scale-up. Scale-out is achieved by connecting many such racks together A large amount of software runs on these switches and network devices, with the software located above all this hardware. Then, you can create a giant system the size of this building, which has a power consumption of about 100 megawatts. A gigawatt data center would require thousands of acres of land. Next, you connect all these data centers through a broader network, allowing them to "think" together. This is what we are building today.

There are several reasons for the rapid infrastructure development, and there are now some questions about bubbles and comparisons to the year 2000. Back in 2000, internet companies like hospital.com and pets.com were mostly unprofitable, and the entire internet industry was valued at around $20 to $30 billion.

Today, the first thing to recognize is that AI is not just about emerging companies like OpenAI and Anthropic. AI is changing the way hyperscalers operate. For example, search is now driven by AI; recommendation systems determine the ads, news, and stories you see, and movies are also recommended by AI; user-generated content is the same. Thus, the businesses of Google, Amazon, and Meta—these hundreds of billions of dollars in revenue—are all driven by AI. Even without OpenAI and Anthropic, the entire hyperscale data center industry is already driven by AI. Therefore, it is essential to recognize that the entire industry needs to shift from using traditional CPUs with classical machine learning to using deep learning with AI. This transformation alone is worth hundreds of billions of dollars.

Secondly, we now have a new market called "AI," which has spawned a new industry for producing AI. Therefore, OpenAI, Anthropic, xAI, Google's Gemini, and of course Meta, will all become producers of AI. This entire tier of AI model manufacturers is also building AI factories. These AIs will power the next generation of new opportunities.

This is why AI-native companies like Harvey, Open Evidence, and Cursor are emerging. They will connect to AI models and, for the first time in history, venture into an industry that has never been touched by technology—the labor industry.

Digital labor, known as "Agentic AI," will complement and enhance the enterprise market. For example, at NVIDIA, 100% of our software engineers and chip designers are assisted by AI. Today, every engineer is enhanced through Cursor, and we use Cursor extensively within the company. We equip all engineers with AI, productivity has improved, and the quality of our work has become much better.

You will also see another emerging industry called "Physical AI." Thus, we have enterprise AI and physical AI to enhance the workforce. For example, autonomous taxis are essentially digital drivers. In the future, we will have AI that can be embedded in any mobile object. In the case of autonomous taxis, the AI is embedded in the steering wheel and wheels; But in the future, you will also see robotic arms for picking and placing, possibly with one arm, two arms, or even three legs—various different physical forms. The market sizes corresponding to these two industries (enterprise AI and physical AI) account for a significant portion of the global $100 trillion economy. For the first time, we have technology that can enhance this part of the economy. That’s why people are so excited about the next wave of AI.

Konstantine Buhler:

Let’s talk about the previous wave, as you mentioned that AI has already brought investment returns. For investors, Meta is a great case. In the fourth quarter of 2022, Apple removed Meta's attribution data, leading to a market value evaporation of hundreds of billions of dollars. The Meta team was then thinking, “How do we solve this problem?” They addressed this issue with AI powered by NVIDIA GPUs and restored their attribution capabilities to previous levels. This saved them hundreds of billions in market value, bringing their total market value to over a trillion dollars higher than its low point. This investment return was entirely driven by your GPUs.

Jensen Huang:

Recommendation systems are among the most complex software systems of the past, and not just used by Meta. It has several foundational technologies: one is collaborative filtering, which is based on my behavior and observes the behavior of everyone else; if we have similar patterns, it will recommend the same movie, the next item on your shopping list, a book, or a video to me. Another is content filtering, which recommends based solely on my identity, preferences, and the specific content of that book. Recommendation systems are the largest software ecosystem in the world, and this ecosystem is rapidly shifting towards AI. Therefore, you will need a lot of GPUs.

Konstantine Buhler:

These systems became famous due to the “Netflix Prize” decades ago. Now, Netflix's recommendations are entirely driven by AI. As you mentioned, at Amazon, a significant proportion of purchasing behavior is driven by recommendation systems.

Jensen Huang:

From search to AI.

Konstantine Buhler:

Shifting from search to AI. Now all of this—like TikTok—is driven by AI.

Jensen Huang:

Yes, shifting to AI. Google Shorts is entirely AI-driven. Now, all personalized advertising will also shift to AI. So, the number of applications for AI is truly incredible. Note that what I just described are traditional use cases. Quantitative trading will shift to AI, and feature extraction that was previously manually designed will also shift to AI.

Konstantine Buhler:

That’s exactly the area Citadel Securities has been pioneering for the past two decades. This is traditional AI.

Jensen Huang:

Citadel has always been a very good customer. Thank you Konstantine Buhler:

That's a classic example. For the investors present, discussing the return on investment in AI, it has already manifested in the form of a trillion-dollar market capitalization. Next, let's talk about future spending. By 2025, total investment in the AI sector is expected to reach $500 billion. Where will the future lead? Will this field become an investment category worth trillions of dollars annually?

Jensen Huang:

Yes. You could say that the manufacturing part of AI, which is the "foundry," consists of model makers. They are like chip manufacturers. One way to understand AI is that large language models are the operating systems of modern computers. You can build applications on top of these AI models, and not just based on a single AI model, but on a system composed of multiple AI models. Therefore, an application will connect and utilize a set of different AIs.

So, what is the upper application space? Aside from all the existing applications we have been discussing that are improved by AI, a reasonable metaphor is "digital humans." For example, digital software engineers (i.e., AI programming), which could represent a multi-trillion-dollar market opportunity. There are also AI digital nurses, AI accountants, AI lawyers, AI marketers. We collectively refer to all of these as "Agentic AI." This technology is developing in a very positive direction.

Thus, technology will no longer merely be a tool used by accountants or software engineers. We will create digital software engineers. I wouldn't be surprised if in the future you authorize the use of some digital humans and also hire some digital humans. It depends on their quality and professional depth. Therefore, the future workforce of enterprises will be a combination of humans and digital humans. Some of these will be based on OpenAI, some may be based on third parties like Harvey, Cursor, or Replit, and some will be cultivated internally by you. We have cultivated many of our own AIs internally because we have a lot of proprietary knowledge and data to protect, and we have the skills to develop these AIs. Over time, more and more people will be able to cultivate their own digital AIs as it becomes easier to do so. Therefore, enterprises and Agentic AI, by enhancing the workforce, bring trillions of dollars in opportunities.

What sets AI apart from previous software is that it needs to continuously process information. You can't pre-compile it like before, put it into a binary file, download it, and then use it. It must always be in a processing state. The reason for this is that it needs to acquire your context, think about what you want it to do, and then generate an output. So it is constantly thinking and generating. This requires machines, requires computers to accomplish. This is why "AI factories" exist. These AI factories will be deployed in the cloud and may also be on-premises, spread across the globe. This can be seen as part of AI infrastructure, where there will be a large amount of "thinking" to produce what we call "tokens," but its essence is intelligence. This is what is known as cognitive AI, which is the digital workforce The second point is robotics technology. Let me give you a thought experiment to explain why robotics technology is so close to us. You can now give AI a prompt, such as "have Jensen Huang pick up a bottle, open it, and take a sip," and it can generate a video of me performing that action. So, if it can generate all of this, why can't it manipulate a robot to perform the same action? This thought experiment suggests that this seems very possible now. If you can design a digital driver that can drive a car, why can't you have a physical robot drive a car? If a physical robot can be given the ability to drive a car, why can't it be given the ability to operate a pick-and-place robotic arm or any type of robot?

We humans have the ability to "embody" almost anything. We can pick up a fork and knife, and they become extensions of our bodies; we can pick up a baseball bat and use it as an extension of our bodies. We are able to "embody" these physical extensions. Future AI will be able to "embody" and manipulate a car, robotic arms, humanoid robots, surgical robots, and so on. Therefore, I believe that both of these markets (digital humans and robots) are within the capabilities of AI.

Finally, let me give an example. When you observe the possibility of something being realized, what remains is just an engineering problem. We have already seen a great example of this, which is AI software coders, and that is why we use it extensively. Since you have AI software coders, why can't it also write software for a marketing campaign, or write software to help you solve any accounting problem, or anything else you want to do? So, the existence of this example itself illustrates that extending it to other fields is just an engineering problem.

Similarly, we now have autonomous taxis. They are "embodied" robots controlling the steering wheel and wheels. Since they exist, why can't they be scaled up? What remains is just an engineering problem. So, this is a good way to infer the potential for this technology to be adopted across various industries and society from first principles. What you need to think about next is how to scale it up? How to deliver this intelligence to all these different applications? The answer is that you need an "AI factory."

Future AI Opportunities: Agents and Physical AI (Trillion-Dollar Market)

Konstantine Buhler:

Let's talk more about robotics technology. You have an excellent robotics team, and an executive responsible for your robotics business is also here today. In a previous conversation, you shared some insights about the future development of robotics technology. Will it be a single humanoid robot project? Will it be multiple open-source projects? How will these open-source projects integrate? How do you think robotics technology will truly manifest in the physical world, and what is the timeline?

Jensen Huang:

Robotic taxis have already emerged, and their ability to generalize across different cities is rapidly improving. The reason is that the underlying technology is the same, and we have all gone through the same development process. For those of you engaged in quantitative trading and algorithmic trading, you have also experienced the transition from manually designed features and machine learning to increasingly using deep learning, embedding specific modalities, and multimodal models Now we have basically achieved an end-to-end process. Moreover, it is multimodal.

In this process, the model's generalization ability is becoming stronger and stronger. The AI models used for autonomous vehicles and those used for humanoid robots or general robots are highly similar, only manifested in two different "embodiments." I can be sure of this because I can drive a car and control my own body with the same intelligence. I can use a knife and fork, pretending to be a surgeon, performing surgery on a steak. You will find that this is the same AI manifested in different "embodiments."

This is the future direction of AI development. Robotics is moving towards increasingly general AI, which are multimodal and multi-"embodied." To create such a future, three things are needed. First, the AI factory I mentioned earlier for training models. Second, a place where AI can learn before entering the real world, allowing it to iterate trillions of times in a virtual world. This virtual world is like a video game, where AI plays a game character and follows the laws of physics. Once it learns how to be an excellent player—thanks to our outstanding simulator, the gap between simulation and reality is minimal—we call it the Omniverse, which is the Omniverse computer. Then, the robot can step out of the virtual world, and the physical world becomes another version of the virtual world it has played in. When it enters the physical world, it also needs a computer.

So, you need three computers: one for training AI, one for simulating the virtual world (i.e., the laboratory), and one as the robot's brain, operating in the physical world. NVIDIA provides all three types of computers, and we collaborate with almost all robotics companies, autonomous driving companies, and various companies with different "embodied" forms. This is likely to become one of the largest markets in history.

Konstantine Buhler:

So, NVIDIA is now touching almost every aspect of technology. As you have said in the past, you start from scratch in a market and help it grow into a trillion-dollar market. Robotics is one of the next frontier markets. What other frontier markets are you particularly excited about? You just mentioned healthcare; is that a field you are passionate about? Are there other areas worth the attention of the investors present?

Jensen Huang:

The technology required for healthcare is very complex, but we are making rapid progress. If you can understand the meaning of words and sequences of characters, you might also understand the meaning of something structured like a virtual world. The reason we can generate videos is that we understand the real world, allowing us to generate its visual representation. So, if you can generate videos, it must be because you understand the world. If you can understand the world, is it possible to understand proteins and chemicals that are similarly structured? The answer is yes We are getting closer to understanding the significance of proteins, thanks to technologies like AlphaFold. We are also able to understand the significance of cells. Recently, we collaborated with ARC, and Evo-2 is one of the first large language foundation models used for cell characterization. Now you can tell it, "I want you to generate other cells with these characteristics." Or you can directly ask the cells, "What characteristics do you have? What can you bind with? What is your metabolism like? What can activate you?" You can converse with cells just like you would with a chatbot. Therefore, understanding the significance of proteins... In short, there has been a lot of progress in this area, with countless examples.

Additionally, I am excited about our work in bringing AI into the telecommunications field; 5G and 6G will undergo revolutionary changes due to AI. I am also excited about our collaboration with quantum computers; by creating quantum-GPU hybrid computing systems, we can advance the process of quantum computing by about a decade. In these systems, we are responsible for error correction, controlling quantum computers, and post-processing. We have launched a new architecture called CUDA-Q, which extends CUDA into the quantum realm and has gained widespread adoption. Now, we can solve many problems that were previously difficult to tackle.

AI Security and Generative Computing

Konstantine Buhler:

Jensen, you recently attended an AI conference at our office and shared some very insightful thoughts on the future of AI security and its importance. This is somewhat related to the previous topic. Some state actors may intervene in AI, and some individual users may misuse AI. What do you think the future of AI security will look like?

Jensen Huang:

The future of AI security will be somewhat similar to cybersecurity. It will require the entire community to work together. You may know that all cybersecurity professionals and chief security officers form a large community. When someone discovers an intrusion, we share it with everyone. When we find vulnerabilities, we also share them with everyone. Therefore, the future of AI security is likely to resemble cybersecurity.

Secondly, if the marginal cost of intelligence, that is, the marginal cost of AI approaches zero, then why wouldn't the marginal cost of AI focused on security also approach zero? This is very clear. It is likely that every AI will be surrounded by a large group of cybersecurity AIs monitoring it. We will have a multitude of AI protectors, thousands, millions, spread throughout and outside the company. That is the future landscape.

The idea that "AI itself must be good" is nice, but we shouldn't rely on it. Just like we hope software will run correctly, we must assume it may have vulnerabilities, viruses, or be compromised. We will push for the development of AI as safely as possible while also deploying a large number of security AIs around it.

Konstantine Buhler:

You have shared that the dynamics of the physical world are decoupled from the digital world. In the physical world, it might be one security personnel for every 100 ordinary people; whereas in the AI world, that ratio might be reversed Jensen Huang:

Yes, just like cybersecurity. The number of cybersecurity agents we have is far greater than the number of people in the company working on cybersecurity.

Konstantine Buhler:

You also shared a perspective that in the future, not only will we have rendering computing, everything will be generative. Can you elaborate on this prediction and what it means for NVIDIA?

Jensen Huang:

One of the best examples is Perplexity. When you ask a question on Perplexity, everything you see is completely generated, 100% of it. Before Perplexity, you would input something, and it would give you a list, and you would click on it. All that content was pre-written or created by someone. So, search is based on stored computing, it's retrieval computing, retrieving information for you to consume. Whereas Perplexity or AI is generative; it researches, reads everything, and then generates answers for you.

So, Perplexity is a great example of the transition from traditional computing methods (where we retrieve a document and read it) to generative methods (AI-based Perplexity). Another example is to look at the videos we see today, like Sora, nano banana, all those pixels are generated. It's up to you to set the conditions and prompts. You might give it an initial seed and then say, "I want you to generate a video of Constantine and Jensen having a fireside chat." Then you prompt it again, saying this time they will talk about something crazy.

Konstantine Buhler:

(By the way, to the online audience) This is real.

Jensen Huang:

Then Sora will generate it. So, every pixel, every action, every word is generated. The way of computing in the future is likely to be generative.

Let me give one last example. The entire exchange we just had was 100% generative. Every question you asked me, I didn't run back to the office to retrieve something to show you and then ask, "Constantine, is this what you wanted?" and then have you read it to everyone. That was computing of the past. Today's computing is us interacting directly like this. We are generating everything in real-time based on the context of the moment, based on the audience, based on what is happening in the world. This is the future of computing.

Your future computer will be like a CEO in front of you, or an artist, a poet, a storyteller, and you will collaborate with it to create unique content for yourself. So, future computing will be 100% generative. Behind it needs to be an AI factory, which is why I am 100% sure we are at the beginning of this journey. We have only built a few hundred billion dollars of infrastructure for this market, which may require trillions of dollars annually. This is the easiest way to understand its prospects Konstantine Buhler:

This calculation paradigm is more like human thinking.

Jensen Huang:

Yes, it is thinking.

NVIDIA's Investment Insights

Konstantine Buhler:

If you're ready, let's answer a few lightning round questions? In the last few minutes. I don't know what question corresponds to that answer, so let's just dive in. What is the key performance indicator (KPI) that Wall Street values the least?

Jensen Huang:

In the future AI factories, your throughput per unit of energy determines your customers' revenue. It's not just about choosing a better chip; it's about determining what your revenue will be. In fact, if you look back at all the cloud service providers (CSPs), those that made the right choices saw revenue growth, while those that were slow to act eventually made the right choices too. So you can see this happening, and people are starting to understand it. Your throughput—the token generation rate of the factory per unit of energy—is your revenue.

Konstantine Buhler:

What is the most underrated part of the NVIDIA platform?

Jensen Huang:

Most people talk about CUDA, and CUDA is very important. But above CUDA, there is a set of libraries. I mentioned one earlier today, called cuDNN. It may be one of the most important libraries ever created. The last one of equal importance was SQL, and this one is cuDNN. There are other ones, like cuDF, and cuLitho, which will be used for semiconductor manufacturing lithography technology. We have about 350 such libraries. These libraries are NVIDIA's treasure.

Konstantine Buhler:

Which technology do you think is severely underrated, and which one might be overrated?

Jensen Huang:

The underrated one, I believe, is the virtual world. We call it Omniverse, and it is a virtual world for physical AI to learn how to become a great physical AI. It's hard to understand, but it is severely underrated. It's not because people are using it or not; it's because they don't yet know they need it. But now, Omniverse is sweeping through the robotics industry, and everyone is starting to realize it. Once you start making robots, you will realize how visionary this is. We started developing Omniverse almost ten years ago. So, Omniverse is very important.

Konstantine Buhler:

Which book has had the greatest impact on your business and leadership philosophy?

Jensen Huang:

One of my favorite books is everyone's first calculus textbook, where you realize that math is an emotion. That was a good book. All of Clay Christensen's books are great; he has passed away but was a good friend. Al Ries's "Positioning" is a very good book If you haven't read it yet, Geoffrey Moore's "Crossing the Chasm" is also a good book. But basically, all of Clayton Christensen's books should be read.

Konstantine Buhler:

What is your favorite comfort food?

Jensen Huang:

Fried chicken.

Konstantine Buhler:

Alright, we got the answer. Last question: If you were a Chief Information Officer (CIO) on-site with a $10 billion budget for artificial intelligence over the next few years, what would you invest in?

Jensen Huang:

I would immediately try to build my own AI. We take pride in how we onboard employees—the methods we use, how we integrate them into the cultural philosophy of the company, the operational methods and practices that make the company what it is, and the long-term accumulated data and knowledge that we make accessible to them. These are the things that have defined a company in the past.

Future companies will certainly include these, but you need to do the same for AI. You need to onboard AI employees. We have a method for onboarding AI employees that we call "fine-tuning," which is essentially teaching them culture, knowledge, skills, and evaluation methods. Therefore, you need to learn how to build the entire flywheel for your "intelligent agent employees."

I tell my CIO that our company's IT department will become the human resources department for intelligent agent AI in the future, becoming the human resources department for digital employees of the future. These digital employees will work alongside our biological employees. This is the shape of our company in the future. So, if you have the opportunity to do this, I suggest you get started immediately.

Konstantine Buhler:

Thank you, Jensen. We heard an incredible story. The story of NVIDIA is an outstanding generalization story: from an accelerated graphics processor to the technology that drives all AI in the world today; from a component and the world's first GPU to all the components that make up the world's AI factory platform. We talked about how services are the cornerstone of this new revolution and how robotics will integrate into all our futures. We discussed foreign policy and even mentioned fried chicken. You covered everything, Jensen. Thank you very much.

Jensen Huang:

Thank you, well done