In an era where AI is reshaping the computer industry, what is the most essential ability for entrepreneurs? According to Jensen Huang, it is not coding, but rather systems thinking, resilience, and the courage to venture into the unknown.

Recently, during a public interview for early-stage entrepreneurs at Startup School, Huang sat down with YC CEO Garry Tan for nearly an hour. They discussed entrepreneurship, the transformation of the AI industry, corporate management, and personal growth, sharing his latest insights on the next generation of opportunities.

With the computer industry undergoing a complete reset, the NVIDIA CEO clearly stated that systems thinking will become a core skill for the future. He identified physical AI and robotics as NVIDIA's next hundred-billion-dollar growth market.

The market is closely watching where NVIDIA's growth potential lies after surpassing a trillion-dollar market cap. Huang provided a clear answer: from AI agents to physical AI, the reset of the tech industry has only just begun.

Physical AI and Robotics: The Next Hundred-Billion-Dollar Breakthrough

Beyond large language models, the capital market is most focused on the commercialisation timeline for embodied intelligence and robotics. Huang offered a highly counterintuitive view during the interview: "I would say the 'ChatGPT moment' for robotics happened a few years ago."

He recalled that when NVIDIA internally first generated a video of a finger moving and a hand picking up a glass, he realised a breakthrough in physical world AI was imminent. "If I can generate a video of a finger moving, why can't I make a robot do the same thing? At that moment, I knew flexible movement of robotic joints was about to be achieved."

Regarding future business guidance, Huang painted a picture of a vast physical AI business landscape for the first time. He pointed out that autonomous driving is currently the only robotics application with a sufficiently large market, relatively standardised technology, real economic value, and the ability to make a data flywheel spin.

"We are in Tesla's cars, in Mercedes' data centres and vehicles. We even open-sourced our self-driving stack because agriculture, mail delivery, and warehouse logistics all need it," Huang revealed a striking figure: "Our physical AI business, including autonomous driving, is currently approaching about $10 billion. It could become one of the largest industries in the world. It won't take two or three years, but it won't take ten years either. This will be our next hundred-billion-dollar business."

AI Agents: A New Software Form and the Challenge of Controllability

Discussing the evolution of the software ecosystem, Huang asserted that "Agents are the new software." He disclosed that NVIDIA is already using various AI agents internally on a large scale to accelerate R&D. "We let a thousand flowers bloom, allowing people to freely choose tools. Cloud code runs autonomously in NVIDIA's internal sandbox. It's fantastic."

However, he also pointed out the core bottleneck facing current agent technology: controllability.

"Whether using RAG or prompts, output control is still too coarse. The biggest breakthrough will be whether we can achieve extremely fine-grained control over agents."

Huang emphasised, "If we change just one word in a plan file, it should produce a specific difference—perhaps changing one pixel, one polygon, one component in a CAD file—and then regenerate everything else. That level of control and collaboration with agents will be a game-changer."

Will AI Take Away Jobs? "This Narrative is Exactly Backwards"

As AI capabilities leap forward, market concerns about "AI replacing humans" are growing. Huang not only denied this but also offered a unique economic perspective.

"The narrative about AI destroying jobs is exactly backwards. AI does eliminate specific 'tasks', but it absolutely does not eliminate 'jobs'."

He argued this logic with specific data: "The task of programming is being automated, but the number of software engineer positions is growing by 10% annually. The task of reading radiology scans is automated, but the number of radiologist jobs has increased by over 20% in recent years. Why? Because there is a massive backlog of demand."

Huang pointed out that due to an immense backlog of creativity, ambition, legal cases, and patients, when AI takes over mundane underlying work, companies need to hire more people to do more ambitious things. "Productivity gains drive growth, and growth drives more employment. That's the fundamental rule."

The Survival Rule of the Era: Systems Thinking and "Founder Mode"

Faced with the classic question of "what should young people learn now," Huang gave a very hardcore suggestion. He stated plainly that sitting in front of a computer and writing code—a "simple job"—will absolutely be automated, much like the "long division" older generations learned has long been obsolete.

"The underlying things will all be handled by agents in the future. Therefore, you must possess the ability for abstract 'systems thinking'. Systems thinking will be the new programming of the future." Huang advised young people in this reset computer era to return to "hard sciences": "Physics, chemistry, biology, computer engineering, and interdisciplinary fields—these sciences that solve extremely difficult problems will never go out of style."

At the end of the interview, Huang shared his very personal management philosophy. He strongly advocated for "Founder Mode," a hot topic in Silicon Valley recently.

In the early days, because of choosing the wrong algorithm for 3D graphics, NVIDIA almost went bankrupt. He saved the company by buying three textbooks on OpenGL from a bookstore. This experience taught him that traditional management dogma is meaningless in the face of rapidly changing technology.

"You are building an F1 race car that you are going to drive. So you must build it in a way that you can drive it. You should make the car adapt to you."

When questioned about what happens when he leaves the company if he doesn't use traditional management techniques, Huang responded candidly: "I tell them, when I one day die at my job, they can just reshape the company for the next CEO. To achieve our mission, you should do whatever it takes to make the organisation adapt to you."

For current entrepreneurs, Huang issued a powerful call: "This is the best time to start a company in the last 60 years. The entire computer industry has been completely reset. As long as you always keep in mind 'How hard can it be?', and let the pain come bit by bit, you will eventually reach your own NVIDIA moment."

Welcome to Startup School 2026. Now, let's begin.

Please join me in welcoming NVIDIA's founder and CEO, Jensen Huang.

Garry Tan: This is a surreal moment for me. Thank you. Thank you for being here, Jensen.

Jensen Huang: I'm happy to do it. It's great to be here. [Cheers] Obviously, if you're here, you'll be successful.

Garry Tan: So, I'm glad I'm here. [Laughter] Oh, Jensen.

NVIDIA's Wrong Algorithm

For students who only see NVIDIA as an AI company, what part of the early NVIDIA story do they most need to know? But most people wouldn't believe that the technology we chose when we founded the company was absolutely wrong. Our initial idea was to reinvent 3D graphics. The company's concept and perspective were that general-purpose computers, CPUs, are very useful, but if we could augment them with accelerators, we could solve problems that were otherwise too difficult.

One of the initial problems we chose was 3D graphics. At that time, in 1993, the personal computer was just being announced. Our great idea was to turn every PC into a game console because we grew up in the era of game consoles. So we thought, what if we could design a system that could fit into a PC and turn it into a game console? We wanted to reinvent the algorithm that required those large supercomputers and cram it into a PC. We came up with some new algorithms, felt excited about them, believed in them, reasoned through them in a thoughtful way, and then founded the company to build it.

It turned out that the algorithm was completely wrong, and the technology the company was founded on was also completely wrong. So in 1995, we realised this, and it was almost too late because there were about 35 to 40 other companies building 3D graphics for PCs. We realised it wasn't working. I came back to the company, and we were all there. I said, "What are we going to do? This isn't working." We all discussed it.

I said, "Look, if we don't face the fact that this isn't working, we will lose the company."

The Real Great Idea: Buying Books to Save the Company

Then we started working hard to find the right algorithm. Someone told me that it turned out none of us knew the correct method. Not only had we chosen the wrong technology, but we also didn't know how to do it correctly. So that was a significant day for me. I had only about $60 in my pocket. So I went to Fry's and bought three textbooks. These textbooks were about OpenGL and how to design the OpenGL pipeline. I brought them back to the company, gave them to the engineers, and we reinvented computer graphics. We are the world leader in modern computer graphics. We have invented most of the major breakthroughs over the past 25 years.

Everyone thinks NVIDIA started as the world leader in 3D graphics, but we learned it from a textbook. So we actually founded the company, raised money, and then bought the textbooks, think about it. For me, the biggest lesson is that technology is constantly changing. As long as you can face reality and have the ability to learn, the technology itself doesn't really matter.

Since then, NVIDIA has been inventing various technologies we had never really done before, approaching everything with the same attitude. If something is important, we need to learn it. How hard can it be? It always turns out to be much harder than we expected. But you go in with the attitude of "How hard can it be?" I mean, in the background, we are discussing… We were talking with some top YC companies, and you said each one has expertise in a certain area. You and NVIDIA also have an expertise, and they all just… I forgot how you put it. Like some kind of algorithm domain. So it sounds like 3D graphics was just the first algorithm domain. And it came from a textbook. But you know, anyone can read that textbook. You created particle physics, fluid dynamics. Yes.

But you created something people wanted, like a final product people are willing to pay a lot of money for. The company's completely correct great idea was that it's possible to solve otherwise too difficult problems by augmenting the CPU. Molecular dynamics is one of them. Image processing is one of them. Inverse physics is another. So there are various different algorithms. Of course, deep learning is one of the major ones. To create the company we have today, we realised early on that the key wasn't making a great chip, but accelerating an algorithm domain. So one thing I've always believed is that what makes a great company is a unique perspective on the world, a perspective you deeply believe in.

It's not just about the technology or the market, although those are important. If you have the right technology for the right market at the right time, your life will be much easier. Having a high-level vision for something important, a unique perspective on it that you firmly believe in, and pursuing that vision is difficult. Those are a good combination. In our case, we realised that accelerated computing would be important, and it turned out to be very important. Our understanding was that it's all about algorithms, not chips, and that proved to be completely correct.

The Sega Story

Garry Tan: So, you've talked a lot about the hardships of founders. Are there a couple of stories that stand out to you? I mean, the people in this room really want to start a company, but do they know they are truly ready to face difficulties, possibly having to shut down the company, things going wrong? What were some key moments that particularly impressed you? I think you were in Japan, right? You seemed to be commemorating… was it Sega? I feel like that was a very powerful story.

Jensen Huang: The project that made us realise the algorithm we had chosen was wrong was the collaboration with Sega. Sega had commissioned us to develop for the console after Saturn, which later became Dreamcast. I don't know if anyone knows what Dreamcast is? So, we didn't build Dreamcast. We were supposed to build Dreamcast, but because our algorithm and technology had fundamental flaws, I went to Japan and told the then-CEO that we would be unable to fulfil the contract, which was about $12 million. I explained why and suggested they find someone else. But then I asked him, unfortunately, I still needed the money. He asked, you can imagine that conversation. You are telling me you can't do what I contracted you to do, but you want all the money from the contract. I said, you're right, absolutely correct. But obviously, I was very polite. I was humble. He realised I was honest, and everything was reasonable.

If he didn't give us the money, we would go bankrupt. I think everyone here will encounter this situation. You are not investing in a company; you are investing in people. What Mr. Iriye recognised was that here was a person and a company he initially trusted, and this contract, he believed in them and wanted to see them survive to the next day. That $5 million kept us alive and gave me enough time to figure out what to do.

The $300 Million IPO

Then I guess if they held on, they later sold for fifteen million, I heard. But they sold the moment we went public. When NVIDIA went public, our valuation was $300 million. $300 million in 1999. That was real money. I think the company's market cap is now over a trillion dollars, roughly. Over a trillion. Yes. It's crazy.

Garry Tan: So, you're the central figure, and we like to say you're the one in control. Before that, I don't think anyone could really foresee how important the GPU and the technology you built would be for the AI revolution. What did you see? I mean, did the accelerator put you in the right place at the right time, or were there definitely many things that led you to be in this position?

Seeing AlexNet Differently

Jensen Huang: Yes. I saw AlexNet like everyone else, but remember, my perspective on the world is always looking for algorithms. The algorithm could be Namd, Vasp, OpenGL, SQL, some domain-specific language, some algorithm. So my worldview is always looking for problems we might be able to help solve. When AlexNet came along, that algorithm was deep learning. So the question is, what is this algorithm? Why is it so important? Why is it so effective? What else can it do? If you scale the algorithm beyond its current size, what problems can it solve that are unsolvable today? Our breakthrough was realising that AlexNet is not just AlexNet. AlexNet is a deep learning method that allows you to learn any function. So 15 years ago, I told everyone, hey, guess what? We just discovered a universal function approximator. We just found a universal function approximator. We can give it, we can give it the answer to almost any function, and it can learn what that function is. For many functions, you don't need to be very precise, in fact, it's impossible to be very precise. So most interesting problems exist in this imprecise way.

The day we realised we had a universal function approximator, the question became, what does this mean for the computing stack? What does this mean for software? Which industries might be impacted? etc. We almost immediately started working on computer vision. Almost immediately started working on robotics, autonomous vehicles, because that basic ability, you can imagine, can solve some important problems in computer vision and robotics. So I think the major breakthrough is just that it was much more fundamental than AlexNet. It's a way of writing software. Its impact on processors, middleware, algorithms, and applications is what I now describe as the "five-layer cake". The entire industrial stack, I envisioned about 15 years ago that it all needed to be reinvented.

How to Build a First-Principles Organisation

It's just about asking questions, reasoning from first principles, asking "If so, then what?" questions, "What if this could be better, then what?", asking all these fundamental questions about something you observe that has a major impact. I mean, one thing that really impressed me is, to what extent do you go deep into the details, read papers, talk directly to the chief scientists who proposed these theories? Do you have any advice for the people in the audience? I mean, that's like true founder mode, but at the same time, you have an organisation, you have executives, and those who say, "Here's the curve, we need to stay on this curve." You know, sometimes that can cause dissatisfaction. Do you have any advice for people on how to build an organisation and navigate this? Like, how to build an organisation that allows you to think from first principles? Because if Fortune 500 companies did this, they might look more like NVIDIA than not, and you seem to have built a very unique company.

My mindset, when I'm in that state, always starts with curiosity. I have many questions myself, and like anyone, I look for the shortest path to the answer. But many times, the answers from people around me may not be satisfactory, I might have other questions, perhaps they are busy doing other things, pursuing other goals. So my first inclination is to discover the answer to my own curiosity. My second inclination is, if I find information, and this information field or specific area might be important to someone, might be important to our company, then my next inclination is, how can I learn as much as possible in order to serve the company and share it with others. It's no different from when you share knowledge. I mean, I watch your podcasts, I watch your videos, I really like them. You are sharing ideas with everyone. In many ways, I think the CEO serves the company and everyone who works there. You want to empower them with some insights.

So, that's really the source of it. It's more of a personality trait than a management technique. You know, I want to empower you. This is a very important observation I just made. Now, the reason for having to stay close to the ground and go into the details is partly because technology is usually complex or changing rapidly. Especially when it changes as fast as our world, unless you have a tactile sense of what's happening, it might feel like it's changing too fast to understand. But if you understand its first principles over time, then everything makes sense. You know, it's a bit like surfing, I imagine. I don't know how to surf, but I can imagine it's a bit like surfing. You go out to surf. To me, that looks like chaos, but to a surfer, somehow, they can orient themselves. They can read the waves, they know how to stay on the crest. So, I think being a CEO is very similar. You have to learn how to surf. To learn how to surf, you must understand the waves. You must be able to read the wind, have good timing. You can't have these unless you try and actually do it. So, part of it is to inform myself, part of it is to try to figure out the problem, break it down so that the company can learn it in an actionable way. Part of it is to inspire others. These are all basic traits of everyone here. You don't need to change your personality or behaviour when you become CEO. You can continue to be yourself. One thing I learned a long time ago, I don't know where I saw it, but you know, the CEO or founder is building a race car that you are going to drive. You are building an F1 car, but you are building it in a way that you can drive it. You should make the car adapt to you. I think someone asked me, Jensen, if you don't use conventional management techniques and organisational techniques, what will happen when you leave the company? Well, you know, when I die at work one day, I told them, they will have to reshape the company for the next CEO.

The wisdom in this is that we are F1 drivers. We are racers. The world is very competitive, and we have to maintain, we have to win the race, we have to complete our mission. So, whatever it takes to make the car adapt to you, whatever it takes to make the organisation adapt to you, that's what you should do. The next CEO, whatever their personality, they can figure it out themselves. That's great. I mean, it really seems like any changes you make to the car would only slow you down and make you lose races that don't suit you. Yes. Or we constantly adjust the car to our needs, and that's what I've been doing. I constantly adjust the company, constantly reshape business processes and ways of working so that I can be more effective for the company. True founder mode. Yes. Founder mode. Founder mode can last 34 years. Correct. From zero to five trillion.

Systems Thinking is the New Programming

Garry Tan: I'd love to change the subject and talk about the frontier algorithms you're most interested in now. I mean, I like how you go deep into materials science, all the way up to the application layer. You were the first person on stage to talk about OpenClaw and now Hermes Agent. I'm wondering if you could take us through how you think about different stages in your day? From materials to chips, to data centres, and even to the application layer, like how people will work? It's a bit like the concept of a full-stack AI factory.

Jensen Huang: This might be one of the most useful skills in the future. In fact, just listening to you talk about technology and how you use it, one of the most important things is system understanding, system awareness, system design, system organisation, but the core is systems thinking. The reason is that most underlying, necessary work will be done by agents anyway. It will be automated anyway. So whether it's my generation, dealing with compiling chips, synthesising transistors, gates, and functional blocks—all of that is now synthesised, and most of our designers are system designers. In the case of software, most software will be done by agents anyway. So you must be more capable of thinking abstractly about systems. What problem do you want to solve? What are the constraints? Where are the inputs, where are the outputs, where does the information come from, what is the rate of information entering and leaving the system, what are the constraints? Is it the processor, memory, or network? Understanding these system issues at a sufficiently technical level will be very helpful for everyone here. I don't think this fundamental knowledge will become useless. I think it will become more useful. So I do my best to understand systems.

Speaking of agents, the fact is, we already have rough, recursive self-improvement. Every time you use it, it improves the markdown file. Every time you use it, it updates its long-term memory. The long-term memory is being processed, either compressed or converted into a knowledge graph, etc. It is constantly being improved, asynchronously, so the agent gets smarter each time. But the problem remains, and I think it's helpful for everyone to solve, is how to have very, very fine-grained control? If you don't use RAG, conditional input, or all our prompts directly fed into the output, that's too coarse.

So we need to be able to condition, to control the agent, all the way down until it comes up with a plan. I change one word in the plan file, and that word produces a tiny difference, not a complete difference, but a specific difference. Maybe it's a pixel, maybe it's a triangle, maybe it's a component in a CAD file, maybe it's a layer, a via, a connection, and then it regenerates all other content. I think that level of control and collaboration with agents will be disruptive. We don't need agents to be 100% accurate or 100% high quality to use them. It could actually be 80%, and we help it finish the rest, or it could be 99%, and we help it finish the rest. So I think controllability might be the biggest breakthrough we need for agents at all levels.

Should People Own Their Own AI?

Garry Tan: Do you think people will… I mean, with Hermes or OpenClaw, it feels like this might actually be somewhat existential. People should control their own personal AGI. They shouldn't outsource this app, just let it run in the cloud, have someone else's agent tell you what to do. You should want it to be your own. Is this one of the driving forces behind NVIDIA's deep involvement?

Jensen Huang: I think first, I need to understand agents because agents are the new software. The way this new software is processed has a significant impact on computer architecture. The deeper we understand the nature of agents and how they differ from chatbots or early reasoning, the better we can design systems. We kind of have to live 5 to 10 years in the future because building a system takes about three years. It takes a few years to ramp it up, and you want them to use the computer for 10 years afterwards. So you must live in the future for a while.

Therefore, for agent systems, from first principles, for us, it's: what is the workload? What is the algorithm? How will it evolve? Where are the bottlenecks? Where is Amdahl's law a problem? How does it scale? What happens with concurrency? How do you handle sandboxes? How do you handle MCP? How do you handle working memory, long-term memory? How do you make all these autonomous, asynchronous systems work all the time? So what design architecture is completely reasonable for this? We have to go explore and discover that.

Then the second thing is, I want to use agents myself to make NVIDIA go faster. So we have, Boris is in the background, we have Cloud Code running autonomously in sandboxes across NVIDIA. That's really great. Some people use CodeC, some use Cloud Code, some use Cursor, some use Cognition. We let a hundred flowers bloom, let people choose the tools they want to use, and then we learn from all the usage.

How NVIDIA Views Agents Internally

The second part is simply helping the company move faster by using these tools. The more they use them, the more we learn how to make them work better in the future. The final part is discovering future solution technologies. Maybe when we saw an early version of chain-of-thought from Stanford about ten years ago, maybe eight years ago, the question was, how effective will it be in reasoning, how scalable is it? For example, in computer vision, what impact would it have if we could reason from prior knowledge? Then came a huge breakthrough. Of course, just thinking within that small domain, you realise maybe we don't need that much data to train self-driving cars. This led us to create Alpameo, the world's first thinking self-driving car. With only about a million miles or a few million miles, it's an excellent self-driving car.

The reason is that it's a bit like us, right? We don't need to drive that many miles to drive reasonably well in our lives. The reason is that we have prior knowledge from language models. We can decompose a situation we've never seen before and construct it using things we understand and are very familiar with. So this is an example of seeing something and then realising its impact later. When agent systems appeared, it was clear that large language models need memory, prior knowledge, tools, and ways to network with other agents. So once you see some early signs and can reason about the future, it helps you leap into the future.

I feel like I'm starting to see a pattern around NVIDIA. Like you see a difficult problem, there's a new algorithm, something new is happening, and then you're actually there with the open-source community. I mean, I remember when OpenClaw came out, people said it was unsafe, but you launched a set of sandbox toolkits that can be put around any tool suite and make it safe. So when I saw OpenClaw, my first thought was, first, I learned about it, and then almost without thinking, you realise we just designed the modern computer. This is the operating system that will host large language models. In many ways, OpenClaw feels like a Linux moment to me. Yes. Now everyone can build their own AI. I'm very excited about this. We contacted Peter and said, "Hey, you know, all of NVIDIA's engineers are your engineers." That's exactly what I told Peter. There's a battleship outside your house. You break down the problem the way you want, and we'll contribute the way you want. Same for the Hermes team. I'm very excited about the work they're doing. I truly believe the world needs everyone to be able to build their own AI.

Of course, you can, and I encourage everyone to use cloud services as much as possible. Everyone should use ChatGPT and Claude. Everyone should use those. But if you need to build your own AI because you are a company, you need to build your own domain-specific AI. Now you have Hermes, you have OpenClaw, you have various methods, you have LangChain, you have DeepAgent, all these different ways to build your own AI. And frankly, it's relatively easy because the software is smart. AI is smart, so AI must be smart enough that you can adjust it easily. So I think we want to encourage everyone and every company to build their own AI. Who knows what innovations open source will bring? I feel like all the alpha comes from building your own AI. I mean, if others are using off-the-shelf, and you have something that can recursively self-improve, it's yours… I mean, these people are casual about markdown files. They say, "Oh, haha, it's just text." But text is intelligence. We are in a different era. Words are thoughts. Yes. Words are thoughts. It turns out you can try to think without words.

Garry Tan: Yes. So, changing the subject again, I mean, every time you move someone else's cheese, everyone gets a little worried. The fact that intelligence will be accessible is really great. I think it bodes well for everyone here. What do you think will change in the economy? What do you think will happen on a broader level?

AI and Jobs

Jensen Huang: Obviously, what I'm going to say is not balanced. You know, we will automate tasks. We will automate cognitive tasks. If that task is someone calling, sending a bunch of text to you over the phone, and your job is to provide a response. If all the information is at your fingertips because you have all the databases here, you should be able to answer that question completely. In that case, that task will be automated. Okay, ignore that for now. Not that we ignore it, but my point is I'm going to answer the question about the truly huge opportunity. So many tasks will be automated. Many jobs, every job will change, and there will be a lot of new jobs. I think that's what we know. The bottom line is this. The evidence shows, and it's completely reasonable, that AI and automation are creating jobs everywhere.

The narrative about AI destroying jobs is exactly the opposite. AI eliminates tasks. AI automates tasks, but it doesn't necessarily eliminate jobs. The reason is that a person's job has a purpose, and that purpose contains many tasks. Some of those tasks can be automated. Many tasks cannot. The evidence shows that we have now automated coding, which is a task, but the job of software engineer seems to be growing. Right? The number of software engineer positions has grown 10% year-over-year. The task of reading radiology scans has been automated, but the number of radiologist positions has increased by about 20% in the past few years, even though AI has taken over the entire field. The reason is the backlog of patients is very high. Now doctors and hospitals can take on more patients. To take on more patients, you need more nurses, more radiologists. The same is true in software. The backlog of ideas, ambitions, and aspirations is so high that if we can automate the task of programming, we can hire more software engineers to do more things, we can be more ambitious. This is true in all fields. They said Harvey would eliminate all paralegal jobs, and the number of lawyers would decrease. It turns out the number of paralegals is growing like crazy. The reason is that the backlog of litigation is very high. Now these law firms can handle more cases. To do that, you have to hire more people. So this is a classic example of productivity gains driving growth, and growth driving more employment. This is why there are more jobs today than when I graduated.

We've talked a lot about software and agents. Another exciting thing NVIDIA is at the forefront of is physical robotics. How soon do you think it will happen? I think you might have even said in the past, just this year, about when we can expect practical robots. What's the latest thinking?

The ChatGPT Moment for Robotics

Yes. The moment I saw we generated a video, that was a great moment for me. The moment I saw we generated a video. I mean, we did the original work on progressive GANs, okay? We did original work on conditional GANs, years before people saw the first generated videos, in our lab. We were driving a simulator entirely generated by video, a computer entirely generated by a neural network. So when I saw we generated motion, if I can generate a video of a finger moving, if I can generate a video of a hand picking up a cup, why can't I make a robot do the same thing? So when I saw that moment of generative AI happening, I realised robot motion was imminent. So now the question becomes, how will a robot understand… generate motion that follows the laws of physics. How does it understand causality? How does it understand friction, tension? How does it understand the laws of physics? This started us on the journey to create what we now call physical AI. Now everyone calls it physical AI.

Physical AI. We started working on a world foundation model, an AI that understands the laws of physics and how the world works. We started the journey of researching robotics. I would say the ChatGPT moment for robotics happened a few years ago. The reason is, remember when ChatGPT first appeared? It didn't do anything productive. It didn't do anything useful, but it opened our imagination to what's possible. I would say a few years ago, you know, robots were moving around, we could do reinforcement learning, fine-tune it and ground it in physics. That really happened a few years ago. So what do we need to do now? We need to do all the things we are now doing for agent systems. We need to create environments for them to learn and evaluate. So we have to do real-to-sim to create environments. We need to do, we need to generate simulators based on simulation physics and generative physics simulation. So Isaac Sim, Cosmos, and all our work in that field are related to simulation.

The final part is simulation-to-real. This part is about reinforcement learning, grounding it in physics, grounding it in all the electromechanical systems required for robots. But I think these three fundamental systems build the evaluation, if you will, the post-training for robotics. I think we will see it soon.

Where Physical AI Will First Appear

Garry Tan: Great. Where does physical AI first become economically realistic? Have you seen it yet?

Jensen Huang: We speculated that robotics would come and decided that the first application of robotics, which must have a sufficiently large market, relatively standardised technology so we can scale and get a flywheel effect, and have real economic value, is the autonomous vehicle. So, inside Waymo, our NVIDIA chips. In Tesla, we are in the car. Now we are in the data centre. Mercedes, we are in the data centre, we are in the car, we are the software stack. We developed Alpameo and open-sourced it. The reason we open-sourced the autonomous vehicle stack is that agriculture needs it, mail delivery needs it, warehouse AMRs need it. There are many different ways to apply autonomous navigation. None of these markets alone is big enough to be the autonomous vehicle market. We think it's so diverse that we will create the entire stack. So we are working with autonomous vehicles in various different places. Our robotics business, autonomous vehicle business, basically the physical AI business, is probably close to $10 billion. So it's already very large. This will likely become one of the largest industries in the world. It will take more than two or three years but less than ten years. So this will be our next hundred-billion-dollar business.

Great. I'd like to take a moment. I think this is exactly the right audience. Perhaps as a stage, we can welcome Jensen to X. Welcome to X. I mean, you posted your first post. Thank you for your leadership.

Jensen Huang's First Post on X

You know, this just shows how introverted I am. I didn't post my first post on X until 2026. You know, I'm probably the last person on Earth to do this. But what I posted was too important to me, too important to the industry, too important to the world. So I overcame my shyness and put my first thing on X.

Garry Tan: No, thank you for your leadership. I mean, open source, open weights, open source models are extremely important for what all of us here want to do. If it weren't for open source, the mobile cloud industry would never have happened. If it weren't for open, not Linux, not Kubernetes, not all these platforms, not TensorFlow, or more importantly PyTorch, right? And early versions, Cafe, right? Torch, I mean all those early versions were open source. If it weren't for all of these, how could we have modern AI? Thank you for your leadership. Your voice is extremely important here.

Before we end, I feel like I really resonate with your story. I think everyone here would be happy to hear your wisdom along the way. I mean, given all the changes you see happening, all the algorithms that will dominate society, what should a young person learn now that, based on your observations, will still be valuable? Some things I see today, some entrepreneurs I meet today, are really, really inspiring.

What to Learn That is Still Valuable

Jensen Huang: An important takeaway is that simple things will be automated. When I say simple things, I mean software, coding. The idea of sitting in front of a computer, actually writing code to solve a problem, will obviously be automated. When I was growing up, we had to do long division. Goodness, who still needs to learn long division, you know? So that was encoded away, automated away. So, I think simple things will be automated. Difficult problems, hard sciences: physics, chemistry, biology, computer science, computer engineering, systems thinking, especially the intersection of these fields. Those difficult problems will never go away. So AI is just an incredible tool to help us become more ambitious, more eager to solve these extremely large and incredibly difficult problems.

So if you look at my generation, when I graduated, a chip designer might design a chip with a thousand transistors. That was already a very large chip. Now, designing a chip with a trillion transistors. If someone told me, Jensen, our next chip is a trillion transistors, I would say, okay, no big deal. The reason is that we are now so ambitious. The scale of the problem, the scale of the task, is no longer the issue. So you don't need to worry about how much code you write, how many engineers you have. You don't need to think about those things anymore. You just need to think about what problem you want to solve. So I think deep tech, deep science stuff, understanding the intersection of technology and social issues, understanding market gaps and opportunities. I think those are still there. The better you are at systems thinking, the more you can orchestrate millions of agents to solve problems autonomously, the more advantage you have. That's why systems thinking will be so important. But beyond that, I think the world will continue to have plenty of great challenges for us to solve. Go to school as usual. Stay in school. Stay in school.

I usually like to end by looking at the crowd. There are many people. I mean, I said at the beginning, I sincerely look at the crowd, and the people I see are no different from us. We are really just people who understand technology and love systems.

Garry Tan: What advice would you give to the people in this room? Do you see yourself in this room? I'm curious, if you could send a telegram to yourself at age 18 to 22, what would you say? I can tell you exactly how I felt when NVIDIA was founded.

The Mindset You Should Have: "How Hard Can It Be?"

Jensen Huang: When we three started, my feeling at the time was that there was so much I needed to know, so much I needed to learn, and I didn't know. I told you before, there was no YouTube, no YC, no one taught you how to start a company. So I went to a bookstore and bought a book titled "How to Start a Company". Unfortunately, that book was 500 pages long. So I thought, by the time I finish reading it, I would be bankrupt, and Lori and I would have no money. So reading it was pointless. But one thing I remember very, very clearly is how scared I was when I went to raise money, because I felt I was going to talk to a group of people, and I didn't know how to answer their questions. And it's true, even today I almost don't know how to answer their questions.

But what I learned is that, it turns out, those things don't matter. You will always have things you don't know. Every day, the world changes. Technology changes. Obviously, this is the best time to start a company in the last 60 years. The entire industry has changed. From a technological perspective, this is a complete reset. The most important technology in human history—the computer—has been completely reset. So this is absolutely the single best time to start a company. I envy all of you and the opportunities before you. I mean, it's going to be incredible. So, on one hand, the timing is perfect. On the other hand, technology is changing so fast. So, the question is, what is the right feeling for you? Finally, I told you the story about buying another book, that textbook. I think my psychology and feeling towards all new experiences, new technologies, new markets, and new dynamics today is, I look at it and say: "This is important. I must learn it. I must act on it. I must start as soon as possible. How hard can it be?" I always have that feeling: how hard can it be? Honestly, it's much harder than you imagine. But you don't want your mind to be there. You want your mind to be "How hard can it be?" Let the pain come little by little. Don't imagine how hard it will be, let that turn into anxiety, and then not do anything. You want to have in your mind, "How hard can it be? Anyway, I have a bunch of AI agents to help me." So, how hard can it be? Then you start doing it. So, that's probably the entrepreneur's attitude. You know, you have to learn a lot along the way. You have to trust your ability to learn. Learning is the single most powerful superpower. If you go in with the attitude of "How hard can it be?", "If others can do it, so can I", and realise it will be hard, you just need to be resilient every day to overcome it. You don't have to overcome your entire life in one day. You just need to overcome that morning. You just need to overcome today. So, no big deal. Get through today first. Towards tomorrow?

Strive towards tomorrow. Continue to pursue your dreams. Everything else, if you persist long enough, NVIDIA happens. So, I think if there's any wisdom I can share, resilience might be the most important thing. If you believe in something, start doing it. Don't let your mind stop you from pursuing it because of fear, anxiety, lack of confidence, or any reason. Tell yourself, I will find my way through by continuous learning. Everyone, keep moving forward.