Jensen Huang at YC Startup School: Three Textbooks from Fry's Fixed NVIDIA
The NVIDIA CEO on learning as the only durable moat, why agents at 80% are already useful, and the $10 billion physical-AI line he calls NVIDIA's next $100 billion business.
Make sure to subscribe at Fireside Alpha (link) to receive daily breakdowns on the best business and tech podcast episodes.
The short version, if you don't have the hour:
- NVIDIA started in 1993 on a 3D-graphics algorithm Huang now calls "exactly wrong," and knew it by 1995 with 35 to 40 rivals already in the market. The fix cost a couple hundred dollars: three OpenGL textbooks from Fry's, handed to the engineers.
- The Sega contract was about $12 million for the console that became Dreamcast. Huang told Sega's president the tech would not work but he still needed the money, and the roughly $5 million that came anyway "kept us alive."
- NVIDIA went public in 1999 at a $300 million valuation. Huang and Garry Tan riffed the arc from there to the trillions, "zero to five trillion" across 34 years of founder mode.
- Huang read AlexNet in 2012 as the "universal function approximator," not a vision result, and started work on computer vision, robotics, and self-driving right away.
- Software engineer jobs are up 10% year-over-year and radiology jobs up about 20% even as AI took over reading scans, and paralegals are "growing like crazy" after Harvey. Huang's line: "the narrative about AI destroying jobs is exactly backwards."
- NVIDIA runs Claude Code, Codex, Cursor, and Cognition in sandboxes across the company, a "thousand flowers bloom." Agents at 80% accuracy are already useful, and humans finish the rest.
- The robotics, autonomous-vehicle, and physical-AI business is already about $10 billion, and Huang calls it NVIDIA's next $100 billion business, arriving in under ten years. NVIDIA chips sit in Waymo, Tesla, and Mercedes.
- "How hard can it be?" is the mindset. Huang says it's always way harder than you think, and the answer is to overcome today, not to imagine the whole climb at once.
In 1995, Jensen Huang drove to a Fry's with a couple hundred dollars in his pocket and came back with three textbooks. NVIDIA was barely two years old, and the technology it was founded on had just turned out to be wrong. The plan had been to reinvent 3D graphics, turn every PC into a game console, and shrink a supercomputer's algorithm onto a card. None of it worked.
"It turns out the algorithm was exactly wrong. The technology that founded the company turns out to be exactly wrong. In 1995, we realized that. It was almost too late because by then there were some 35, 40 other companies that were building 3D graphics for PCs."
Jensen Huang
▶ Watch · 00:02:43
The textbooks were about OpenGL and how to design its pipelines. Huang handed them to the engineers, and NVIDIA relearned computer graphics from scratch, then spent the next 25 years inventing most of what the field runs on.
"We actually started the company, raised money, and bought textbooks when you think about it. The big lesson for me is technology's changing all the time. So long as you're able to confront reality, so long as you are able to learn, the technology itself actually doesn't matter."
Jensen Huang
▶ Watch · 00:04:17
NVIDIA has run that play ever since. Something looks worth doing, the company goes and learns it, walking in with the question Huang asks about everything: how hard can it be. He's the first to say the honest answer always turns out much harder than he expected. He asks it anyway.
The idea was the algorithm domain, not the chip
The technology was wrong, but the underlying idea was not. NVIDIA's founding idea was that you could pair an accelerator with a CPU and crack problems otherwise too hard, molecular dynamics, image processing, inverse physics, later deep learning. What the company worked out early is that the product was never the chip.
"In order to create the company that we have today, we realized early on that it's not about building a great chip. It's about accelerating an algorithm domain."
Jensen Huang
▶ Watch · 00:06:03
3D graphics was just the first algorithm domain NVIDIA owned, learned from a textbook anyone could have bought, and what it added was the product people would pay real money for. The accelerated-computing call turned out to matter enormously, and so did choosing the algorithm over the chip.
The Sega call that kept the lights on
Sega had hired NVIDIA to build the console after the Saturn, the machine that became the Dreamcast, for something like $12 million. When Huang realized his technology couldn't deliver it, he flew to Japan and told Sega's president, Irimajiri-san, two things at once: pick another vendor, and please still pay us.
"So what you're telling me is what I contracted you to do, you can't do, but you would like all the money on the contract. And I said, you got it. That's exactly right."
Jensen Huang
▶ Watch · 00:08:43
Sega paid.
"You don't invest in companies, you invest in people."
Jensen Huang
▶ Watch · 00:09:12
The roughly $5 million that came in kept the company alive and bought time to figure out what to do next. Without it there's no NVIDIA.
From a $300 million IPO to five trillion
NVIDIA went public in 1999 at a $300 million valuation, which Huang says felt like real money then. Sega sold its stake the moment the company went public. Tan had heard it went for around $15 million and put NVIDIA today north of a trillion dollars, and Huang agreed it was more than that. Tan had a name for where that left him.
"You're the man who controls the spice."
Garry Tan
▶ Watch · 00:10:05
The two riffed the whole arc as founder mode holding for 34 years, from zero to five trillion.
AlexNet as the universal function approximator
Huang saw AlexNet in 2012 the same week everyone else did. What he took from it was different. He was always scanning for an algorithm to accelerate, OpenGL or SQL or a molecular-dynamics code, and through that lens AlexNet read as something much bigger than a computer-vision result. He took it as a general machine for learning any function.
"We just discovered the universal function approximator. We could give it the answer for almost any function, and it could learn what the function is. And for a lot of functions, you don't have to be precise. And in fact, it's impossible to be precise."
Jensen Huang
▶ Watch · 00:11:50
Once deep learning was a way to write software rather than a trick for images, the question became what it does to the whole computing stack, the processor and middleware and algorithms and applications he calls the five-layer cake. He says he pictured reinventing that entire stack about 15 years ago, and started NVIDIA on computer vision, robotics, and self-driving almost right away. Underneath it is a set of plain questions he keeps asking: if this, then what, and if this can get better, then so what, run against anything that looks impactful and reasoned out to first principles.
Curiosity first, and the surfing problem
Tan asked how you build a company where the CEO reads the papers and talks to the principal scientists, the thing he called true founder mode, without the org fighting it. Huang says it doesn't start as a management technique. It starts with his own curiosity and a habit of chasing his own questions when the nearest answer isn't satisfying. The second move is turning what he learns into something the company can use, which is how he defines the CEO job, in service of the people working there and out to empower them with some insight.
Being in the weeds isn't optional when the technology moves as fast as his does. Without a tactile feel for what's actually happening, he says, it all reads as chaos. He pictures surfing, a thing he admits he can't do.
"You have to learn how to surf. And in order to learn how to surf, you have to understand the waves, and you have to be able to read the wind, and you have to have good timing. And you can't have any of that unless you try and unless you actually do it."
Jensen Huang
▶ Watch · 00:16:36
Build the car you can actually drive
You shouldn't run someone else's management system. Think of an F1 car built to be driven by you specifically, not by an idealized driver.
"You're building a car that you are going to race. You're going to build an F1 racer, but you're going to build it in a way that you can drive. You should adapt the car to you."
Jensen Huang
▶ Watch · 00:17:33
When people ask what happens to the company after him, he says they'll reshape it for the next driver, and that's fine. He and his co-founders are the racers, the world is competitive, and the job is to fit the car to whoever's driving now. Huang says he's constantly tweaking the company and its processes to be more effective, the same move as tuning the car every race.
Systems thinking is the new coding
Ask Huang what skill matters most now, and it's systems thinking. The reason is straightforward: the low-level work is getting done by agents anyway, the way transistor synthesis already automated chip design, so most of NVIDIA's designers are systems designers.
"Most software is going to be done agentically anyhow. So you have to be much more able to think abstractly about systems."
Jensen Huang
▶ Watch · 00:20:40
He argues agents already do a coarse form of recursive self-improvement, updating markdown files and long-term memory that gets compacted into knowledge graphs asynchronously, so they get a little smarter every time you use them. The piece he wants most is fine-grained control, being able to change one word in a plan file and get a specific difference rather than a full regenerate. The bar for using them is lower than people assume.
"We don't need the agents to be 100% accurate, 100% high quality, in order for us to use it. It could literally be 80%. And then we help it the rest of the way. Or it could be 99%, we help it the rest of the way."
Jensen Huang
▶ Watch · 00:23:02
How NVIDIA runs agents on itself
NVIDIA lives this close to agents for reasons that feed each other. Agents are the new software, and how that software gets processed drives computer architecture, so the more intimate NVIDIA is with the workload, the better it designs systems. That's why Huang insists on living five to ten years ahead, since a system takes about three years to build, a couple more to ramp, and customers use it for ten after. Speed is the next piece, and NVIDIA doesn't standardize on one tool.
"We've got Claude Code autonomously running in sandboxes all over NVIDIA. That's really fantastic. Some people use Codex, some people use Claude Code, some people use Cursor, some people use Cognition. We let a thousand flowers bloom, let people select the tools they want to use. And then we learn from all of that."
Jensen Huang
▶ Watch · 00:25:21
The last piece is discovery. Tracing the early Chain of Thought work out of Stanford, eight to ten years ago, Huang landed on the idea that a car might not need much data if it can reason from prior knowledge, which produced Alpamayo, a thinking self-driving car he says drives well on just a million or a couple million miles. He compares it to a human, who needs few miles because language gives you prior knowledge to decompose a situation you've never seen into pieces you already know.
Own your AI, and the open-source lineage
Tan pushed on whether people should own their own AI rather than rent an agent that tells them what to do. Huang separates the layers. Use the cloud services heavily, and also build your own when you're a company that needs domain-specific AI.
"I do think that the world needs the ability for everybody to build their own AI. I encourage everybody to use cloud services as much as possible. Everybody should use ChatGPT and Claude. Everybody should use that. But if you need to build your own AI because you're a company and you need to build your own domain-specific AIs, now you have Hermes, and you have OpenClaw."
Jensen Huang
▶ Watch · 00:29:06
Huang called OpenClaw a Linux moment, the operating system that'll hold a large language model, and said NVIDIA's engineers are Peter's engineers on it, same offer to the Hermes team. He grounded the point in the open-source stack modern AI already sits on, running the list himself: Linux, Kubernetes, TensorFlow, and PyTorch, plus the early versions of Caffe, Torch, and Theano. Without open source, he says, there's no mobile cloud industry and no modern AI. That lineage was the subject of his first-ever post on X, made in 2026 after years of holding out, when he figured he was probably the last human on earth to do it.
The jobs argument, backwards
Huang inverts the standard fear about AI and employment. He grants that automation is uneven and that cognitive tasks, like answering a call when all the information is at your fingertips, will get automated. Then he separates task from job.
"The narrative about AI destroying jobs is exactly backwards. AI eliminates tasks. AI automates tasks away, but it doesn't necessarily eliminate jobs."
Jensen Huang
▶ Watch · 00:31:54
He points to fields where AI took over a core task and headcount went up anyway.
"The number of software engineer jobs year over year has increased 10%. The task of reading radiology scans has been automated, but the number of radiology jobs has increased some 20% in the last several years, even though AI has taken over the whole field."
Jensen Huang
▶ Watch · 00:32:32
The driver he points to is backlog. Hospitals have a queue of patients, so cheaper reads mean more admissions and more nurses and radiologists, not fewer. Law is the same story with Harvey, where he says paralegals are growing like crazy because firms can finally push more of a high backlog of cases through. Software works the same way in his telling: the backlog of ideas and ambition is so high that automating programming lets you hire more engineers to do more, not fewer to do the same.
The ChatGPT moment for robots already happened
Huang dates the turn in robotics to the moment NVIDIA could generate video of articulation. NVIDIA did early work on autoregressive and conditional GANs and was driving a simulator generated entirely by neural networks before public video generation existed. The leap was seeing a generated hand and asking why a real one couldn't follow.
"If I can generate video of a finger moving, if I could generate video of a hand picking up a glass, why can't I cause a robot to do the same? The moment I saw that generative AI happening, I realized that robotics articulation was around the corner."
Jensen Huang
▶ Watch · 00:35:02
The realization kicked off physical AI, starting with a world foundation model that understands the laws of physics, friction, tension, and causality. Robots already had their ChatGPT moment a couple of years ago, Huang says, the opening act that expanded imagination before it did anything useful. What's left, he says, is the same loop NVIDIA runs for agents: build environments to learn and eval in, real-to-sim and sim-to-real, grounded in physics, with Isaac Sim and Cosmos doing the simulation.
Where physical AI pays first: $10 billion now, $100 billion next
Huang says NVIDIA picked self-driving cars as the first robotics market on purpose, because it was large, standardized enough to scale a flywheel, and economically real. NVIDIA sits in Waymo, in Tesla, and in Mercedes, in the car and in the data center. It open-sourced the Alpamayo self-driving stack, he says, because agriculture, mail delivery, and warehouse AMRs each need autonomous navigation and none of them alone is a big enough market to justify building it.
"Our robotics business, autonomous vehicle business, basically physical AI business, is probably almost $10 billion. So it's really, really big already. Likely this will be one of the largest industries in the world, and it'll take longer than two or three years. It'll take less than 10. And so this will be our next hundred billion dollar business."
Jensen Huang
▶ Watch · 00:38:38
What still matters to learn
For an 18-to-22-year-old room, Huang splits the world into what automates and what doesn't. The simple stuff goes, and he means coding specifically, the way long division went in his generation. He's blunt that sitting in front of a computer writing code is the concept that gets automated away. What stays is the hard science, and especially the places where fields cross.
"The hard problems, the hard sciences, physics, chemistry, biology, computer science, computer engineering, systems thinking, and particularly the domains that are intersecting, those hard problems will never go away."
Jensen Huang
▶ Watch · 00:41:50
Ambition scaled with the tools. Huang points out that a large chip in his early career held about a thousand transistors, and a trillion-transistor chip is now unremarkable, because the scale of the task stopped being the constraint. You don't have to worry about how many engineers or how much coding, he says, only about what problem you're solving. The better you are at orchestrating millions of agents, the better off you are, which is why he keeps landing on systems thinking.
The mindset: how hard can it be
Huang closed on the feeling he had when NVIDIA was three people, that there was too much he didn't know. There was no YouTube, no YC, no one teaching you to start a company. He bought a 500-page book on how to start one, figured he'd be out of business before he finished it, and never read it. What he learned is that not knowing doesn't matter, and he says he barely knows how to answer investors' questions even today. For the room, he told them, this is the greatest time in 60 years to start a company, because the computer, the most important technology in human history, has been completely reset.
"I always have this feeling, how hard can it be? And truth be told, it is way harder than you think. But you don't want your mind to be there. You want your mind to be, how hard can it be? Let the suffering come to you a little bit at a time."
Jensen Huang
▶ Watch · 00:46:55
Learning, he says, is the single greatest superpower, and resilience the single most important thing. The way you get resilient is by shrinking the horizon until it's survivable.
"You don't have to overcome life in one day. You just have to overcome that morning. You have to overcome today."
Jensen Huang
▶ Watch · 00:48:02
Stick with it long enough, he says, and NVIDIA happens: a company that started on the wrong algorithm, relearned graphics from three Fry's textbooks, kept the lights on with $5 million of Sega money it couldn't technically earn, went public at $300 million in 1999, and now runs Claude Code in sandboxes across the building while it stands up a physical-AI business he expects to clear $100 billion.