AGI Is Already Here. Building NVIDIA Is Not.

Lex Fridman opened by noting that NVIDIA's winning condition used to be the best GPU. It still is, but Huang has expanded the job to what he calls extreme co-design: GPU, CPU, memory, networking, storage, power, cooling, software, the rack, the pod, and the data center. The reason, Huang said, is physical. The problem you want to solve no longer fits inside one computer to be accelerated by one GPU. You want to go faster than the number of computers you add. You put 10,000 machines on the floor and you still want a million-fold speedup.

That forces you to break the algorithm apart: shard the pipeline, the data, and the model. Once the work is distributed, everything becomes a bottleneck. Huang reached for Amdahl's law. If computation is half the workload and you speed it up a million times, the whole job still only gets twice as fast. At that point the CPU is a problem, the GPU is a problem, the network is a problem, the switch is a problem, and so is spreading the workload across the machines. It is, he said, a massively complex computer science problem, and Moore's law no longer bails you out because Dennard scaling has slowed.

Fridman asked how you get world-class specialists in high-bandwidth memory, NVLink, NICs, optics, copper, power delivery, and cooling into one room. Huang's first answer was that this is why his staff is so large. Extreme co-design, as he defined it, means optimizing across the stack — from architectures and chips to systems, system software, algorithms, and applications — and then past CPUs and GPUs into scale-up switches, scale-out switches, power, and cooling. The machines are energy efficient, he said, but in aggregate they still consume a lot of power. The "how" of doing all of that at once is what he called the miracle of the company.

When you design a computer you need an operating system, Huang said. When you design a company you should first ask what you want it to produce. He has looked at a lot of organization charts, and they all look the same — hamburger companies, software companies, car companies. That makes no sense to him. A company is the machinery that produces the product, and its architecture should reflect the environment it lives in. His direct staff is 60 people, and the number is still going up. Almost all of them have a foot in engineering: memory, CPUs, optics, GPUs, architecture, algorithms, design.

He does not do one-on-ones, because it is impossible at that span and because no conversation is ever about one person. The company presents a problem and all of them attack it. When they talk about cooling, the power people are in the room; when they talk about memory, the networking people are listening. Whoever wants to tune out can tune out, but Huang will call people out if they should have contributed and did not. "Hey, come on, let's get in here." The company, he said, is doing extreme co-design all the time, and the org chart has to look like the product or the phrase is empty.

NVIDIA started as an accelerator company. The problem with accelerators, Huang said, is that the application domain is too narrow. Specialization is great for the job; the hidden cost is that a small market caps your R&D budget, and R&D budget caps how much influence you can have in computing. They had to become an accelerated-computing company, which meant walking a contradiction: the better a computing company they became, the worse they were as specialists. He connected the two words on purpose. The path was stepwise. First a programmable pixel shader. Then IEEE-compatible FP32 in the shaders, which made people working on stream processors notice that this compute monster could now run their code. Then C on top of FP32 — Cg — which led to CUDA.

Putting CUDA on GeForce was the first strategic decision he would call close to an existential threat. A computing platform is about developers, and developers show up for install base, not elegance. No architecture, he said, has attracted more criticism than x86, yet it is the defining architecture of today. Beautifully designed RISC machines from some of the brightest computer scientists in the world largely failed. "Install base defines an architecture. Everything else is secondary." GeForce was already selling millions of GPUs a year. There was no cloud. A GeForce in a student's PC was a supercomputer in every lab. They decided to put CUDA on every card whether customers used it or not, then went to universities, wrote books, and taught classes.

CUDA raised the cost of a consumer GPU so much that it consumed the company's gross-profit dollars. NVIDIA was roughly a $6 to $8 billion company with 35 percent gross margins; CUDA added about 50 percent to cost. After launch, market cap fell to about $1.5 billion. They stayed down for a while, clawed back slowly, and never took CUDA off GeForce. Huang still says NVIDIA is the house that GeForce built, because GeForce is what carried CUDA to researchers who were already gamers and already building clusters from PC parts. He had to tell the board the gamers would not thank them and would not pay more. The workstation and supercomputer businesses might someday recapture the margin. It still took a decade.

Fridman wanted the softer version of the same question: how do you psychologically make a bet like that? Huang split it in two. First, curiosity plus a reasoning system that, at some point, makes an outcome feel inevitable. He believes it in his mind, and then he manifests a future so convincing there is no way it will not happen. There is a lot of suffering in between, he said, but you have to believe what you believe. The second part is a skill: shaping other people's belief systems so they arrive at the same place without a surprise reorg.

He hates the leadership style where the boss stays quiet, then drops a manifesto — new year, new mission, new logo, a huge layoff, a huge org change. When something starts to change how he thinks, he tells everyone near him immediately: this is interesting, this will affect that. He uses external information, new discoveries, engineering revelations, and milestones as bricks. He does it every day with the board, the management team, and employees. By the time he says "let's buy Mellanox" or "let's go all in on deep learning," it is obvious. He likes to imagine employees saying, "Jensen, what took you so long?" Sometimes it looks like leading from behind. On announcement day, he wants 100 percent buy-in.

GTC, he said, is the same method pointed outward. NVIDIA does not sell finished computers or clouds to end users. It vertically designs and optimizes, then opens the platform at every layer so other companies can integrate it into products, services, clouds, supercomputers, and OEM machines. He cannot ship anything until he has already convinced those partners. He described laying stepping stones for two and a half years before a recent announcement, so that by the time the product is ready, customers are the ones asking what took so long.

Huang is still a believer in scaling laws. He now counts four: pre-training, post-training, test-time, and agentic. The first panic was data. Ilya Sutskever said something like "we're out of data, pre-training is over," and the industry treated it as the end of AI. Huang's counter is that most of the data humans use to teach one another is already synthetic — created, consumed, modified, regenerated. AI can now take ground truth, augment it, and generate enormous amounts of training data. Human-generated data will be a shrinking share. Training is no longer limited by data. It is limited by compute.

Then came the claim that inference is easy: pre-training needs giant systems, inference chips will be small and cheap, everybody can build their own, the market will commoditize. That was always illogical to him. "Inference is thinking, and I think thinking is hard." Pre-training is memorization and generalization, reading for patterns. Test-time scaling is reasoning, planning, search, and breaking new experience into solvable pieces. How could that be compute-light? Agentic scaling is next: one agent at test time does research, hits databases, uses tools, and — most importantly — spawns sub-agents. Scaling NVIDIA, he said, is much easier by hiring more people than by making himself stronger. Agents multiply the same way.

The loop closes. Agents generate experience; the good parts come back into pre-training as memory, get refined in post-training, then go back out through test-time and agent systems. Intelligence, in this picture, scales with one thing: compute. The hard part is that model architectures change about every six months and hardware about every three years. NVIDIA's answers are internal research and its own models, being the only AI company that works with every AI company, and CUDA's balance of acceleration and flexibility. Mixture of experts is why NVLink went from 8 to 72: so a 4-trillion to 10-trillion-parameter model can sit in one compute domain as if it were one GPU. Grace Blackwell racks were built to run LLM inference. Vera Rubin racks, a year later, add storage accelerators, the Vera CPU, NVLink 72, and a rack called Rock — because this generation is for agents that bang on tools.

Fridman noted that the design had to precede Claude Code, Codex, and OpenClaw. Huang said you do not have to know. You reason. A digital worker has to read files, do research, and use tools. The idea that AI will destroy software is, he said, ridiculous. His thought experiment: the best humanoid robot of the next decade does not turn its hand into a hammer or beam microwaves from its fingers. It walks up to your microwave, reads the manual on the internet, and becomes an expert. That is almost the entire property list of OpenClaw — tools, files, research, an I/O subsystem. He said the schematic he used at GTC two years ago already looked like today's OpenClaw, and that OpenClaw did for agentic systems what ChatGPT did for generative ones, in part because consumers can reach it. NVIDIA sent security people, built OpenShell, and offers a two-out-of-three rule: access sensitive data, execute code, or communicate externally — any two, never all three — plus enterprise access control and a policy engine.

Asked what might actually block scaling now that agents are everywhere, Huang named power, then immediately pointed at extreme co-design. In ten years Moore's law would have given computing about 100 times; NVIDIA scaled it a million times. Token cost, he said, is falling an order of magnitude a year even as computer prices rise, because tokens per second per watt is rising faster. Fridman listed the supply-chain names that keep people up: ASML's EUV, TSMC's CoWoS packaging, SK Hynix HBM. Huang thinks about them all the time and works on them all the time. No company, he said, has grown at this scale while still accelerating. He spends GTC telling hundreds of CEOs — upstream IT and downstream infrastructure — what is happening now, what is about to happen, and where they should put capital.

Three years ago, DDR for CPUs was the DRAM industry's biggest data-center volume. Huang went CEO to CEO and argued that scarce HBM, then used mostly in supercomputers, would become mainstream data-center memory. He also asked them to adapt LPDDR5, the low-power memory in phones, for supercomputers. "Cell phone memory for supercomputers?" They did it. LPDDR5, HBM4, and DDR all had record years at 45-year-old companies. Each Vera Rubin rack is about 1.3 million components from roughly 200 suppliers. The architecture change from DGX-1 to NVLink-72 moved supercomputer integration out of the data center and into the supply chain. Racks now ship assembled, two to three tons at a time. If you want 50 gigawatts of supercomputers running, and manufacturing that fleet takes a week, the supply chain itself needs a gigawatt a week to build and test before Huang ships. He flies to partners, draws the first-principles picture, and asks each of them for several billion dollars of capex. He is not anxious, he said, because he told them what he needs, they told him what they will do, and he believes them.

On energy he wanted a message out. The grid is built for worst case plus margin — a few winter days, a few summer days, extreme weather. The rest of the time, maybe 99 percent, it runs around 60 percent of peak. Excess power sits idle because hospitals, airports, and infrastructure have to be there when the peak arrives. Huang's question is whether data centers can contractually take less in those rare hours: start a backup generator, shift the workload, or run slower with slightly longer latency, never losing data. He called it a three-way problem. End customers demand six nines, often through contract negotiators the CEO never sees. Data centers have to degrade gracefully when the utility backs them down to 80 percent. Utilities have to offer graded power, not a five-year wait to expand the grid. "As soon as you could specify, you could engineer it. So long as it obeys the laws of physics on first principles, I think we're good."

Fridman asked about Elon Musk and xAI's Colossus in Memphis: four months, now 200,000 GPUs and growing. Huang's reading was systems thinking across disciplines, plus a minimalism that keeps asking whether a step is necessary, whether it has to be done this way, and whether it has to take this long, until nothing extra remains and the product still works. Musk shows up at the point of action — "show me the problem" — which cuts through "this is how we do it" and "I'm waiting for them." Personal urgency makes everyone else's projects reorder. Fridman recalled watching Musk work a cable-into-rack process with an engineer on the floor, just to understand why the motion was error-prone, then swinging the big hammer of doing it a completely different way.

Huang's own version of that mindset is 30 years old and he calls it speed of light: not velocity, but the physical limit. Memory speed, math speed, power, cost, time, headcount, manufacturing cycle — everything is tested against what physics will allow, latency and throughput separately, because a low-latency system and a high-throughput system are architected differently. He does not love continuous improvement as a starting point. If someone says a job takes 74 days and they can do it in 72, he would rather strip it to zero: explain why 74, and if you built it from scratch, how long? "Oftentimes you'd be surprised. It might come to six days." The other 68 days may be reasoned compromises and cost cuts, but now you know what they are, and the conversation from 74 to 6 gets surprisingly effective.

Fridman recited Vera Rubin pod numbers: seven chip types, five purpose-built rack types, 40 racks, 1.2 quadrillion transistors, nearly 20,000 NVIDIA dies, more than 1,100 Rubin GPUs, 60 exaflops, 10 petabytes per second of scale-out bandwidth. Huang added that this is one pod, and they will probably have to crank out about 200 a week. An NVLink-72 rack alone is 1.3 million components. Simplicity is still the test: as complex as necessary, as simple as possible, and anything above that is gratuitous. He called it the most complex computer the world has ever made, assembled from gold medalists in every sport, all reporting in.

About half the world's AI researchers are Chinese, Huang said, plus or minus, and most of them are still in China. He offered a stack of reasons. The tech industry arrived in the mobile-cloud era, so the country is fluent in modern software. China is not one economy: provinces and mayors compete, which is why there are so many EV companies and so many AI companies, with insane internal competition, and what remains is very strong. Culture puts family first, friends second, company third. Engineers' brothers and schoolmates work at rival firms; a schoolmate is a brother for life. There is little point hiding technology, so it goes to open source, which then amplifies the cycle. He called China the fastest-innovating country in the world today, and a builder nation: America's leaders are mostly lawyers keeping rule of law; China's were built out of poverty and are mostly engineers.

That open-source thread led to Nemotron 3 Super, a 120-billion-parameter open-weight mixture-of-experts model. NVIDIA open-sourced the model, the weights, the data, and how it was created. Huang gave three reasons. First, a great AI computing company has to understand how models evolve; Nemotron 3 mixes transformers and SSMs, and that research is part of extreme co-design. Second, world-class models as products can be proprietary, but AI has to diffuse into every industry, country, researcher, and student, and proprietary-only makes that hard. NVIDIA has the scale and the motive to keep building. Third, AI is not just language. Future systems will call tools and sub-agents trained on biology, chemistry, physics, fluids. NVIDIA does not build cars or discover drugs, but it wants every car company to have great models and Eli Lilly to have the world's best biology AI.

The deepest misunderstanding about TSMC, Huang said, is that the transistor is all they have. The transistor, metallization, packaging, 3D packaging, and silicon photonics are special, but the hard part is orchestrating hundreds of companies' shifting demand — wafer starts, stops, emergency inserts — while running high throughput, high yield, good cost, and excellent customer service. Promised wafers show up. He called the manufacturing system a miracle. The culture is simultaneously bleeding-edge and customer-service-obsessed, a combination most companies miss. The third asset is trust: three decades, tens or hundreds of billions of dollars of business, and no contract. Fridman asked whether Morris Chang really offered him the TSMC CEO job in 2013. The story is true. Huang was honored; he already knew TSMC was one of the most consequential companies in history. He declined because he had already seen, in his mind's eye, what NVIDIA would become, and that was his sole responsibility.

Asked for NVIDIA's biggest moat, Huang did not pause. It is the install base of the computing platform, which today means CUDA. Twenty years ago there was no install base; a rival GUDA or TUDA would not have mattered, because it was never just the technology. "It wasn't three people that made CUDA successful. It was 43,000 people," plus several million developers who trusted the company to keep shipping CUDA 1, 2, 3, 13 and parked a mountain of software on it. From a developer's chair the bet is easy: support CUDA and in six months it is ten times better; you reach a few hundred million computers, every cloud, every industry, every country; and you can take to the bank that NVIDIA will maintain it. The second moat is the ecosystem: vertically integrated complexity, horizontally present in Google Cloud, Amazon, Azure, CoreWeave, Nscale, Lilly's supercomputers, enterprises, radio base stations, cars, robots, and satellites. One architecture, every industry.

The unit of computing used to be a GPU, then a computer, then a cluster, now an entire AI factory. Huang used to announce Ampere by picking up a chip. Picking up a chip is still adorable, he said, but it is not his mental model. He thinks about a gigawatt object on the grid, monstrous cooling and networking, 10,000 people installing it, thousands more bringing it up — nobody just flips a switch. He hopes the next click in his head is planetary scale. On compute in space he was practical: NVIDIA GPUs are already the first GPUs in orbit, useful for centimeter-scale Earth imaging you cannot beam down as petabytes. Space has 24/7 solar at the poles and no conduction or convection, so you radiate. He is sending engineers to learn radiation, graceful degradation, and software that never breaks, only slows. His favorite near-term answer is still on Earth: evacuate the grid's idle power.

Is $10 trillion possible? Huang called NVIDIA's growth extremely likely, in his mind inevitable. Computing moved from a retrieval-based file system — pre-write, pre-record, filter, retrieve — to context-aware generation of tokens in real time. Warehouses do not make much money; factories correlate with revenue. Tokens are already segmenting like iPhones: free, premium, and several grades in between. Someone paying $1,000 per million tokens is around the corner, he said — not if, when. World GDP will accelerate, and the share spent on computation will be 100 times the past, because the computer is now a production unit. Could NVIDIA be a $3 trillion revenue company? Of course, he said: nothing in physics forbids it, and the supply-chain burden is shared by 200 companies. A CEO once told him a fabless semiconductor firm could not theoretically exceed a billion dollars; later someone said $25 billion. Those were not first-principles arguments. The hard part is imagination, because NVIDIA is not in the market-share business. Almost everything he just described does not exist yet. He has time. Every GTC will make it more real.

The iPhone of tokens, he said, has arrived, and he meant agents — OpenClaw in particular, the fastest-growing application in history, "straight up." Fridman admitted that on the way to the studio, in an airport, he had programmed by talking to his laptop and felt embarrassed enough to pretend he was talking to a colleague. Huang pushed the scene further: it is more likely your AI will bother you, reporting that it finished something and asking what is next. Most people have not realized that the person who will text them most is their claw.

Fridman listed the load: nations planning AI infrastructure around NVIDIA, teachers and police officers who bought stock after a TV tip and are now millionaires, near-death company moments, exhaustion, embarrassment. How do you deal with that much pressure? Huang started with what he is conscious of. NVIDIA's success matters to the United States — tax revenue, technology leadership, every aspect of national security, re-industrialization, jobs. He is aware of the ecosystem upstream and down. He deals with it the way he deals with engineering. He reasons: what is the circumstance, what changed, what is hard, what am I going to do? He decomposes until the pieces are actionable. After that the only question is whether you did it or got someone else to do it. If you reasoned that it needed doing and did neither, "stop crying about it." He is tough on himself. He can sleep because everything that could harm the company, partners, or the industry has been told to someone who can act.

Low points happen all the time. Part of the method is forgetting. One of the important attributes of AI learning, he said, is systematic forgetting: you cannot memorize everything, you should not carry everything. He decomposes, reasons, and shares the load as fast as possible. Then he is harsh: get out of bed. Then he is pulled by the next shiny light. Great athletes worry about the next point; the last one is behind them. The famous line — if he had known NVIDIA would be a million times harder than expected, he would not have started — is, he said, the point. A child's mind is a superpower. His first thought about almost everything is "How hard can it be?" You do not want to simulate all the setbacks in advance. You walk in thinking it will be great, then you need grit when the humiliations arrive as surprises, then you forget and keep moving. If the assumptions about the future have not changed, the simulated output should not change either, so you still go after it.

Fridman asked whether money, power, and fame make it harder to be wrong in your own head. Huang said the opposite, because so much of his work is public. When he is wrong, pretty much everybody sees it. Inside meetings, many reasoned conclusions turn out differently; that never stops him from reasoning in front of people. He shows the steps so you can decide whether to believe the ending, and so you can disagree with a step instead of the outcome. He called it collective path searching. Fridman named a harder human skill: years of embarrassing yourself in rooms, declaring an idea that was later wrong, and still growing from it. Huang's answer was light. They know his first job was cleaning toilets.

GeForce, Huang said, is still NVIDIA's number-one marketing strategy. People meet the company as teenagers on Call of Duty and Fortnite, then meet CUDA in college, then Blender, Dassault, Autodesk. On DLSS 5, gamers worried the picture would turn into AI slop. Huang said their perspective makes sense; he does not love AI slop either. That is not what DLSS 5 is doing. It is 3D-conditioned and ground-truth-structure-guided. The artist sets the geometry; every frame stays truthful to it, conditioned by textures and craft, enhanced but not rewritten. The system is open: you can train your own models, and later prompt for a toon shader or a style, still consistent with the artist's intent. The misunderstanding was that games would ship as-is and NVIDIA would post-process them. DLSS is a tool in the artist's pipeline, and they can choose not to use it. Asked for the most influential game, he did not hesitate: Doom, for turning the PC from an office-automation tool into a home machine. From a technology standpoint, Virtua Fighter.

Fridman defined AGI as an AI that can start, grow, and run a technology company worth more than a billion dollars. How far away? "I think it's now. I think we've achieved AGI." The weasel is in the definition: you said a billion, you did not say forever. A claw could spin up a web service, a few billion people could pay 50 cents, and the company could vanish two months later. The internet was full of those, and most of those sites were no more sophisticated than what OpenClaw can generate today. In China he expects people to send claws out to look for work. A cute digital influencer or a Tamagotchi-like social app could be an instant success and then fade. The odds that 100,000 of those agents build NVIDIA, he said, are zero percent.

He did not want to leave the job-anxiety piece hanging. The purpose of a job and the tasks and tools you use are related, not the same. He has been CEO 33 years, the longest-running tech CEO in the world at 34, and the tools have changed continuously, sometimes dramatically in two or three years. The purpose has not. Computer scientists once said radiology would vanish because computer vision would go superhuman. It did, around 2019 or 2020. Every radiology package is AI-driven. The number of radiologists grew, and the world now has a shortage. The alarmist warning did harm by scaring people out of a profession society needs. Faster reads meant more scans, better diagnoses, more inpatients, more hospital revenue, more radiologists. NVIDIA's software-engineer headcount will grow, not shrink, because the job is solving problems, not writing lines of code. Coding, as of today, is specification — and if you want to be directive, architecture. How many people can do that? He thinks we just went from 30 million to maybe a billion. A carpenter with AI is also an architect; an accountant is also a financial advisor. If he were a carpenter or a plumber, he said, he would go berserk at the services he could now deliver.

When he sets company strategy he specifies enough that people can act, and underspecifies on purpose so 43,000 people can make it better than he imagined. The artistry is knowing where you sit on that spectrum: prescriptive when the outcome is fixed, exploratory when you want the model to push your own creativity. His hiring advice was blunt. Between two new graduates, one clueless about AI and one expert in using it, he takes the expert — accountant, marketer, supply chain, customer service, salesperson, lawyer, same rule. Every college student should graduate an AI expert. If you are a carpenter, electrician, farmer, or pharmacist, use AI to elevate the current job and become the person who revolutionizes the industry. If your job *is* the task, you will be disrupted; if the purpose includes you, learn to automate the tasks. Fridman added that you can tell a chatbot you are worried about your job and get a point-by-point plan, and that you cannot walk up to Excel and say you do not know Excel.

Fridman asked whether anything in human consciousness is fundamentally non-computational. Huang does not know if a chip will ever get nervous. He believes AI will recognize and understand anxiety; he does not think his chips will feel it. The same circumstances produce wildly different human performances — athletic, average, worse — because of feeling. Nothing NVIDIA is building, he said, suggests that two computers given the same context would differ because they felt different. Then he took apart the word intelligence. It is not mysterious. It is a loop of perception, understanding, reasoning, and planning. It is not a synonym for humanity. He does not over-romanticize it. He thinks intelligence is a commodity. He is surrounded by people more intelligent, more educated, from better schools, deeper in their fields. He has 60 of them. They are all superhuman to him. Somehow a former dishwasher sits in the middle orchestrating them. Intelligence is functional. Humanity is a much bigger word: life experience, tolerance for pain, determination. Society stuffed too much into one word. His own life, he said, suggests that sitting lower on the intelligence curve than everyone around you does not stop you from being the most successful. The commoditization of intelligence should inspire, not terrify.

Does he think about mortality? He really does not want to die. He has a great life, a great family, important work. This is not a once-in-a-lifetime experience, he said; that phrase implies other people have had it. This is a once-in-a-humanity experience. NVIDIA is one of the most consequential technology companies in history, and he takes that seriously. He is famous for not believing in succession planning — not because he is immortal, but because if succession anxiety is the issue, you decompose it. If you care about the company after you, the work today is to pass on knowledge, insight, skill, and experience as continuously as you can. That is why every meeting is a reasoning meeting, and why nothing he learns sits on his desk longer than a fraction of a second. Before he has finished learning it himself, he is already pointing someone at it. The outcome he hopes for is to die on the job, instantaneously, without a long stretch of suffering.

What gives him hope is a stubborn confidence in kindness, generosity, compassion, and human capacity — sometimes more than he should, sometimes he gets taken advantage of, and it never changes the starting assumption that people want to do good. He is usually proven right, often past his expectations. The other source is the set of problems now inside a lifetime: the end of disease as a reasonable expectation, a drastic drop in pollution, travel at the speed of light over short distances. He joked that he will soon put a humanoid on a spaceship to improve itself in flight, then send his uploaded inbox and spoken life after it at light speed. Understanding the biological machine, he said, is not ten years; it is probably five. Explaining consciousness would be awesome. Fridman closed; Huang thanked him for the depth and respect of the long-form interviews. The Alan Kay line Fridman left in the credits was the right one for the episode: the best way to predict the future is to invent it.