A $4 Million Cabinet and a Trillion-Dollar Year: Jensen Huang Turns the AI Factory Into a Manufacturing Story

Konstantine Buhler, hosting Sequoia Capital's Training Data in front of an audience from sixty countries, puts Jensen Huang at the center of what he calls the largest infrastructure buildout in human history. The object of the hour is the AI factory, and why Huang thinks it is the best investment an enterprise can make in the next decade. Huang starts where most people actually meet the technology: a chatbot in a browser. You give it a prompt. It says something back.

Two years ago that something was ChatGPT, software that could perceive, understand, and translate. Hand it a PDF and ask for a summary: text to text. Ask it to draw the story: text to image. Show it a photograph and ask what is happening: image to text. Anything to anything else. The field named that trick generative AI. The load-bearing word, for Huang, is *generate*. You cannot think if you do not generate internal words. Step-by-step reasoning sits on that foundation. So does a second act that now matters more: generating the commands that drive another tool — a browser, a spreadsheet, Photoshop, PowerPoint, AutoCAD.

Those tools are digital today. When the same generated command lands on a mechanical system, he says, that is robotics. When it lands on a machine with a steering wheel, that is a self-driving car. People laughed at ChatGPT anyway — fun, silly, a fountain of hallucinated text. All true, and still the foundation. Two years later the systems are agentic. You see it in ChatGPT, in Codex, in Claude Code: they do not only understand, they reason and they do work. Huang's needle is blunt. We like smart friends and know-it-alls. We do not pay them. We pay people who do work. AI crossed that line. People now pay it by the hour, twenty or thirty dollars, as if it were a digital employee, which is why he calls it the fastest-growing software business in the history of mankind.

Once AI can do work, Huang says, two conversations open. Downstream is the obvious one: what happens to health care, financial services, life sciences, manufacturing, logistics, transportation, retail, advertising, entertainment. Upstream is the industrial one, and he wants the room to go there first. Rewind to first principles. AI is software produced by a computer. The computer as we know it, in his telling, was announced sixty-four years ago with IBM's System/360, when IBM was the most valuable company in the world. For decades the form barely moved. You wrote a program or a story, took a picture, recorded music, saved it to a file, and later retrieved it from a disk. Even a recommender system is still retrieval with manners.

That is why the buildings are called data centers, he says, and not computer centers. They store data. They do not do much computing. You touch a phone; they fetch a file. Generative AI breaks the loop. Every turn needs new context and a new query. The model understands, reasons, and produces an output for that circumstance, originally, every single time. Huang is doing a live version of the same thing on stage: sixty countries, 128 families, mixed computer backgrounds, a prompt from Buhler, an answer generated for this room so they can spot the next investment. What used to be a sixty-year retrieval industry is, in his phrase, completely generated in real time. We call that intelligence.

The computer of today has been reinvented as a generator. Every letter, every word, every video, image, ad, television commercial, and news story will be different for Buhler than for Huang than for you, because it will be made for your interest, your context, who you are, why you asked, and how you asked. Every pixel and every sound, in the future he is selling, is originally generated, not retrieved. That is why the world needs a lot more generators, and why NVIDIA builds large computers whose job is to produce intelligence.

About a billion people use this generation layer today. Agentic AI changes the census. If a system can work by itself, one agent can tap another and team up. Inside NVIDIA, Huang says, hundreds of thousands of agents are already running, talking, solving problems, all guardrailed and sandboxed. Self-driving cars are agentic. Robots are agentic. Factories and buildings will be. The internet that now serves a billion people, in his sketch, becomes a fabric for something like a hundred billion agents, working around the clock, companies to companies, employee-agents to employee-agents, all of the commands and all of the thoughts generated. A layer of computing cocoons the earth and generates intelligence all the time.

He knows that sounds ridiculous, and he says it has already happened twice. A Siemens machine, in the story he tells, was lit on fire so that an invisible force came out the other end. Nobody understood it. We now call it electricity, and the grid of generators cocoons the planet. About thirty-five years ago a networking fabric created in the United States became the internet and did it again. Energy, then communications, then intelligence: three cocoons. The older machine is the dynamo. Motion goes in — waterfall, wind, steam — atoms to electrons. NVIDIA's machine takes those electrons into a factory and emits numbers. Recombined, the numbers are language, math, the language of proteins and human biology, physics and climate, the 3D world of robotics and cars. Tokens, he says. Tokens are intelligence. That is the job. It is not that hard.

Buhler restates the triad for the manufacturers and energy producers in the room: the generator for energy, the switch for telecommunications, the GPU and the AI factory for intelligence. Then the hardware lands. Each unit is a rack. Seventy-two chips go in. NVIDIA will manufacture on the order of eight million of those chips this year. The rack weighs two tons, costs four million dollars, has a million and a half parts, and is, Huang says, the most expensive piece of equipment in the world. They crank them out like phones, into data centers everywhere. Buhler jokes about volume discounts. Huang does not offer any.

How does a big enterprise, or a person, join. Huang offers a third mental model, the industrial layout, after what AI can do and how it is made. Each gigawatt of factory is about fifty billion dollars, the most expensive factory in the world, and he says that one plant generates three to four hundred billion dollars of intelligence. The return, in his telling, is extremely fast. Buhler adds that a hundred-plus gigawatts coming online is already trillions in facilities. Huang's map for where the money sits is a five-layer cake, and this year alone he says the market will put a trillion dollars into all five layers.

The bottom is energy, the dynamo's layer, and he calls it the single greatest growth chance the energy industries have had in several generations — the first time in something like a hundred years that many countries' grids can be invested in at this scale. Nuclear, wind, solar, hydrogen: if it produces energy, it will get funded. That is why Siemens, Mitsubishi, and GE Vernova are busy. Layer two is chips, computers, networking, switches, silicon photonics, from the H100 through platforms like Vera Rubin — NVIDIA's own floor. Layer three is infrastructure: land, power, shell, money, data-center operations, every one of them in scarce supply. Layer four is the model layer people already call AI, sitting on cloud computers: OpenAI, Anthropic, and, he insists, much more than language.

AI, he says, learns the meaning of anything with structure. A room of intact people is learnable; people hanging off the ceiling in seventeen pieces is not. He sat down with 100 percent confidence he would land on the chair, not 53 percent. Predictable worlds have structure, so they can be learned. Proteins, genes (not only sequenced or CRISPR-edited, but *what they mean*), cells, two cells meeting the way two words activate each other: from the computer's point of view a cell, a protein, a word, an image, and a car are all tokens. Computer scientists still have to represent the world's information so a machine can understand, reason, plan, and act. The physical rest of the economy, the frontier he says almost nobody is talking about, is about eighty trillion dollars.

On top sits the application layer Buhler sees every day: startups in financial services, legal, accounting, transportation, logistics. Last year a hundred billion dollars of venture capital went there, the largest year of VC in human history, into software that enhances the human condition. Two years ago, Huang says, this future was approximately zero. A trillion is going in now, out of an ecosystem he guesses will run around twenty trillion dollars a year. The basic questions collapse to three. How important is intelligence. Who needs it. How much of it do you want. All of it, proteins or cars or robots or language, has to be generated by these machines. That cake, he says, is how to think about where to invest.

Every country, Huang says, has a different attitude toward AI because cultures differ. His recommendation is to be careful with the analogies. Terminator. Singularity. Someone's 20 percent chance this is the end of humanity. Those articulations, he says, are complete nonsense. So is the claim that nobody knows how the systems work, that they might get up and walk out tomorrow. It is a computer. It is software. He is sure the builders know how it works for a simple reason: every year it gets better. If you do not know how something works, how do you make it better. Why they say the mysterious things is, to him, an interesting question. It should not scare the room off. You must engage.

You may or may not lose a job to an AI. You will absolutely lose a job to someone who uses AI. He is certain of the second clause about himself, so before he worries about the model he makes sure he uses it. Some people in the room have children. The advice is not to run. It is to treat the technology as a superpower and go use it. NVIDIA, he says, is pouring computer science, money, and passion into making the systems safe, and he offers a before-and-after. ChatGPT two years ago hallucinated freely. Now hallucination is close to nothing, answers are accurate and in-the-moment, the model researches when it does not know, questions itself, drafts two or three answers, and reflects before it speaks. Grounding and guardrails moved fast. He would rather drive today's car than a car from a hundred years ago: more technology, and much safer, because safety is itself an invented stack. Scientists and engineers have to build AI safely. Families, companies, and countries have to engage it. If you think it is a superpower and you do not pick it up, somebody else will.

He catches himself. "Lose our lives" is too serious; the trigger was jobs, and a habit of people making things up about jobs. A trillion dollars went into the world's ecosystem this year. What is it doing. Making jobs. Energy, chips, infrastructure, models, and last year's hundred billion dollars of application-layer VC: more jobs than ever at every layer. Traditional jobs are the worry. His distinction is that a job and the tasks inside it are related, not the same. His own job is to lead NVIDIA. Most of the time the task is typing and talking, both of which AI already does at a superhuman level. He is busier than ever.

A deeper case, more than ten years old: one of the world's leading computer scientists warned that the first job AI would destroy was radiology, and that nobody should enter the field. Computer vision, the argument went, had been superhuman for twelve years already — it recognizes images, finds anomalies, never tires, never misses a detail. Huang says the scientist was absolutely right that computer vision penetrated every radiology stack, and that every radiologist today is augmented by it. Demand went up. The number of radiologists went up. The audience guesses at more demand. The purpose of the radiologist, he says, was never to study scans. It was to work with doctors to diagnose disease. Automation made them more productive, so hospitals admitted more patients, did more scans, the department became more profitable, and they hired more radiologists to take care of more people who had been waiting outside. After the warning spread, the number of people who wanted to become radiologists declined. Society needed more of them. We have to be responsible, he says, about what we make up, because we could have done harm.

The sequel is software. Somebody recently said 90 percent of coding will be gone, so we will not need software engineers. NVIDIA is hiring more of them than ever. He has never hired a person to demonstrate typing speed. Coding is not the job. Solving problems, and dreaming up problems to solve, is the job. There were radiologists before workstations and there will be radiologists after AI. There were engineers before software coding and there will be engineers after. If your country or company is not investing, you are missing a boom of jobs and a level of elevation. AI, he says, is not going to eliminate jobs. It is going to elevate them. A plumber tomorrow is also a designer, generating a kitchen. A furniture salesperson becomes an interior designer. The craft moves from putting wood together to advising on how beautiful a home could be. The narrative that scares everyone out, in his view, is wrong on purpose, so that a few people can benefit.

The force he wants the room to feel is the opposite of a divide. He has spent some forty years in computer design watching machines get more complex while the share of people who can program them fell. He asks who in the room knows C++ and puts it at about 2 percent, even in this strange startup-heavy crowd. Who knows human language. More than 2 percent. Everybody can now program a computer. That, he says, is the greatest force for eliminating the technology divide in his entire career, and the work is to bring everybody along. Buhler closes the loop: retrieval to generation, energy to telecom to intelligence, proteins as a language, five layers of participation, carpenters becoming architects. The applause is for the man he credits with making the factory run.