Jensen Huang on AI Factories, Digital Workers and Why the Build-Out Has Barely Begun

Asked what insight gave him an edge at 30, Huang describes the early 1990s, when Moore's Law, CPUs and microprocessors absorbed nearly all of Silicon Valley's investment. His team noticed that general-purpose technologies are rarely excellent at very hard problems. They also believed transistor shrinking would eventually hit limits, so a more domain-specific accelerator could supplement CPUs on problems nearly infinite in scale.

The difficulty, he explains, was a chicken-and-egg problem: a new computing platform needs a large market, and no such market existed without the architecture. Nvidia therefore had to invent the technology and the market together. He recalls pitching Sequoia's Don Valentine, who questioned the killer app, and Huang pointed to Electronic Arts and 3D games, which Nvidia helped turn into one of the world's largest entertainment industries.

Huang sees CUDA as more than a technical generalization of the GPU. It was also new products, go-to-market strategy and an ecosystem flywheel, all invented from scratch. He notes that outside ARM and x86 almost no computing platform is used by nearly everyone, and that building one took Nvidia almost 30 years.

Huang credits two serendipitous moments and one sharp observation for Nvidia's role in the AI revolution. He had been frustrated by how brittle computer vision was, and meanwhile his "CUDA everywhere" strategy had him visiting universities. That put Nvidia in contact with researchers like Geoffrey Hinton, Andrew Ng and Yann LeCun, who were all working on vision ahead of the ImageNet contest.

Where Nvidia went further, he says, was asking what else the breakthrough could do. Deep networks can be trained end to end with backpropagation, so he concluded they were universal function approximators, and with state added through RNNs, LSTMs and eventually transformers, few problems would be out of reach. The company then reasoned about where deep learning might be in 10 or 20 years and decided every chip, system and software layer could be reinvented.

In 2016 Nvidia built the DGX-1, which Huang describes as a computer unlike anything the world had seen. At the GTC launch the audience had no idea what he was talking about, but Elon Musk asked for one, and Huang delivered it to a nonprofit in San Francisco that turned out to be OpenAI. He says every model since has been built on Nvidia.

Huang says people picture small GPUs, but a modern Nvidia GPU is rack-scale: two tons, 120,000 watts and about $3 million. The first AI factory, at $300,000 per node, was not very successful, and he concluded it was not big enough, so the bigger second version took off. A one-gigawatt AI factory, he says, runs to about $50 billion.

The company's pace, in his account, comes from designing the whole infrastructure at once, down to the networking, switches, CPUs and GPUs, on one software stack. Each year's design stays software-compatible, which he compares to how Windows compatibility let PCs move fast. By co-designing algorithms, software and hardware he says Nvidia breaks past Moore's Law and delivers roughly 10 times more performance each generation.

The economics are the reason, he argues. A one-gigawatt site cannot draw more power, so a chip with three times the performance per watt lets the owner earn three times the revenue, which is why he calls it a factory. He describes the platform as CPUs, GPUs, network processors and three kinds of switches, with scale-up turning a rack into one computer and scale-out linking many racks and even separate data centers.

Addressing comparisons to the year 2000, Huang notes that most dot-com companies were unprofitable and the entire internet industry was perhaps $20 or $30 billion. Today, he says, AI is not only new labs like OpenAI and Anthropic: search, ads, news and user-generated content at Google, Amazon and Meta are already powered by AI recommender systems. Those hundreds of billions of dollars in revenue must move from classical CPUs and machine learning to deep learning, a transition worth hundreds of billions on its own.

The second layer is the model makers, which he likens to wafer makers and which are also building AI factories. The third is AI-native application companies such as Harvey, OpenEvidence and Cursor. Buhler's Meta example is that after Apple removed attribution data in Q4 2022, Meta rebuilt it with AI on Nvidia GPUs, and its market value later rose more than a trillion dollars above the low.

Huang calls the recommender system the largest software ecosystem in the world, and says moving it to AI will need a mountain of GPUs. He expects quantitative trading to shift from human-engineered features to AI as well, which he says Citadel Securities has pioneered for 20-some years. The ROI from this earlier wave, he says, is already visible in trillions of dollars of market cap.

Huang argues that agentic AI addresses, for the first time, an industry that software never could: labor. He says Nvidia already uses Cursor across its software engineers and chip designers, and expects AI coding alone to be a couple of trillion-dollar opportunity, with digital nurses, accountants, lawyers and marketers to follow. Together with physical AI, he sizes these opportunities at about $100 trillion of the world economy.

Future enterprise workforces, he predicts, will mix humans and digital humans that companies license, hire or grow in-house. Nvidia home-grows many AIs to protect proprietary data, and he expects it to get easier for others to do the same. Unlike pre-compiled software, AI must think continuously on your context, which is why it needs AI factories in the cloud, on-premises and around the world.

On robotics, Huang offers a thought experiment: if AI can generate a video of someone opening a bottle, why can't it drive a robot to do so? Because we can embody tools as extensions of ourselves, AI will embody cars, arms, humanoids and surgical robots. Seeing one working example, such as an AI coder or a robotaxi, he says, means the rest is engineering.

Robotaxis are already here, Huang says, and their ability to generalize from city to city is improving quickly. The industry moved from human-engineered features to deep learning to largely end-to-end, multimodal models. He argues the model for a self-driving car and for a humanoid is highly similar, only in different embodiments, just as he can both drive a car and cut a steak.

Building this future takes three computers: an AI training computer, a simulation computer for a virtual world that obeys physics, and an onboard computer where the robot's brain runs. In the virtual world, which resembles a video game, the AI can try trillions of iterations before entering the real one, and Nvidia's Omniverse keeps the simulation-to-real gap low. Nvidia offers all three computers and works with nearly every robotics and self-driving company.

Beyond robotics, Huang says that if models can understand video and the virtual world, they can understand structured things like proteins and cells. He cites a partnership with Arc and Evo 2, a foundation model for cells that lets you talk to a cell like a chatbot. He also names AI for 5G and 6G, and CUDA Q, which pairs quantum computers with GPUs and could pull in the quantum timeline by about a decade.

On sovereign AI, Huang says no country can afford to outsource all of its data and import its intelligence back. Countries can buy and import some capability, but he would not give up producing national intelligence, helped by plentiful open source. He expects each country to import some, buy some and build some, and mentions activity in the UK, France, Italy, Spain, Germany, Japan and Korea, including Mistral, Nscale and Nebius.

On exports to China, Huang says policymakers want America to win and that Nvidia is a proud American company, but argues that winning developers creates future platforms. He notes China has about 50 percent of the world's AI researchers and calls it a mistake to keep them from building on American technology. He advocates a nuanced, changing strategy that keeps the US ahead while the world builds on its stack.

Huang says Nvidia went from 95 percent market share in China to zero, and that all of its forecasts assume zero, so any change would be a bonus. He calls China the second largest computer market and thinks staying out is a mistake. On security, he expects AI security to resemble cybersecurity, with shared intelligence across the community and every AI surrounded by many protector AIs.

Huang contrasts old retrieval-based computing, in which someone writes content in advance and you click through, with generative computing. Perplexity generates every answer it shows, and video models like Sora generate every pixel. He points out that the conversation itself was generated in real time rather than retrieved from his office, and argues that the future computer will be more like a collaborator, such as an artist or storyteller.

Given that shift, he says the world has built only a few hundred billion dollars of infrastructure for what will likely be trillions of dollars each year. In the lightning round, his underrated KPI is throughput per unit of energy, because token generation rate per watt determines a customer's revenue. He also singles out the roughly 350 libraries on top of CUDA, such as cuDNN and cuLitho, as Nvidia's treasure trove.

Asked what a CIO with $10 billion should do, Huang says to experiment right away with building your own AI. Companies onboard employees into their culture and knowledge, and must now onboard AI employees through fine-tuning, skills and evaluation. He tells his own CIO that IT will become the HR department for digital employees working alongside biological ones.