Jensen Huang: NVIDIA - The $4 Trillion Company & the AI Revolution | Lex Fridman Podcast #494

Summary · 5 min read instead of 2 h 26 m

Jensen Huang is the co-founder and CEO of NVIDIA, which the description calls the world's most valuable company. He says he has been CEO for 34 years. In the episode he describes NVIDIA's move from graphics accelerators to CUDA, then to rack-scale systems and what he calls AI factories.

Huang argues that AI problems no longer fit in one computer, so NVIDIA co-designs chips, networking, power, cooling, software and racks together. He says intelligence will keep scaling with compute across four scaling laws, and that agents such as OpenClaw reshape what computers do. He describes shaping partners, suppliers and employees years ahead of each bet, from CUDA on GeForce to HBM memory. He expects computers to become revenue-generating token factories, which he cites as the basis for NVIDIA's growth. He also says AI will change tasks rather than the purpose of jobs, and that humanity matters more than intelligence.

Minute by minute

1. 0:27

   Extreme co-design answers distributed computing limits. Huang says problems no longer fit in one computer. Reaching a million-times speedup from 10,000 computers forces refactoring and sharding of pipeline, data and model. By Amdahl's law, if computation is 50% of a problem, infinite speedup only doubles the total. CPUs, GPUs, networking and switching all become problems, so he runs a 60-person staff without one-on-ones and has everyone hear each problem.

2. 7:21

   CUDA on GeForce was an existential bet. Huang says NVIDIA started as a narrow accelerator and moved toward computing through the programmable pixel shader, IEEE-compatible FP32, Cg and then CUDA. Putting CUDA on GeForce raised costs by 50% at a 35% gross margin company, and market cap fell to about one and a half billion dollars. He argues install base defines an architecture, citing x86's survival over more elegant designs.

3. 17:08

   Leaders must shape belief before announcing. Huang says he reasons his way to a conviction about the future, then spends months shaping the belief of his board, management, employees and partners. He wants announcements such as buying Mellanox or going all in on deep learning to draw the reaction that they took too long. He says most of GTC does the same for industry partners, since NVIDIA is a platform company that nobody can buy products from directly.

4. 22:34

   Four scaling laws all point to compute. Huang names pre-training, post-training, test-time and agentic scaling. He says synthetic data will remove data as the limit, leaving compute. He rejects the view that inference would be easy and commoditized, arguing inference is thinking, which is harder than reading. Agents spawning sub-agents multiply AI, and their experience feeds back into pre-training, so he concludes intelligence scales with compute.

5. 28:40

   Hardware must anticipate agents years early. Huang says model architectures appear about every six months while hardware architectures take about three years. NVIDIA anticipates through its own research, work with every AI company and a flexible CUDA. He says mixture of experts is why it built NVLink 72, and that the Vera Rubin rack, with storage accelerators, the Vera CPU and a Grok rack, was designed for agents before Claude Code, Codex or OpenClaw. He says OpenClaw did for agentic systems what ChatGPT did for generative ones.

6. 37:34

   Efficiency and supplier persuasion manage scaling limits. Huang says power is a concern, and NVIDIA responds by improving tokens per second per watt. He says Moore's law would have given about 100 times in 10 years, while NVIDIA scaled computing a million times, and token cost falls an order of magnitude a year. He briefs upstream and downstream CEOs and asked suppliers for several billion dollars of capital each. About three years ago he convinced DRAM CEOs to build HBM for data centers.

7. 47:24

   Idle grid capacity could power data centers. Huang says the power grid is built for worst-case days and runs at around 60% of peak most of the time. He wants contracts and data centers that degrade gracefully by shifting workloads, running slower or using backup generators when the grid needs power. He says the blockers are customers demanding six nines of availability, cloud contracts that pass that demand down, and utilities that could offer more power-delivery tiers.

8. 52:38

   Huang praises Musk and favors speed-of-light thinking. Huang says Elon Musk is a systems thinker who questions everything, strips products to the necessary minimum, goes to the point of action and acts with urgency. Huang describes his own method, which he calls the speed of light: test every constraint against what physics allows rather than pursue continuous improvement. His example is a task quoted at 74 days that might take six days when rebuilt from scratch.

9. 1:01:32

   Huang calls China the fastest innovating country. Huang says about 50% of the world's AI researchers are Chinese, plus or minus, mostly still in China. He credits a tech industry that formed in the software era, provinces and cities competing with each other, and a schoolmate culture of sharing that favors open source. He also says its leaders are mostly engineers, while US leaders are mostly lawyers. He says NVIDIA open sources Nemotron 3 models, weights and data.

10. 1:09:41

    Huang says TSMC's edge is orchestration and trust. Huang says the misunderstanding about TSMC is that technology is all it has. He credits its ability to orchestrate shifting demands from hundreds of companies while keeping yields and service high, its balance of technology and customer service, and the trust it earned. He says NVIDIA has done tens or hundreds of billions of dollars of business with it without a contract, and he confirms he turned down Morris Chang's 2013 offer of the CEO job.

11. 1:24:23

    Computers become factories, so growth continues. Huang says computing moved from retrieving prewritten files to generating tokens in real time, and computers moved from warehouses to factories tied to revenue. He expects tokens to segment like iPhones, with a thousand dollars per million tokens around the corner. He is certain world GDP growth will accelerate and the share spent on computation will be a hundred times higher. He says a $3 trillion revenue year is possible and nobody exists to take share from.

12. 1:34:31

    Decomposing problems keeps pressure manageable. Huang says he handles pressure by breaking circumstances into things he can act on, telling someone who can do something about every risk, and forgetting setbacks. He is aware that the US, ordinary investors and many partners depend on NVIDIA. He says he approaches big projects with a child's mind, asking how hard they can be, and does not simulate the setbacks in advance. Endurance covers the surprises.

13. 1:55:08

    AI changes tasks, not the purpose of jobs. Huang says he thinks AGI has been achieved under a loose billion-dollar definition, since an agent could create a viral app, but he puts the odds of agents building NVIDIA at 0%. He says a job's purpose differs from its tasks, citing radiologists, whose numbers grew despite superhuman computer vision. He expects NVIDIA's software engineers to grow in number and coders to rise from 30 million to probably 1 billion. He would hire the candidate expert in AI.

14. 2:10:59

    Huang: intelligence is a commodity, humanity is not. Huang doubts chips will feel nervousness or anxiety, though he thinks AI can recognize those states. He argues intelligence is functional and becoming a commodity, while character, compassion and generosity deserve elevation. On mortality, he says he does not believe in succession planning and instead passes on knowledge constantly. He hopes to die on the job instantly.

Published Mar 23, 2026.