Jensen Huang | Does AI Revenue Justify Capex, '27 AI Bottlenecks, Singularity, Open vs Closed & more

Summary · 5 min read instead of 1 h 40 m

Huang leads NVIDIA, which he says started out building a graphics chip for gamers and now supplies computing infrastructure to every cloud. The episode was recorded days after he received the National Medal of Science and attended the White House summit on the accord on superintelligence.

Huang argues that reasoning agents need far more compute than earlier models and that the cost of intelligence will keep falling while demand rises. He says safety is an engineering problem solved through containment and out-of-band monitoring, and that it has to accelerate with capability. On the capex question, he says AI labs account for only about a quarter to 30% of NVIDIA's business, with the rest in classical computing and startups renting capacity. He names land, power and shell as the main 2027 constraint and expects NVIDIA to gain share because it delivers the most tokens per watt. He says he is 99.99% confident in the outcome of the decade ahead.

Minute by minute

1. 8:17

   Reasoning agents multiply compute demand. Huang says compute needs are about a billion times higher than last year because step-by-step reasoning, planning and pruning take far more inference than recalling memorized answers. He adds that better results lead companies to give AI more jobs, which compounds demand. He credits last year's gains to working memory and tool use, with coding first and design work next.

2. 15:55

   The cost of intelligence will fall fast. Huang says the Navier-Stokes result built on earlier research and needed vast exploration. He expects intelligence costs to fall each year like Moore's Law, and wants a drop of a factor of two or 20 per year. Gerstner puts the compute cost of that result at about $4 million. Huang says training the original GPT would fall from about $20 million to roughly $100 over 10 years, even as models grow to 3, 10 and 30 trillion parameters.

3. 24:31

   Physical AI is already a large business. Huang says NVIDIA did $10 billion in physical AI in the first six months of this year, after its first $10 billion year and $5 billion the year before. He lists medical imaging, radar, self-driving cars, robots and manufacturing. He says a physical AI platform needs three computers: one for training, one running a digital twin such as Omniverse, and one inside the robot or car.

4. 30:39

   Safety is solved by engineering, not perfect alignment. Huang says alignment will be a forever problem, so safety cannot depend on perfecting it. He proposes starting every AI with zero trust and minimum rights, then expanding access as it is tested. He also wants monitoring that runs out of band, so the system behaves the same when watched, and can escalate or shut the system down. He says NVIDIA's OpenShell has over 100 companies joining.

5. 36:39

   Safety and capability accelerate together. Huang compares safety and security to a car's airbags and brakes: they are part of the technology, not separate from it. He says accelerating AI means accelerating the whole package, especially safety. He describes the room at the White House as unified on this. He defines America winning as every industry and every citizen winning, not only the technology sector.

6. 42:52

   The three-legged framework needs speed over perfection. Gerstner describes three legs: Commerce Department review of frontier models, the White House accord with audits and board oversight, and the SI force led by Jay Clayton. Huang says it is better to act quickly than argue about perfection, and that the measures have real teeth. He asks critics to learn about the problem and the solution, and to be constructive.

7. 46:54

   Land, power and shell limit 2027 build-out. Gerstner cites SemiAnalysis, which expects 43 gigawatts of new AI compute next year. Huang says demand is there and more, but permitting, power, skilled construction labor and moratoriums are obstacles. He says NVIDIA is the most resilient because it sells to every cloud and region. He says content per gigawatt rose from about $18 billion to over $40 billion, and its share of CSP capex is about 40%.

8. 55:14

   SpaceX and data centers in space. Huang says Elon Musk's operation moved from NVIDIA's enterprise category to its cloud provider category, because he trains, deploys and rents compute. He says NVIDIA and SpaceX engineers are working on a new server architecture for space. Huang argues that NVIDIA's performance per unit of energy makes launch cost a small share of the return. He leaves any 500 megawatt timeline for Musk to give.

9. 1:00:30

   AI labs are a minority of NVIDIA demand. Gerstner says lab revenue of about $200 billion would need to double twice to pay back the capex. Huang splits NVIDIA's business into three parts: classical computing growing 30% to 40% a year, labs at about a quarter to 30%, and startups renting capacity. He says he sees five or six labs, not two. He adds that 80% of startups use open models, often Nemotron, to build domain-specific systems.

10. 1:08:17

    Build-out does not hinge on two labs. Huang says the infrastructure build-out will reach between $3 trillion and $4 trillion a year by the end of the decade, with about $1.5 trillion of it compute. He cites a half trillion before AI and says the end of Moore's Law raises costs. He says cost per gigawatt rose from $5 billion to $50 billion. He says it is unnecessary to extrapolate lab revenue to multiple trillions and that too much pressure sits on two labs.

11. 1:13:21

    Orders keep growing while supply stays tight. Huang says NVIDIA already held giant order books when it gave its forward numbers, and that he allocated customers less than they wanted. He says new orders arrive almost weekly, and cites a 2 gigawatt Australia deal worth about $80 billion. He names land, power and shell, substrates and silicon photonics as constraints. He expects the supply gap to take a few years to close.

12. 1:17:35

    Consensus lagged and the moat is widening. Gerstner shows a chart where consensus for NVIDIA moved up 100% to 200% in a year, while the stock rose only 20%. Huang says the competitive moat is increasing. He says only NVIDIA can credibly take a $50 billion or $100 billion order sight unseen. He says general-purpose architecture gives fungibility across workloads, models and customers, and that specialization is its enemy.

13. 1:26:33

    Software licensing and revenue share open new growth. Huang describes two new business models beyond hardware sales. One is software licensing for enterprises, which he illustrates with a couple of thousand dollars per GPU per year. The other is revenue share with clouds that rely on NVIDIA for offtake and software support. He says investments will shrink as a share of cash generation. He would not be surprised if annual infrastructure spending reached around $10 trillion.

14. 1:35:22

    Huang is 99.99% confident in the decade. Gerstner asks whether Huang is more confident than 12 months ago that the world will be better in five years. Huang says he is 99.99% confident, because many people care about a good outcome and are working on it. He says worries should be voiced but predicting the end of the world underestimates human capacity. He says that stopping today means losing.

Published Oct 15, 2026.