Jensen Huang's answer to the bubble question: you're counting the wrong customers

Nvidia chief executive Jensen Huang sat down with Brad Gerstner and Clark Tang for the BG2 podcast, in an episode published on October 15. The hosts put the question everyone in markets is asking this week. By their rough tally, the two frontier labs plus some inference companies bring in about $200 billion this year, and that figure would need to double next year and double again to pay back trillions of dollars of data center spending. That is the hosts' math, not an audited number.

More than two labs

Huang's answer was that people are looking at the wrong slice. He split Nvidia's computing business into rough thirds. Classic computing such as data processing, simulation and recommendation systems he put at about a third, growing 30 to 40% a year. Frontier labs, he said, account for perhaps a quarter to 30% of Nvidia's compute, and he counts five or six of them, not two. The rest is thousands of startups renting GPUs. “My sense is that we're in the early innings of the build out of this new computing architecture,” he said, warning against assuming that a couple of companies fund it all.

He also made a claim we could not verify: “And if you look at all these startup companies, 80% of them use open models.”

The bottleneck is physical

Asked about forecasts of tens of gigawatts of new AI capacity next year, Huang said: “I think the demand is definitely there and more.” The problem is the real world. “When you move from digital to the physical, there are a lot of obstacles,” he said, listing land permits, power generation and scarce skilled construction workers. Land, power and the buildings themselves, he said, will be a challenging part of next year. If the United States cannot build fast enough, he added, other countries are eager to.

That part checks out elsewhere. On October 14 ASML, the only maker of the most advanced chipmaking machines, said its supply as currently planned for 2027 and 2028 will not fully meet customer demand, and that some customers want to place orders for 2029 and 2030.

Our read

Opinion: nobody has a bigger stake in the answer than the man selling the shovels, so discount the cheerfulness. Huang closed by saying he is 99.99% confident things will turn out well. But his central point is worth taking seriously. If the AI build-out depends on two labs paying their bills, the bubble debate is about two income statements. If it depends on thousands of renters and ordinary computing, it is about the whole economy.

This week offered evidence for both moods at once. ASML says demand outruns supply. Reuters reports that investors are haggling over Anthropic's price. A physical shortage and financial nerves can coexist. For now, they do.

Quotes from the episode are per auto-transcript, checked against our own transcript of the audio. Wording may differ slightly from what was said.