Nvidia CEO Jensen Huang sat down with Dwarkesh Patel for a conversation posted Wednesday, and spent a good chunk of it explaining why almost every competitor to his company's GPUs owes its existence to a single customer: Anthropic. The interview is 103 minutes long, frequently tense, and in places genuinely revealing about who actually has pricing power in AI compute.
The TPU concession
The quote that will get clipped: "Without Anthropic, why would there be any TPU growth at all? It's 100% Anthropic. Without Anthropic, why would there be Trainium growth at all?" Huang said it twice, because Dwarkesh pushed, and because it's the kind of thing a CEO in his position usually lets a comms team soften.
He's not wrong, and that's the uncomfortable part. Google and Amazon invested billions in Anthropic early, structured the deals so Anthropic would consume their custom silicon, and the ASIC business has been riding that single contract ever since. Huang says he missed it. VCs, he figured, would fund the frontier labs. They couldn't, not at the scale Anthropic and OpenAI needed, and the hyperscalers filled the gap with compute credits instead of cash. Nvidia has since poured a reported $10 billion into Anthropic and up to $30 billion into OpenAI, at valuations that are now, let's say, not the bargain they would have been three years ago.
The China argument, which got ugly
The longest exchange, over thirty minutes, was about whether the US should keep selling chips to China. Dwarkesh came in hard, using Anthropic's Mythos disclosure as his wedge: if a frontier model can autonomously find thousands of zero-day vulnerabilities, including a 27-year-old bug in OpenBSD, then handing Chinese labs more compute is handing them offensive cyber capability. That's the steelman for export controls, and it's a decent one.
Huang's counter was that the premise is wrong. China, he argued, already manufactures over 60% of the world's mainstream chips, employs roughly half of the world's AI researchers, and sits on enough surplus energy to run ghost data centers that are currently just waiting for silicon. The restrictions haven't slowed China; they've forced the country to build Huawei and SMIC into something more resilient than they would have been otherwise. Peter Wildeford pointed out on X that the "ghost datacenters, fully powered" line is arguably also an argument for controls, which tells you how cleanly this debate splits.
Bloomberg ran with the framing that Mythos itself proves the US needs dialogue with China, which is Huang's line, not Dwarkesh's. You can decide which you buy.
Benchmark trash talk
Huang was less diplomatic about TPU and Trainium performance claims. He specifically called out Amazon's claim of a 40% cost advantage on Trainium and said neither Google nor Amazon has submitted results to public benchmarks like InferenceMAX or MLPerf. "I would love to hear them demonstrate the cost advantage of TPUs. It makes no sense in my mind. It makes absolutely zero sense."
This is the kind of thing CEOs say when they're right on the numbers and annoyed that nobody has to prove it. Or it's the kind of thing CEOs say when a benchmark would be inconvenient for them too. Hard to tell from outside. But the absence of public head-to-head data on the highest-volume AI workloads, three years into this cycle, is itself a data point.
What to watch
Nvidia reports fiscal Q1 results on May 28. Anthropic's next compute deal, whenever it lands, will tell you more about the ASIC trajectory than any podcast. And the Trump administration's position on H20 and successor chip exports to China is still in flux. Huang is lobbying openly for looser rules. Whether Mythos-class capabilities tilt the other way is now a policy question, not a technical one.
