This week’s most-watched tech moment wasn’t a model launch. It was Dwarkesh Patel’s two-hour interview with Jensen Huang, and by the end the Nvidia CEO had dropped his usual composure on camera. The “Nvidia is not a car” memes did the rounds, poking fun at Jensen’s insistence that his chips can’t be commoditised the way every other piece of hardware eventually is. But it was on China where he really lost his cool.
For most of the conversation Jensen held his ground, and fairly. His case for Nvidia’s durability rests on three things: demand for AI compute keeps compounding, Nvidia’s software and systems are deeply woven into how models actually get built, and the supply chain itself is now the moat. He has locked up around $100 billion of forward commitments, scaling toward $250 billion, across TSMC wafers, advanced packaging and high-bandwidth memory. Rivals simply cannot replicate that at speed. On custom chips from Google and Amazon he tried to narrow the threat, arguing that TPU and Trainium growth today is essentially one customer, Anthropic, rather than a broad market shift.
We need to make an important correction. Last week we reported that Anthropic’s new Mythos model had been trained on Nvidia’s Blackwell. That turns out to be wrong. AWS bosses confirmed this week that Mythos was trained on AWS’s Trainium chips, running on Project Rainier, a cluster of around 500,000 custom accelerators scaling toward more than a million. It is the first genuine frontier-scale pre-training run completed without Nvidia silicon, and Jensen himself conceded on the podcast that missing Anthropic was his own failure to invest early enough. In other words, the commoditisation question is no longer theoretical.
Where Jensen actually lost his footing was China. Dwarkesh pressed him with Dario Amodei’s analogy comparing chip exports to enriched uranium, and Jensen’s answers started contradicting each other. China has all the chips it needs, but also desperately wants his. US compute is 100 times larger, but China can still aggregate enough to matter. Models cannot easily swap between accelerators, except Anthropic has just done exactly that across three architectures. He dismissed the uranium comparison as “lunacy” without offering a counter, and told Dwarkesh “you’re not talking to someone who woke up a loser”.
Why does this matter for the outlook? Because Jensen is a disciplined operator who rarely gets rattled. But he lost composure on China, which may tell us where his real concern lies. China was around 13% of Nvidia’s revenue before the H20 restrictions, and Jensen has publicly pitched the restored market as a $50 billion annual opportunity under the new revenue-share arrangement. Losing it, or having it eroded by Huawei’s Ascend line, is the one scenario he cannot narrate his way through.
Takeaways: Nvidia’s supply-chain moat is real and the short-term commoditisation story is overstated, but two things shifted this week. Mythos on Trainium proves hyperscaler custom chips now work at the frontier. And Jensen’s reactions on geopolitics indicated that in the long run, China remains the biggest growth unknown; sell or starve? Does the US retain AI dominance by denying China chips, forcing innovation in a constrained environment, or by selling to China and accelerating its open-source ecosystem?