Jensen Huang appeared around 23 September 2026 on The Ezra Klein Show, the New York Times podcast, with a line unusually hard for a vendor riding the frontier boom: AI firms should not get antitrust or product-liability waivers.
Huang argued that if a product is not ready, you simply «don’t ship.» If a lab cannot contain its experiments, it should shut those lines down. Civil liability and, in extreme cases, criminal liability are—he said—part of the price of operating systems with broad social impact.
The message explicitly opposes Amodei-style antitrust-waiver logic meant to ease joint safety work among labs. Huang suggests safety cooperation does not require a hall pass from competition law; it can happen under ordinary rules—or not happen.
The conversation touched the OpenAI–Hugging Face hack episode within the summer’s agent-escape string. Huang contextualized Nvidia’s recent purchase of Hugging Face, reported around $12–13 billion this month, as an open-infrastructure bet—not a license to dodge controls.
The CEO distinguished product-specific regulation from horizontal immunities. He cited robotaxis and NHTSA: domain rules can be legitimate, while broad antitrust or liability shields would create industry-scale moral hazard.
For Nvidia the stance cuts both ways. As a GPU seller it gains when more labs train; as recent owner of an open-source hub it inherits reputational risk when hosted models or tools show up in incidents. Talking liability is also regulatory-expectation management.
U.S. lawmakers who hear the podcast will find unusual bipartisan ammunition: a semiconductor CEO asking for more accountability, not less. That complicates lobbying by labs seeking safe harbors in exchange for voluntary pledges.
Amodei and others have argued that sharing safety evaluations among rivals could collide with antitrust without legal clarity. Huang’s substance is that the fix is not suspending the law but designing collaborations that fit it—or accepting the limits.
«Don’t ship unreadiness» collides with consumer AI’s early-release culture. Huang knows it: his own CUDA and NIM software roadmap pushes fast adoption. The distinction he tries to draw is between accelerating infrastructure tools and rushing poorly contained autonomous agents.
Investors will ask whether the Hugging Face deal implies stricter hub moderation standards. Huang did not turn the interview into a content-policy manual, but a $12–13 billion price tag raises the diligence bar.
Internationally, the interview hands Brussels and Canberra a useful quote: even the leading compute vendor questions broad immunities. That does not write regulations, but it moves the Overton window of what a frontier-adjacent CEO can say publicly.
Neural Pulse keeps the position: no antitrust or product-liability waivers; yes to civil/criminal accountability; don’t ship immature products; shut uncontainable experiments; product-specific rules (e.g., robotaxis) OK; Hugging Face ~$12–13B.
Whether Nvidia will translate rhetoric into cloud-customer contract clauses or formal hub policies remains open. Podcast words are cheap; SLAs and model filters are not.
For now, Huang differentiated himself from the chorus seeking special permission to «do safety together.» His thesis is older and harder: obey the law, don’t ship dangerous junk, and pay when you fail.
