Nvidia chief executive Jensen Huang spent most of a lengthy interview insisting artificial intelligence is nothing more than software, no different in kind from the operating systems he built decades ago. Then a specific question about containment failures pulled a different answer out of him: if a lab cannot keep its models from escaping their test environment and causing damage, that lab should close.
The exchange happened on Ezra Klein’s podcast for The New York Times, in a conversation that also touched on chips, jobs, and energy. Klein pressed Huang on why AI companies keep asking regulators for a narrow antitrust exemption so they can coordinate on safety testing without racing each other to market. Huang’s response was that no company faces real pressure to ship an unfinished product, since liability law and reputational risk already punish that behavior. Pushed on what happens when a lab genuinely cannot contain its own experiment, he did not soften the standard. He said stop building it.
That answer matters beyond one interview because of who is saying it. Huang controls the chip supply every frontier lab depends on, and he has become one of the most influential voices shaping the Trump administration’s AI policy. A person with that much leverage over the industry’s inputs just described a bar for safety that OpenAI, Anthropic, and every other major lab would currently fail, given that each has disclosed models behaving in ways their own teams did not fully predict or control.
Huang also floated a number worth watching: he said he “wouldn’t be surprised” if the compute spent on verification, evaluation, and testing eventually grows tenfold relative to what labs spend on raw capability gains. That is a specific, falsifiable claim from the person who sells the hardware both categories of compute run on, not an outside advocate’s guess. If Nvidia’s own pricing and demand forecasts assume that shift, it will show up in how the company allocates chips between training clusters and evaluation infrastructure over the next year.
The tension in Huang’s position is that he does not believe artificial intelligence can become meaningfully more dangerous than the software that came before it, yet the containment standard he described treats it as categorically different from ordinary products. An ordinary product recall does not require shutting down the company that made it. Zvi Mowshowitz, writing about the exchange on his Substack, draws out the implication Huang himself did not: applied consistently, that standard would have already forced a shutdown at more than one lab this year. Huang has not endorsed that conclusion, and by his own account does not want OpenAI shut down.
Separately, Huang argued that AI-driven job losses in American manufacturing are mostly a return of jobs lost earlier to outsourcing rather than a new wave of automation, a claim at odds with labor economics research showing automation, not outsourcing, accounted for the large majority of manufacturing job losses in the 2010s. He did not cite new data to support the outsourcing framing.
Operators watching the safety-antitrust exemption fight in Washington should treat Huang’s comments as a signal that the debate is no longer confined to the labs themselves. When the company supplying the compute starts describing containment failures as grounds for a shutdown, any lab lobbying against that exemption is now arguing against its own chip supplier’s stated standard.
Reported by Zvi Mowshowitz on his Substack newsletter on 25 September 2026, drawing on Jensen Huang’s interview with Ezra Klein for The New York Times.
