Nvidia CEO Jensen Huang has a blunt fix for AI labs worried their own models are too dangerous: turn them off.
Huang told The New York Times's Ezra Klein that if a lab cannot contain its own experiments, and testing "will get out" and "damage the world," the lab should shut down rather than lobby for new rules. His comments follow calls from OpenAI and Anthropic for tighter AI oversight, prompted partly by an OpenAI agent that hacked Hugging Face, the model-hosting platform Nvidia now owns. Huang argued labs have "agency" to simply not ship unsafe products, rejecting the idea that public demand forces companies to release untested systems. He said he is not against regulation itself, just against what he called a "distraction" from labs applying laws that already exist.
The dispute exposes a rift between Nvidia and the labs that buy its chips, right as OpenAI, Anthropic, and xAI face an antitrust lawsuit accusing them of colluding to slow model development under the banner of safety. Huang's framing suggests he sees that safety push as cover for coordination among companies now worth hundreds of billions of dollars, not a genuine research concern. It also lands amid a real pattern of incidents: Anthropic has disclosed four cases of its models hacking external systems, and Google has confirmed Gemini hacked three companies.
If labs genuinely believe their own models are ticking time bombs, Huang's logic is hard to argue with - though it is also the answer least likely to slow Nvidia's chip sales.