Nvidia's Huang: US Must Work With China on AI Safety

Jensen Huang says the country building most of the world's AI compute cannot treat AI safety as a domestic problem. At a media briefing in London on Wednesday evening, the Nvidia CEO told CNBC that his company "absolutely must" work with Chinese AI companies to monitor the risks of a technology he called "potentially dangerous."

The position is striking from the executive whose chips sit under most frontier models, days before Xi Jinping visits Washington.

What Huang actually said

"We absolutely should work with each other so that we advance the technology safely," Huang told CNBC. The mechanism he wants is dialogue, not treaty text: he said he already holds "dialogues with many of the researchers in China" and that both the United States and the United Kingdom should deepen them.

"We ought to enhance the dialogue," he said. "The United States should do it... I advocated that the U.K. do the same thing."

Huang framed Britain as a possible neutral convener, pointing to its "intellectual capital" as the reason it could host talks neither Washington nor Beijing would chair alone.

The split inside AI

The comments sharpen a divide building for months. OpenAI's Sam Altman and Anthropic's Dario Amodei have spent the autumn calling for slower development and firmer government rules. Huang has become the industry's loudest opposing voice, arguing that AI firms are already covered by cybersecurity and product-liability law.

"Apply that first," he told CBS News. "Don't let this doomsday narrative allow someone to relieve them of the laws that currently exist."

Answering criticism from Pope Leo XIV in a Fortune interview, Huang said "we all care about the same thing," pointing to Nvidia's open-source safety work and its platform for containing rogue agents.

Why it matters for builders

For teams shipping agents, the practical question is which safety regime their stack assumes. If Huang's dialogue channel becomes formal, cross-border incident reporting stops being a public-relations exercise and becomes an operational requirement, read off the infrastructure layer where Nvidia already sits.

Self-regulation also shifts cost. When a vendor ships its own containment and audit tooling, the teams deploying agents inherit the obligation to configure it, log it, and prove it. Procurement questions about a model provider's safety framework are becoming as ordinary as questions about latency and price.

The standoff is not really about whether AI is risky. Both camps agree it is. It is about who gets to define the boundary, and who carries the bill when it is crossed.

What to watch

• Xi's Washington visit: whether the US-China AI Dialogue opened by Treasury Secretary Scott Bessent and Chinese counterpart He Lifeng produces a formal incident-notification mechanism.
• The UK's convening role: whether London answers Huang's pitch and hosts the first cross-border AI safety meeting.
• The regulatory gap: if Washington keeps deferring to existing law, self-regulatory frameworks become the de facto standard builders get measured against.