DeepMind and Anthropic Warn AI Chip Exports to China Decide the AI pace

Something unusual happened at Davos this week. Two of the biggest voices in frontier AI stopped talking about models and started talking about hardware, as if the real race is being decided long before any chatbot reaches a user.

Anthropic CEO Dario Amodei and Google DeepMind CEO Demis Hassabis turned Davos into a forum on China and advanced AI chips, framing the discussion as a direct argument about how much chip access still shapes the global AI balance.

What Anthropic said at Davos about China AI chips

Amodei called the idea of selling advanced AI chips to China “crazy” and compared it to “selling nuclear weapons to North Korea.” This was not just abstract trade speak. The Davos coverage tied the debate to specific frontier accelerators, such as Nvidia’s H200 and AMD’s MI325X, which is why Amodei framed the decision as a direct shortcut for closing any current lead in high-end compute.

He also argued that the United States maintains a meaningful lead in the most advanced chips and that easing export controls would allow China to close the gap faster than policymakers expect.

What DeepMind said at Davos about China progress and chip constraints

Hassabis noted that Chinese AI companies are about six months behind the leading Western labs, while emphasizing that US restrictions limit access to the most advanced semiconductors needed to develop and run frontier AI systems.

That gap sounds small in calendar time, but in modern AI cycles it can represent an entire generation of capability, the difference between a model that competes and one that redefines how companies automate work at scale.

Hassabis alsocalled DeepSeek’s work “impressive,” which sharpens the tension in his claim. If Chinese teams can be impressive under constraints, the unspoken threat is clear: what happens if the constraints loosen and the same work gets fed with top-tier compute?

Why these Davos comments matter on a global scale

This is a power story. Chips decide who can train, who can scale, and who can iterate fast enough to stay at the frontier. Amodei is arguing that letting high-end accelerators flow to China is the fastest way to compress any current lead. And that the lead exists partly because chip access is already constrained.

That is why the impact is worldwide even if the policy tool is national. Training and inference do not grow on ideas alone. They grow on steady access to computing power, and computing power depends on cross-border supply chains, export rules, and which markets can buy the newest chips in large numbers. When CEOs talk about chips at Davos, they are speaking to investors, governments, and business buyers all at once.

Why this is a boardroom issue, not just a policy issue

For most companies, this will not appear as a geopolitical headline. It will appear as whether AI moves from tests to real use. When computing power is scarce or politically uncertain, the cost of trying things goes up, rollout schedules get longer, and the safest choice becomes doing nothing.

The difference between having computing power and waiting for it becomes the difference between launching and falling behind.