The US-China AI race sharpened on Monday, when the US Department of Homeland Security ordered civilian agencies to stop using DeepSeek, even as Washington and Beijing negotiate AI cooperation ahead of the Xi-Trump meeting in Shenzhen next month. The two tracks show how the world’s two AI powers now compete and talk at the same time.
DeepSeek ban: what Washington ordered
A DHS notice issued on October 19 tells civilian agencies to remove DeepSeek products, and tools that third parties built on DeepSeek, from their IT devices. DHS can direct civilian agencies but not the military or intelligence services, and the Defense Department did not respond to AFP’s request for comment. The notice says DeepSeek’s models present multiple threat vectors and could send sensitive data to China, where the company’s servers sit and where the legal system gives the government wide access to private infrastructure.
DeepSeek released its V4 model in April and updated a faster, lighter version this month. Washington accuses DeepSeek and other top Chinese labs of systematically copying the capabilities of US frontier models through distillation. Beijing calls distillation a common practice in the AI sector.
Xi-Trump Shenzhen summit: AI cooperation on the agenda
The South China Morning Post reports, citing sources, that US and Chinese officials are meeting regularly to flesh out a framework announced in September for handling AI-related incidents and holding a standing AI dialogue. Technical teams want AI cooperation to be a deliverable when Xi Jinping and Donald Trump meet on the sidelines of the APEC summit in Shenzhen next month, their third in-person meeting this year.
The contest over rules continues in parallel. Washington is building a network of trusted partners called Pax Silica, while Beijing promotes the World Artificial Intelligence Cooperation Organisation. The Trump administration has rejected calls to slow frontier AI development, pointing to competition with China.
How big is the US-China AI gap?
Estimates put the US lead at roughly six to seven months. The Institute for Progress, in a new report, puts it at about 6.5 months and argues the gap is a policy choice, shaped by China’s access to chips and its distillation of US models. It estimates China will have access to about 4.6 million H100-equivalents of AI compute by the end of 2026, counting domestic production, legal imports, smuggled chips and US chips accessed remotely in data centres abroad. Epoch AI’s tracking has Chinese models lagging by about seven months on average since 2023.
NYU professor Gary Marcus told Bloomberg he expects a tie rather than a clear winner, like Coke and Pepsi: US models are generally better, while Chinese models are cheaper.
An entangled race
The Washington Post reported on October 11 that US data centers are full of China-made equipment and that many US companies rely on Chinese AI, even as American officials describe the contest as winner-take-all. That overlap complicates talk of a clean break.
What to watch
- Shenzhen outcomes. Whether the two leaders formalise the AI incident channel and the AI dialogue.
- Follow-on bans. Whether other agencies, contractors or states copy the DHS order.
- Distillation fights. US officials are pressing the case, and China rejects the charge.
- Compute access. Chips and remote access to data centres abroad are the levers analysts keep pointing to.
Figures are as reported and may be revised.