Washington Weighs Curbs on China's Open AI Models as They Gain Ground

The Trump administration is once again weighing how to slow the spread of Chinese open-weight AI models inside the United States, according to reporting from Tom's Hardware and others. The trigger this time was the launch of Moonshot AI's Kimi K3, and the stated worry is cybersecurity: officials fear downloadable Chinese models could carry hidden flaws or backdoors. It is worth being precise here. That concern has been raised repeatedly, but no public evidence has established such backdoors exist.

Why the timing is awkward

The push comes exactly as these models are winning on merit. Kimi K3, released in July, is the largest open-weight model shipped to date at 2.8 trillion parameters, and it debuted third on the Artificial Analysis leaderboard, behind only Anthropic's Claude Fable 5 and OpenAI's GPT-5.6 Sol, while posting the top score of any open model. DeepSeek, meanwhile, has become one of the most-used models on the developer platform OpenRouter. Over the first half of 2026, the combined share of tokens running through Google, Anthropic and OpenAI on that platform fell from roughly 55 percent to about a third. Chinese open models absorbed much of the difference.

That popularity is precisely what makes a ban hard to enforce. A closed API can be switched off. Open weights, once downloaded, live on thousands of private servers and laptops. Officials appear to know this, which is why the emerging approach looks less like an outright prohibition and more like pressure applied at the edges: federal procurement rules, the threat of the Entity List, and formal advisories warning companies about foreign software.

An industry that does not agree with itself

The American AI industry is not speaking with one voice on this. Reporting suggests some frontier labs are comfortable with tighter controls, while chipmakers and others who benefit from a thriving open ecosystem are wary of anything that pushes developers away from a shared, global pool of models. Nvidia in particular has spent the year arguing that cutting American tools off from foreign markets mostly helps foreign competitors.

There is a real tension underneath the politics. Open models lower costs and spread capability, which is good for developers and bad for anyone hoping to keep frontier AI inside a small number of trusted hands. It is the mirror image of a story we covered last month, when Beijing began weighing its own curbs on AI exports. Both capitals are discovering that model weights are far harder to police than chips, and that the country doing the restricting risks slowing itself down first.