US opts for targeted bans on Chinese AI models rather than a blanket prohibition

The Trump administration favours targeted bans on specific Chinese AI models over a blanket prohibition, The New York Times reported on 26 July 2026. National security concerns and suspicions of intellectual-property theft are central to the deliberations. The US Treasury Department is investigating whether prominent Chinese AI models were built unlawfully on the output of American systems.

Running in parallel is a public debate on open-weight models, in which companies including OpenAI and Google DeepMind signed an open letter opposing broad regulation, while those same companies are lobbying in Washington for restrictions. The gap between their public and private positions is attracting considerable attention in the US tech sector.

Distillation at the heart of the dispute

Central to US concerns is the technique of model distillation: a smaller AI system learns from the output of a more capable system, thereby acquiring characteristics of that stronger model without direct access to its underlying weights or training data. Michael Kratsios, director of the White House Office of Science and Technology Policy, accuses Chinese companies of what he describes as large-scale, covert industrial distillation of patented American technology, citing Moonshot AI as an explicit example.

Moonshot AI developed the Kimi K3 model, a system with 2.8 trillion parameters that the company presents as the largest open-weight model in the world. According to Moonshot's own benchmarks, the model performs close to Anthropic's Claude models. Dean Ball, head of strategic futures at OpenAI, questions this, arguing that those results cannot be fully explained by distillation alone.

That the debate is not purely theoretical is illustrated by Thinking Machines Lab, which demonstrably trained its open model Inkling using Moonshot's Kimi 2.5. The way Chinese open-weight models spread in practice and find their way into other systems makes the case for targeted measures more concrete than an abstract policy question.

Open letter attracts more than fifty signatories

On 24 July 2026, an open letter titled Open Weights and American AI Leadership was addressed to US policymakers, urging them to avoid broad restrictions on open-weight AI models. Initial signatories included Hugging Face, Meta, Microsoft, Mistral, Nvidia and Replit. Nvidia CEO Jensen Huang shared the letter on X.

Two days later, OpenAI, Google DeepMind, AMD, Cisco, Cloudflare, GitHub, Block and Ollama had also added their signatures, bringing the total to more than fifty signatories, according to Forbes on 25 July. OpenAI president Greg Brockman had already stated on 23 July that he was not involved in discussions about a ban on Chinese open-weight models and that AI use should be accessible to a broad public.

Replit CEO Amjad Masad put it sharply: in practice, a ban on Chinese open models amounts to a ban on open models in general. Amazon and Anthropic did not sign the letter.

Anthropic in a dissenting camp

On this point, Anthropic is taking a different line from most major tech companies in Silicon Valley. The Information reported on 26 July that the company is actively lobbying for restrictions on open-weight models, a position that sets it apart within the industry. Anthropic does not offer open-weight models and bases its commercial position on API access to the closed Claude models. From that position, the company has a direct financial interest in limiting cheaper Chinese alternatives.

OpenAI is also lobbying privately for restrictions, while publicly signing the open letter. The fact that these two positions coexist raises questions about the extent to which the public stance reflects internal policy. Sanctions have not yet been announced. Treasury Secretary Scott Bessent indicated that sanctions against Chinese AI companies are being considered if intellectual-property theft is confirmed.

Which models and companies are in the frame

In addition to Moonshot AI and its Kimi K3 model, US officials are also naming DeepSeek as a candidate for targeted measures. Alibaba, the parent company of Moonshot AI, is also in scope; Alibaba's Qwen 3.8 Max model scores just behind Claude in benchmarks. Chinese AI company Z.ai has also released an open-weight model under the name GLM 5.2.

The Treasury Department is preparing to examine these models for evidence of unauthorised distillation. Placement on the Entity List, which would exclude US companies from cooperation, is among the instruments the government is considering. No concrete decisions had been taken at the time of writing.

What this means for European AI developers

The US debate is relevant to European companies and policymakers for several reasons. Open-weight models of Chinese origin, such as DeepSeek and the Kimi family, are actively used and integrated into products across Europe as well. If the US introduces targeted export restrictions or sanctions, European companies building on these models could face legal and commercial complications in their dealings with American partners.

At the same time, the open letter shows that major US technology companies, including Microsoft and Nvidia, both of which have significant operations in Europe, are publicly opposing broad restrictions on open-weight models. That position aligns with the preference of many European AI companies and policymakers for open models as the foundation for sovereign AI development. How the European Commission and national governments weigh the security risks associated with Chinese open-weight models against the benefits of openness is a question that becomes more pressing in light of the US policy shift.