US weighs AI model restrictions as Chinese open-source share hits 80%

The US government is weighing restrictions on Chinese open-source AI models after Moonshot AI's Kimi K3 triggered distillation accusations, a move that could reshape a supply chain where 80 percent of developers use Chinese tools.

"We have information that Moonshot AI distilled Anthropic's Fable for the development of its K3 model," Michael Kratsios, White House science and technology policy advisor, posted on X. Treasury Secretary Scott Bessent followed with a warning that "sanctions and Entity List designations will be on the table" for companies found to be improperly distilling AI models.

The accusations center on distillation, a technique where smaller models learn from the outputs of larger frontier models. Anthropic said its Claude capabilities were distilled on an "industrial scale" by DeepSeek, Moonshot and MiniMax, which used about 24,000 fake accounts generating 16 million exchanges. Chinese open models accounted for more than 50 percent of all AI use on OpenRouter in some weeks this summer, up from a weekly average of roughly 13 percent in 2025. Microsoft is exploring a fine-tuned version of DeepSeek V4 as a low-cost alternative for Copilot Cowork, with DeepSeek costing less than a tenth of Anthropic's Opus.

A ban would raise startup costs and concentrate pricing power among a handful of US incumbents, according to Selina Xu, a fellow at the Penn Project on the Future of US-China Relations. Nvidia, Microsoft, Meta and Palantir joined more than 20 other companies in a letter urging policymakers to avoid "premature restrictions" on open-weight AI models. Anthropic, valued at close to $1 trillion, said stopping illicit distillation was a matter of national security.

Distillation: Common Practice or IP Theft?

Distillation is widely used across the AI industry. Google's Jeff Dean said in February that the company discovered distillation techniques while looking to improve performance without relying on one large image recognition model. Nvidia used distillation as part of the training process for its Llama Nemotron series, as detailed in an accompanying research paper.

"Distillation, or the practice of using one model's outputs to help train or improve another, is a widely used technique for model improvement, evolution, and validation," the industry letter stated.

The line between legitimate distillation and IP theft is contested. OpenAI and Anthropic have banned the practice in their terms of service, arguing that using their larger models without authorization represents potential IP theft. But both companies have themselves relied on external content to build their models and have been sued for doing so. Max Pritt, an attorney for Boies Schiller Flexner who represents book authors in copyright litigation against AI firms, said the administration has focused on protecting technology companies' IP while remaining silent on creators' IP used without authorization.

The Open-Source Economics at Stake

Chinese open-weight models have become integral to the US AI supply chain. When some OpenAI models recently hacked Hugging Face, the victim had to turn to a Chinese open-weight model to defend itself because the safety guardrails of closed US models blocked its requests. Thinking Machines Lab, the startup founded by OpenAI's former chief technology officer Mira Murati, used Moonshot AI's Kimi K2.5 to generate some of its early post-training data.

The economics are stark. Moonshot AI's annual recurring revenue spiked to a record high after Kimi K3's release, according to company president Zhang Yutong. Chinese open-source AI is effectively subsidizing the rest of the world through cheap access to intelligence.

For US startups, the stakes are existential. Pukar Hamal, founder of AI security firm SecurityPal, said he would use Chinese open-weight models like Kimi K3 at his company because they could save significant money. "We would make sure that there's no nefarious backdoors in the code," Hamal said. "But hosting it on our own infrastructure after we've done an assessment, why not?"

The White House has already tightened export rules covering frontier AI models, banning Anthropic's Fable from export and requiring OpenAI and Google to release frontier models in preview to selected partners. China is considering restricting the export of its own frontier AI models in response.

Selina Xu argues the US should exclude foreign models from intelligence systems, military networks and critical infrastructure, but a model downloaded, independently evaluated and operated on a US company's own infrastructure poses a different risk from a Chinese-hosted service. "The answer to China's growing AI strength isn't to construct an American version of the Great Firewall," she wrote. "It is to outcompete Chinese models."

For investors, the policy outcome carries direct implications. A ban on Chinese open-weight models would strengthen the pricing power of OpenAI and Anthropic, which charge premium rates for frontier access. But it would also raise costs for the broader AI startup market that depends on cheaper Chinese alternatives. Microsoft, which is exploring DeepSeek V4 for Copilot Cowork, could face higher infrastructure costs if the option is removed. Moonshot's Kimi K3 release already caused a dip in US stocks before the industry pushback letter was published, and further policy moves could trigger additional volatility across AI-linked equities.

This article is for informational purposes only and does not constitute investment advice.