US Threatens Sanctions on Moonshot's Kimi K3 Model as China-US AI Distillation Allegations Escalate

On July 21, 2026, US Treasury Secretary Scott Bessent publicly stated that watermarks from US large language models were found in multiple Chinese models, and said Washington may impose sanctions on foreign models suspected of stealing US companies' intellectual property. The next day, White House Office of Science and Technology Policy Director Michael Kratsios accused Moonshot AI of distilling Anthropic's Fable model to build Kimi K3, and setting up an internal platform to extract data at scale, while switching access methods to evade detection.

The core mechanism of the allegations lies in model distillation technology. Developers typically use the outputs of stronger models to train smaller models, a common practice in the industry. But US officials believe Moonshot generated over 16 million interactions with Claude through thousands of fraudulent accounts, constituting industrial-scale covert extraction. In a February 23 report, Anthropic had already disclosed that three Chinese labs including Moonshot created about 24,000 fake accounts to extract Claude's reasoning, coding, and computer use capabilities.

Kimi K3 was released on July 16, 2026, with 2.8 trillion parameters. Michael Kratsios also pointed out that Moonshot obtained servers based on Nvidia GB300 and used GB300 hardware in Thailand for training, which involves the Blackwell series chips that US export controls prohibit from selling to China. If the allegations are substantiated, the distillation dispute would trigger two legal fronts: intellectual property infringement and export control violations.

For Chinese developers, the threat of sanctions could cut ties with the US financial system and restrict US suppliers from providing goods without authorization. Moonshot plans to release full model weights on July 27, 2026, and seek a Hong Kong listing. If sanctions are imposed, these plans would face direct obstacles. Open-source advocates worry that further US restrictions on using Chinese cutting-edge open-source models domestically would weaken developers' access to low-cost options.

For US enterprise users, OpenRouter data cited by KuCoin shows that the token share generated by Chinese models for US enterprises rose from less than 5% in early 2025 to 46% in April 2026, with DeepSeek accounting for 17.6%. Price is the main factor—DeepSeek V4 Flash costs $0.14 per million tokens, far lower than comparable US models. If sanctions expand, users may be forced to reassess supply chain costs and compliance risks.

Beijing has responded. China's Ministry of Commerce warned that if US actions substantially harm China's interests, it will take all necessary measures to protect legitimate rights and interests. Both sides have linked this incident to the September China-US AI negotiations. The US move may increase its bargaining leverage, while Beijing may retaliate in areas such as chip design exports.

Historical comparisons show that previous chip export restrictions did not prevent Chinese labs from advancing their own alternatives, but instead accelerated related development. The core evidence for the allegations has not been fully disclosed. Fable was released on July 1, 2026, and Kimi K3 on July 16.

The most likely development is that the US will maintain pressure before the September negotiations while observing whether Moonshot releases its weights as planned. If the share of Chinese models in US enterprise token usage continues to rise, or more public watermark evidence emerges, it will be a key signal to verify the necessity of sanctions.