U.S. authorities have filed official charges against Chinese startup Moonshot AI. According to a statement from White House spokesperson Michael Kratzios, the company's developers secretly copied Anthropic's Fable 5 model to create their own neural network, Kimi K3. This is not just another patent dispute—it is a direct threat to U.S. technological sovereignty.
Michael Kratzios published a resonant post on social media platform X, stating that the Chinese side had created a separate infrastructure to massively bypass protective barriers. The developers constantly changed network routes to mask suspicious requests.
The White House's fight against illegal neural network distillation
The conflict erupted around a popular method called distillation. This technology allows training new algorithms on the responses of a more advanced system. Legal use of the method has long become an industry standard. However, covert large-scale data collection is equated by U.S. experts to outright theft.
"We have data that Moonshot AI conducted distillation of Anthropic's Fable to create its K3 model... However, large-scale covert distillation aimed at stealing American technology and undermining the U.S. research base is unacceptable," said Michael Kratzios.
Anthropic holds the same position. According to the company's February report, specialists tracked over 3.4 million interactions with Claude linked to Moonshot. Traffic came from hundreds of fake accounts—credentials matched the public profiles of senior Moonshot employees.
As a result of the attack, key areas of Claude's work were affected:
- Image processing
- Logical reasoning algorithms
- Software code generation
Additionally, Kratzios addressed the issue of illegal hardware imports. Moonshot utilized powerful NVIDIA GB300 servers in Thailand. Such equipment is banned from import into China due to existing sanctions. In connection with this, Taiwanese law enforcement conducted a raid on supplier Super Micro in June.
The impact of Kimi K3 on the U.S.-China technology race
The presentation of Moonshot's flagship project took place on July 16. The architecture of the new model contains 2.8 trillion parameters, making it the largest among open neural networks. Heightened audience interest forced the authors to freeze registration after 48 hours. The developers promise full access to the weights by July 27.
The publication date is significant: Anthropic warns that copied models lose their safety restrictions. Once released into the public domain, these versions could spread without control.
This is not the first such complaint. In January 2025, White House AI chief David Sacks accused DeepSeek of copying OpenAI models. DeepSeek denied this. Beijing called Anthropic's findings, published in February, baseless.
Kratzios did not announce sanctions, and Moonshot has not commented on the situation.
Cryptalist Analysis: This incident is just the tip of the iceberg in the escalating technology war. The U.S. advantage in AI investment remains enormous—spending is nearly 23 times higher than in China. However, Chinese developments are rapidly closing the gap, and this case shows that Beijing is not above using any methods. The next move is up to Washington, and it could be harsh.