U.S. Accuses Moonshot AI of Industrial-Scale Theft of Anthropic Tech
The White House alleges the Chinese startup used covert distillation and proxy servers in Thailand to bypass export controls.
The U.S. government has accused Chinese AI startup Moonshot AI of stealing proprietary technology from American firm Anthropic to build its own large language model. The allegations, delivered via the White House Office of Science and Technology Policy (OSTP), mark a significant escalation in the geopolitical struggle over artificial intelligence supremacy.
According to the administration, Moonshot AI utilized a process known as "large-scale industrial distillation" to copy Anthropic's Fable model, using the stolen data to develop its Kimi K3 model. U.S. officials allege that Moonshot AI built a sophisticated internal platform specifically designed to conduct this distillation at scale, rotating its access methods to evade detection by American security systems.
Beyond software theft, the U.S. claims Moonshot AI bypassed hardware export restrictions to obtain high-end computing power. The administration alleges the company illegally accessed Nvidia GB300 Blackwell chips—which are prohibited for export to China—by routing access through servers located in Thailand. Treasury Secretary Scott Bessent reinforced these claims, stating that watermarks from U.S. large language models have been discovered across numerous Chinese models, calling the trend "unacceptable."
The Mechanics of Distillation
Distillation is a standard AI technique where a smaller, more efficient model is trained using the outputs of a larger, more complex "teacher" model. While the practice is common in open-source development, the U.S. government distinguishes between legitimate research and what it terms "covert industrial distillation." In this instance, the White House argues the process was used as a tool for intellectual property theft to undermine American research and development.
Strategic Implications
This confrontation highlights two critical vulnerabilities in the current AI landscape: the ease with which proprietary model weights can be mimicked through distillation and the difficulty of enforcing hardware sanctions. The use of Thailand as a proxy for Nvidia chips demonstrates how third-party nations can create loopholes in U.S. export controls, allowing restricted technology to reach prohibited actors.
Future Outlook
Michael Kratsios, Director of the OSTP, has characterized the theft of proprietary U.S. technology as unacceptable. The accusations signal a shift toward more aggressive enforcement, with the potential for targeted sanctions against Chinese AI firms found to be engaging in intellectual property theft. As the administration increases scrutiny on tech loopholes, the industry is watching to see if these allegations will lead to formal trade penalties or further restrictions on AI-related exports.