The US government has alleged that Chinese AI firm Moonshot illicitly copied Anthropic's Fable LLM to develop its Kimi K3 model. Experts are questioning the technical feasibility of the claim, while the incident highlights growing tensions over semiconductor export controls and intellectual property in the AI sector.
Detailed Coverage
The U.S. government has raised serious allegations against Moonshot, a Chinese artificial intelligence company, accusing it of building its Kimi K3 model by illicitly copying Anthropic's Fable large language model. Michael Kratsios, a U.S. science advisor, stated that the development of Kimi K3 involved industrial-scale distillation aimed at stealing U.S. proprietary technology. These remarks follow similar concerns expressed by Treasury Secretary Scott Bessent, who noted that certain Chinese models appear to carry characteristics of American AI architectures.
Technical Doubts and Industry Expertise
Independent AI researchers have expressed skepticism regarding the specific claim that simple model distillation explains the rapid advancement of Kimi K3. Braden Hancock, co-founder of Snorkel AI, pointed out that the timeline is extremely tight, as Kimi K3 surfaced just weeks after the public release of Fable on July 1st. Many experts believe that achieving such high performance in such a short window is mathematically unlikely through basic distillation alone. Nathan Lambert from the Allen Institute for AI added that as models reach frontier performance levels, simple fine-tuning techniques become less effective, suggesting that Chinese firms may be employing more complex reinforcement learning methods that require significant independent computational resources.
Furthermore, industry observers have noted that underestimating the capabilities of Chinese AI talent may be a strategic error. Moonshot’s founding team includes researchers with significant academic credentials, and some analysts suggest that China’s progress in the sector is increasingly driven by native innovation rather than solely relying on the replication of Western models. Distillation itself is a widely used practice in the global AI industry, sometimes used to create synthetic training data, making it difficult to distinguish between legitimate model training and intellectual property theft.
Hardware Restrictions and Black Market Risks
Beyond software, the U.S. allegations include the claim that Moonshot gained access to restricted Nvidia Grace Blackwell 300 chips through servers located in Thailand. The U.S. maintains strict export controls on high-end semiconductors to prevent their use in military or advanced AI applications in China. The presence of a hardware black market that bypasses these restrictions has become a major concern for U.S. policymakers. Researchers are now calling for stricter oversight of data centers, including potential regulations that would require providers to verify the end-users and purposes for which advanced computing hardware is being deployed.
Moonshot has faced similar accusations in the past. Earlier this year, Anthropic identified instances where it alleged that Moonshot and other firms like DeepSeek and MiniMax systematically extracted data from its models. As the global AI race intensifies, investors and stakeholders in the technology sector should continue to monitor how these geopolitical disputes impact cross-border research collaboration, the availability of high-end chips, and the potential for new trade restrictions in the artificial intelligence supply chain.