US accuses China's Moonshot of stealing Anthropic tech for Kimi K3

The US has accused Moonshot, a prominent Chinese AI startup, of conducting what the White House described as "industrial-scale" theft of American intellectual property to create its latest model, Kimi K3. According to the allegation, Moonshot improperly used Nvidia chips and Anthropic's Fable model in the development of Kimi K3. The US indicated it is considering sanctions in response to this conduct.

The announcement comes in the wake of market turbulence triggered by Kimi K3's advanced capabilities, which caused a US tech selloff last week. That selloff echoed the shock to markets from DeepSeek, a rival Chinese AI company whose capabilities similarly rattled investor confidence in American AI dominance. The timing suggests that the White House is responding not only to the intellectual property theft itself but also to the competitive threat that successful Chinese AI models pose to US commercial and national security interests.

Central to the US concern is a technique called "distillation," in which one AI model is trained on the outputs of another, more advanced model. Washington believes China is systematically exploiting this method to develop frontier AI capabilities that could endanger US national security. The White House is stepping up efforts to deter distillation and similar practices.

However, experts acknowledge a fundamental enforcement challenge. Semafor's tech editor noted that "you can't stop software at the border," underscoring the reality that digital models and code are difficult to contain through traditional border controls. Instead, the editor suggested, the US should focus on "hardening your own defenses"—implying that technical and operational security measures within American institutions may be more effective than external enforcement.

The allegation against Moonshot reflects an escalating pattern of US-China tension over AI development and intellectual property. After the market disruption caused by DeepSeek's capabilities, the White House is signaling a harder line against what it views as systematic theft by Chinese competitors. By naming both Nvidia chips and Anthropic's Fable model as sources of misappropriation, the US is drawing attention to two vulnerabilities in its AI supply chain: hardware access and algorithm leakage through model outputs.

Washington's focus on deterring distillation is strategically significant because the technique is difficult to prevent through traditional legal or export controls—once a model's outputs are public, reverse-engineering and fine-tuning become feasible. The US threat of sanctions signals an intent to raise the cost of such practices, but the underlying challenge remains structural: as Semafor's tech editor observed, software cannot be stopped at borders in the way physical goods can. This suggests that US policymakers may need to shift focus toward securing their own systems and limiting the information leakage that makes distillation effective in the first place.