Michael Kratsios, director of the White House Office of Science and Technology Policy, accused the Beijing startup Moonshot A.I. on Wednesday of covertly extracting knowledge from an Anthropic model to build its newly released Kimi K3, and of skirting U.S. export controls by acquiring restricted Nvidia servers routed through Thailand. Posting on X, Kratsios cited “information that Moonshot AI distilled Anthropic’s Fable for the development of its K3 model,” and said the company had “developed a sophisticated internal platform to conduct large scale distillation against U.S. models, allowing them to quickly switch between multiple methods of access to avoid detection.” Moonshot, he added, “acquired GB300-equipped servers and has accessed GB300s in Thailand, likely to train its AI models.”

The accusation lands four days after Kimi K3 shipped as a 2.8-trillion-parameter open-weight system. Demand was heavy enough that Moonshot suspended new subscriptions within 48 hours, citing compute constraints. Moonshot, founded in 2023 and backed by Alibaba, Tencent, and Meituan, claims the model matches or surpasses OpenAI’s GPT 5.6 Sol and Anthropic’s Claude Fable 5 across several benchmarks. It has denied training on American systems, and a spokesperson for the Chinese embassy in London said Beijing “opposes baseless allegations and malicious smears against its AI development.”

Treasury Secretary Scott Bessent had already primed the escalation. On Fox Business Tuesday, and again on X, he wrote that “when PRC firms conduct covert, industrial-scale distillation attacks that cross the line into IP theft, sanctions and Entity List designations will be on the table.” He went further: “We are finding watermarks of our U.S. large language models on many of the Chinese models, and that that’s unacceptable.”

Independent analysts read the fingerprints similarly. Hamish Low, a research associate at the Institute for AI Policy and Strategy, told The Hill it’s “pretty overwhelmingly likely they are distilling from leading U.S. models,” pointing to Kimi’s strong coding performance alongside comparative weakness in cybersecurity as a pattern consistent with selective extraction from frontier systems rather than ground-up training.

The structural picture is familiar to anyone who watched the 2019 Huawei Entity List cycle: a Chinese product outperforms expectations, Washington responds with an attribution narrative and a threat of designations, and the enforcement question, whether the export regime can actually police GPUs moving through Southeast Asian intermediaries, gets deferred to a later news cycle. That deferral is where the Trump administration’s AI posture has sat since the abrupt May cancellation of a landmark executive order. Chris Lehane, OpenAI’s chief global affairs officer, said the episode “only reinforces the importance of the U.S. really establishing a coherent framework and process” for model testing.

It’s a framework the administration hasn’t yet produced, and Kimi K3 is now the artifact it has to explain around.

Sources