The White House director of science and technology policy, Michael Kratsios, on Wednesday publicly accused Chinese startup Moonshot AI of distilling Anthropic’s Fable model to build its newly released Kimi K3, alleging the firm used a sophisticated internal platform to evade detection. The charge immediately rippled through financial spread-betting markets: on London-based platform IG, implied valuations for Anthropic slumped by $232bn to $1.56tn and OpenAI by $160bn to $1.16tn between Friday and Tuesday, though both companies remain privately held and the figures represent market bets rather than actual capitalisation.

Kratsios said Moonshot developed a covert, large-scale distillation operation that could rapidly switch access methods to avoid security systems, and that it acquired servers with Nvidia GB300 chips deployed in Thailand, likely to train its models. While distillation—training a smaller model on a larger one’s outputs—is a legitimate technique, the US official drew a sharp line at “large-scale, covert industrial distillation aimed at stealing proprietary U.S. technology.” The statement follows a State Department diplomatic cable in April that ordered a global campaign to highlight what Washington views as systematic Chinese IP theft from American AI labs, and Treasury Secretary Scott Bessent has raised the prospect of sanctions.

Viewed from Washington, the accusation marks a rare alignment between the Trump administration and Anthropic, which had previously seen its own most advanced models suspended by the US government on security grounds. In Sydney and Canberra, the episode lends momentum to Australia’s push to set AI regulations that big US developers have welcomed as a way to create predictable investment environments and, critics say, to lock out Chinese competitors. From New Delhi, analysts describe Kimi K3 as a strategic headache, likening AI models to fifth-generation fighter jets and warning that India lacks the compute, funding, and talent to field a sovereign alternative.

Kimi K3, a 2.8-trillion-parameter open-source model, has been benchmarked at or near the level of Anthropic’s Fable and OpenAI’s leading systems, and its launch last week forced Moonshot to temporarily pause new subscriptions due to demand. The model’s open-weight release, scheduled for next week, will allow local deployment and customisation, further eroding the moat of proprietary US systems. Anthropic, meanwhile, is navigating its own pressures, having agreed on Monday to a $1.5bn settlement with authors over copyright infringement.

The next concrete milestone is Moonshot’s planned release of K3’s full model weights, which will test whether the US responds with export controls, sanctions, or diplomatic pressure before the technology becomes widely embedded.