The US says Moonshot secretly cloned Anthropic's Fable

China's open models are catching up to America's frontier labs at breakneck speed, and Washington is running out of patience.

On Wednesday, White House tech policy chief Michael Kratsios publicly accused Chinese AI startup Moonshot of illicitly distilling Anthropic's flagship Fable model to build its newly unveiled Kimi K3. Kratsios also alleged that Moonshot obtained servers equipped with Nvidia's restricted GB300 chips and accessed them in Thailand to train its systems.

Treasury Secretary Scott Bessent quickly turned up the heat, warning that open source is not open season on American IP and confirming that the administration is weighing sanctions and Entity List designations against Moonshot.

A disputed timeline and Thai servers

Model distillation, which involves training a model on the outputs of a larger one, has long been a standard optimization technique across the AI industry. But Washington is now framing it as covert, industrial-scale theft.

Anthropic policy executive Sarah Heck joined the fray on X, arguing that Chinese theft of U.S. frontier models creates grave national security risks. The company previously claimed in February that Moonshot, DeepSeek, and MiniMax used roughly 24,000 fake accounts to generate more than 16 million unauthorized interactions with Claude.

Yet technical observers note that the timeline does not entirely add up. Anthropic's Fable only became publicly accessible on July 1, making it technically implausible that Moonshot could have trained its massive 2.8 trillion-parameter K3 model primarily through distillation in just three weeks.

Chinese officials pushed back forcefully against the allegations. An embassy spokesperson in Washington called the accusations entirely unfounded, while Beijing's foreign ministry accused the U.S. of politicizing and weaponizing technology.

Why it matters

We think this clash exposes a deep vulnerability in the frontier AI race: massive capital investments in proprietary models can be challenged by massive open-weight systems built at a fraction of the cost.

For enterprise developers, open systems like Kimi K3 offer near-frontier intelligence without vendor lock-in. But if Washington follows through with Entity List designations or broad restrictions on Chinese open models, using these weights could soon carry serious legal and compliance penalties, especially for multinational firms.

For the broader industry, the battle lines are shifting from silicon to software. Silicon Valley figures, including OpenAI Head of Strategic Futures Dean Ball, are already urging policymakers to restrict foreign open-weight models to protect America's edge.

Washington's playbook looks predictable yet increasingly hard to enforce. If hardware restrictions in third countries like Thailand fail to contain access, turning intellectual property disputes into sweeping geopolitical bans risks fragmenting the global AI research community beyond repair.