Fifteen days. That's reportedly how long it took between Anthropic's Fable model coming back online after export-control restrictions and Moonshot AI shipping Kimi K3. The White House looked at that timeline, did some math, and decided it added up to something a lot more pointed than coincidence.

An Accusation With a Very Specific Number Attached

Michael Kratsios, director of the White House Office of Science and Technology Policy, publicly accused Chinese AI firm Moonshot AI of running large-scale covert distillation against Anthropic's Fable model to build Kimi K3. Kratsios was careful to note that ordinary distillation — the standard technique labs use to compress a big model's knowledge into a smaller one — is completely normal. What he's alleging is different: an internal Moonshot platform built specifically to run large-scale extraction against American models while switching between access methods to dodge detection.

He separately claimed Moonshot obtained restricted Nvidia GB300 chips accessed via Thailand. Treasury Secretary Scott Bessent followed up with the part that actually has teeth: a threat of sanctions and export-control blacklisting. Notably, the OSTP post offered the timeline and the accusation, but no public forensic evidence showing exactly how the extraction supposedly happened.

The Real Fight Isn't About One Model

This is playing out against a US-China AI rivalry that's already tense enough without accusations of industrial-scale IP theft flying between a government office and a foreign lab. Whether or not Moonshot did exactly what's alleged, the accusation itself is a policy weapon, and it lands at a moment when open-weight Chinese models keep matching or nipping at the heels of their closed-source American counterparts.

Even OpenAI's Greg Brockman got dragged into the discourse, allowing that Kimi K3 is “pretty good” while declining to confirm or deny the distillation claim — which is corporate-speak for “we noticed too, but we're not touching this with a ten-foot pole.” That non-answer is arguably more interesting than a denial would have been.

What nobody disputes is that the walls around frontier models are looking more like chain-link fences than vaults. Export controls can restrict who buys the chips, but they can't stop a sufficiently motivated lab from prodding a rival's API until it spills its secrets one distilled response at a time.

Sanctions or not, this is the shape of AI geopolitics now: less about who trains the biggest model, and more about who can prove — or plausibly allege — that somebody else just borrowed theirs.

Source: TipRanks