You Can't Distill Fable in Two Weeks

Michael Kratsios just put a name on Washington's Kimi K3 panic — and the calendar may be his weakest witness. On July 22, the White House science chief accused Moonshot AI of covertly distilling Anthropic's Fable to build Kimi K3, training on smuggled Nvidia GB300 chips in Thailand. Treasury doubled down hours later. Researchers say the timeline doesn't work.

What Kratsios alleged:

• "We have information that Moonshot AI distilled Anthropic's Fable for the development of its K3 model," Kratsios wrote on X July 22.
• He said Moonshot built "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."
• Kratsios added Moonshot "has acquired GB300-equipped servers and has accessed GB300s in Thailand, likely to train its AI models" — implying a workaround of U.S. export controls on Nvidia's Blackwell-generation chips.
• He drew a line: "Legitimate AI distillation used to create smaller, more efficient models plays a vital role in this open innovation ecosystem. However, large-scale, covert industrial distillation aimed at stealing proprietary U.S. technology and undermining American research is unacceptable."
• Kratsios did not publish evidence or explain how the government reached its conclusion. Moonshot has not responded. The Chinese embassy in Washington has not commented.

Treasury followed within hours:

• "Open source is not open season on American IP," Treasury Secretary Scott Bessent posted on X July 22.
• "When [Chinese] firms conduct covert, industrial-scale distillation attacks that cross the line into IP theft, sanctions and Entity List designations will be on the table."
• Bessent had warned July 21 that officials were "finding watermarks of our U.S. large language models on many of the Chinese models" — without specifying what those watermarks are.

The timeline problem:

• Anthropic's Claude Fable 5 became publicly available July 1, according to TechCrunch.
• Moonshot released Kimi K3 — a 2.8-trillion-parameter open-weight model — on July 16, during Shanghai's World Artificial Intelligence Conference.
• "I don't think you get a model this strong and this quickly on the heels of Fable doing strictly distillation," Braden Hancock, a researcher at the Laude Institute and co-founder of Snorkel AI, told TechCrunch. "There's just not even frankly time, right? Fable's only been publicly available since July 1st. You can't distill that much data, train a model, and release it in two weeks."
• Nathan Lambert, an AI researcher at the Allen Institute for AI, said on his podcast that distillation "has become less and less impactful over time as the Chinese models get closer to the frontier and the training regime shifts to reinforcement learning."
• Advanced distillation at frontier scale would require tens of millions of agent runs — "insanely expensive" via API and "a time bottleneck," Lambert said.

The older case is stronger:

• In February, Anthropic publicly accused Moonshot AI, DeepSeek, and MiniMax of industrial-scale distillation — roughly 16 million exchanges through about 24,000 fraudulent accounts.
• In June, Anthropic told the Senate Banking Committee that Alibaba's Qwen lab ran 28.8 million exchanges through 25,000 fake accounts — the largest known distillation campaign Anthropic has documented.
• Kratsios did not say whether Washington's Fable claim rests on new forensic evidence or extrapolates from those earlier campaigns.
• On July 22, Sarah Heck, Anthropic's head of public policy, thanked Kratsios on X: "Illicit, adversarial distillation is IP theft and industrial espionage that supports adversary military and intelligence capabilities." Anthropic did not respond to TechCrunch's questions about Fable-specific distillation.

The chip-smuggling angle:

• Georgetown CSET research fellow Sam Bresnick told TechCrunch a black market exists for banned Nvidia chips, and that "know your customer" rules for global data centers remain thin.
• In May, Supermicro founder Charles Liang was indicted for allegedly smuggling advanced chips into China — a case Bresnick cited as evidence the supply chain leaks regardless of export controls.
• If Moonshot trained on GB300s in Thailand, the violation is hardware smuggling — a provable export-control crime — not a two-week distillation sprint.

Washington's enforcement stack:

• Reps. Andrew Garbarino (R-NY) and John Moolenaar (R-MI) launched a joint House investigation April 29 into PRC AI distillation, naming Moonshot alongside DeepSeek, Alibaba, and MiniMax.
• The probe targets "the redistribution of those stolen capabilities as open-weight models available for global download" — exactly what Kimi K3's July 27 weight drop will deliver.
• Sens. Bill Hagerty (R-TN) and Andy Kim (D-NJ) are drafting an NDAA amendment to blacklist firms found harvesting U.S. model outputs at scale.
• Moonshot is separately racing toward a reported $30 billion Hong Kong IPO — the same company whose K3 launch helped erase more than $3 trillion in global chip market value last week.

The hypocrisy nobody wants to litigate:

• Elon Musk testified this year that SpaceX's xAI distilled OpenAI models to develop Grok — and called the practice common in the industry.
• Hugging Face CEO Clem Delangue told TechCrunch: "We know distillation to be a very small factor in the ability to create good models, and it's a practice that everyone is doing, including companies in the U.S."
• Microsoft CEO Satya Nadella wrote in July that it is "ironic" for frontier labs to claim fair-use training rights on public data, then "impose restrictive terms on distillation."
• On July 20, a federal judge approved Anthropic's $1.5 billion copyright settlement after ruling book training is fair use — while pirated downloads never reached a jury.

What enterprises should price in:

• Named accusations — even without public evidence — are enough for procurement teams to freeze Chinese open-weight pilots before July 27.
• Sanctions risk attaches to vendors, not downloads. A Treasury designation turns a cost hack into a compliance violation overnight.
• The cheaper the model, the higher the geopolitical discount rate just became — whether or not the Fable timeline holds up in court.

Convina's view: Kratsios did what Bessent's watermark threat could not — he named Moonshot before the ZIP file ships. That alone re-prices Kimi K3 for any enterprise with a sanctions compliance desk. But Washington's specific Fable claim looks like a press-release timeline, not a training log: fourteen days is not enough to distill, train, and ship a 2.8-trillion-parameter frontier model. The stronger case is February's documented query campaign and whatever GB300 receipts Treasury may hold. Until those surface, treat this as enforcement theater ahead of a geopolitical product launch — not a proven heist.