On July 22, White House science and technology adviser Michael Kratsios accused Chinese AI startup Moonshot of covertly extracting capabilities from Anthropic's Claude Fable 5 through a process called distillation -- training a weaker model on a stronger rival's outputs without authorization -- to build its viral Kimi K3 model. Treasury Secretary Scott Bessent escalated immediately, posting on X that 'open source is not open season on American IP' and that sanctions and Entity List designations 'will be on the table' for firms conducting what he called 'covert, industrial-scale distillation attacks.'
The accusation didn't stop at distillation. Kratsios alleged Moonshot had acquired Nvidia GB300-equipped servers -- part of Nvidia's export-controlled Blackwell generation -- and accessed additional GB300s in Thailand, 'likely to train its AI models,' raising a separate and arguably more serious question about whether Moonshot circumvented US export-control rules entirely.
This is the latest escalation in a fight that started when Kimi K3, a 2.8-trillion-parameter open-weight model, landed with benchmark claims putting it close to Fable 5 and OpenAI's GPT-5.6 -- while reportedly costing a fraction to train. That combination single-handedly helped trigger a bear market in the Philadelphia Semiconductor Index days earlier, as investors questioned whether frontier-level performance really requires frontier-level compute spend, undercutting the scarcity thesis propping up Nvidia and AMD valuations.
Not everyone buys the distillation narrative at face value. Fable 5 has only been generally available since July 1, and several AI researchers have publicly pushed back on the idea that three weeks is enough time to distill a 2.8-trillion-parameter model this capable -- suggesting either Moonshot's own research progress is further along than Washington wants to admit, or the accusation is convenient cover for a broader export-control crackdown.
The competitive and political backdrop is important: this comes the same week Jensen Huang publicly argued that competitive Chinese AI, including open-weight models, grows the total compute market rather than shrinking Nvidia's share of it -- a framing directly at odds with the Treasury's escalating rhetoric. It also follows a pattern this year of Washington oscillating between wanting American labs protected from IP theft and wanting American chip companies to keep selling into the Chinese market.
For founders and investors, an actual sanctions action or Entity List designation against a major Chinese lab would be a meaningful escalation beyond existing chip export controls -- it would target the model and research layer directly, not just hardware, and could accelerate the bifurcation of the global AI stack into separate US- and China-aligned ecosystems that founders building cross-border AI products would need to navigate.
Watch for: whether the administration follows through with a formal Entity List designation rather than just rhetorical pressure; Anthropic's own public response, which has so far been notably muted; and whether independent researchers can actually verify or debunk the distillation claim on technical grounds, since right now this is a political accusation without a published forensic analysis behind it.