The Trump administration does not have one policy on Chinese open-weight models. It has several, and over the past few days they have collided very publicly in a pretty big way.
Within days of Moonshot releasing Kimi K3, the White House science chief accused Moonshot of theft, the Treasury Secretary threatened sanctions, and the president’s most influential AI adviser basically told everyone to calm down.
The best line came from inside the administration itself. A senior official told Wired:
“It’s not an argument with two sides, it’s an argument with 10 sides.”
Who Is Actually in Trump’s AI Brain Trust?
There is no single desk where China policy gets decided, and barely any coordination between the desks that think they own it. What the Trump administration does have, is a cast of officials with clashing instincts, all lobbying the same president.
At the hard end sits National Cyber Director Sean Cairncross, who helped build Trump’s 2 June executive order on frontier models and has made curbing distillation, the practice of training a model on a rival’s outputs, his priority. Wired notes that Cairncross is a former Republican campaign lawyer with no technical background, and now he is running the government’s national-security response to AI.
Treasury Secretary Scott Bessent has taken the most aggressive public line, and White House science chief Michael Kratsios has supplied the specific accusation:
At the loose end sits David Sacks, who left his role as AI czar in March but kept a direct line to the president and a 1.6-million-follower megaphone on X. He wrote that the “Kimi panic needs to stop” and that America stays ahead only by continuing to release.
He was backing Commerce Secretary Howard Lutnick, who had played down Kimi K3’s benchmarks and floated incentives for US labs to build their own open-weight models. Lutnick is no dove, though. His department runs export controls through the Bureau of Industry and Security, and it was Commerce that imposed the order which took Anthropic’s own Mythos-class models offline weeks earlier.
Above all of them is Chief of Staff Susie Wiles, who vets every proposal before it reaches the Oval Office. Bessent, Lutnick and Cairncross all take their cases to her. Her call is the one that counts. Ten sides, one desk.
What Are the Hardliners Claiming?
The hardline case treats Kimi K3 as evidence, not competition.
Kratsios accused Moonshot of distilling Anthropic’s Fable through a concealed internal platform built for large-scale extraction, and of acquiring NVIDIA GB300 servers, some accessed in Thailand, to train its models. He drew a line between routine distillation, which he called legitimate, and covert industrial-scale copying, which he called unacceptable.
Bessent went further on Fox Business, branding the practice IP “theft” and threatening sanctions or a place on the Entity List. He said officials keep finding watermarks from American systems inside Chinese ones. Undersecretary of State Jacob Helberg continued this line of thinking when he called it a “heist of invaluable American Intellectual Property.”
The chip half of the allegation matters as much as the model half. If Moonshot really did route around export controls through GB300s in Thailand, the story stops being about copying and becomes about smuggling, which is terrain Washington already knows how to fight on.
Where Is the Pushback Coming From?
The counter-argument is that a crackdown would do the closed labs’ competitive work for them. The Little Tech Association, representing close to 200 companies including Y Combinator and Proton, warned the administration that cutting off models available abroad would entrench incumbents and act as a tax on intelligence.
NVIDIA’S Jensen Huang told Axios that American firms should absolutely be allowed to use Chinese models, calling them “excellent”.
Sacks makes the competition subtext explicit. He has accused the closed labs of regulatory capture and explains Anthropic doesn’t want the competition. Critics go further, noting a ban on the cheapest rivals would land just as Wall Street scrutinises the two firms’ margins ahead of their IPOs, and that Anthropic and OpenAI are the ones lobbying hardest for it.
Does the Distillation Claim Hold Up?
Moonshot denies it, and its defenders lean on the timeline: a company employee mocked the idea of distilling a whole frontier model in the fortnight between Fable’s release and K3’s launch, and independent experts told SCMP they doubt distillation alone could explain the result.
But that speed is exactly what the hardliners find suspicious. Their case was never that Moonshot only had two weeks with Fable. Kratsios alleged a hidden system for repeated, disguised access, and Anthropic said earlier that Moonshot had pulled millions of Claude conversations through fake accounts.
Why Does the Policy Keep Fighting Itself?
The deepest problem is not China. It is that Washington’s positions cancel each other out.
The same administration now casting the theft of Fable as a national emergency is the one that throttled Fable in the first place. Wired reports that the previous head of the Commerce Department’s AI standards centre, Chris Fall, resigned after weeks spent pressing Anthropic to strengthen safeguards on Fable 5, and that his staff negotiated the protections that allowed the model back online. Weeks later, its capabilities are the thing Beijing is accused of stealing.
A clean ban is close to unenforceable anyway. Once weights are public they can be mirrored and run offline. Entity List action mostly raises the legal cost for US and allied companies that want to use the models openly. It does very little to Moonshot.
What Does This Mean for Europe?
A US crackdown would not bind Brussels, and that is the part European operators should watch.
Kimi K3 is cheap, capable and self-hostable, and European startups use it for the same reasons American ones do. The EU has no equivalent Entity List mechanism for models, and the EU AI Act treats open-weight general-purpose systems more leniently than closed frontier ones. Europe’s own Mistral signed the open-weights letter that Anthropic refused.
Even Bessent, the administration’s most aggressive voice, has reportedly told colleagues that the US, Europe and Japan need to build their own open-weight models faster or lose the race outright. So the hardliners’ actual remedy is more open models, not fewer.
That leaves Europe with a fork: follow a Washington ban and lose the cheap weights its own builders depend on, or keep them and split the Western position. Either way, the decision gets made for Europe, not by it.
What Should Europe Watch Next?
Three signals will show which side is winning.
- Whether Treasury or Commerce leads the next formal move.
- If any Chinese lab actually lands on the Entity List before Xi Jinping’s expected White House visit in September.
- And whether the White House ever issues guidance separating the distillation it calls legitimate from the distillation it calls theft.
Until then, the most honest description of US policy on Chinese open-weight models is the one the administration gave about itself. Not two sides. Ten.
See Also:
Why Did Anthropic Refuse to Sign the Open Weights Letter?
