The open-weight race just became a sanctions fight.

On Wednesday, the director of the White House Office of Science and Technology Policy publicly accused Moonshot AI of stealing Anthropic's Fable model to build Kimi K3. Within hours, the Treasury Secretary threatened sanctions and Entity List designation, the same instrument used against Huawei in 2019. It is the first time a senior US official has named a specific Chinese lab for copying a specific American model.

Then the awkward part. Researchers started doing the arithmetic, and an OpenAI executive publicly said he does not think the accusation holds up. Kimi K3's weights go fully public on Monday regardless. Here is the case, the counter-case, and what actually happens next.

Michael Kratsios, who directs the White House Office of Science and Technology Policy, laid it out on X. He said his office has information that Moonshot AI distilled Anthropic's Fable to develop K3, and that the company built a sophisticated internal platform to run large-scale distillation against US models, switching between multiple access methods to avoid detection.

He drew a deliberate line between legal and illegal versions of the practice. Distillation, training a smaller model on a stronger model's outputs, is common, legal, and useful at normal scale. What Kratsios described is industrial-scale covert extraction aimed at lifting proprietary US capability wholesale. He added a second allegation with separate legal weight: that Moonshot obtained Nvidia GB300 servers and accessed more of them in Thailand, Blackwell-generation hardware barred from sale to Chinese companies, which would make it an export-control matter as well as an IP one.

Treasury Secretary Scott Bessent escalated the same day. He said the administration can sanction foreign AI models built on improperly obtained American technology, claimed officials keep finding watermarks from US language models embedded in Chinese systems, and put sanctions and Entity List designations on the table. His framing was the quotable one: open source is not open season on American IP. Entity List placement would cut Moonshot off from American hardware, software, and cloud infrastructure entirely.

Here is where it gets complicated, and where most coverage stops short.

Fable 5 only returned to public availability on July 1, after the US export-control ban that took it offline was lifted. Kimi K3 launched on July 16. That is a fifteen-day window in which Moonshot would have had to run industrial-scale extraction against Fable and then train, tune, and ship a 2.8-trillion-parameter model.

Elie Bakouch, a researcher at Prime Intellect, said the quiet part plainly: with only fifteen days between the Fable ban lifting and K3's release, claiming K3's performance comes from Fable distillation does not make sense technically, even if distillation did occur. More striking, Dean Ball, OpenAI's head of strategic futures, said he does not believe K3's performance can be explained away by distillation or anything like it. That is an executive at Anthropic's biggest rival declining to pile on a Chinese competitor when handed a free shot.

The evidentiary picture is thin in other ways too. Kratsios did not say how the government knows any of this. Anthropic itself has not claimed it holds evidence tying K3 to Fable, though it did disclose in February that it traced roughly 3.4 million Claude exchanges to Moonshot, which is the groundwork this accusation is built on. Moonshot has not responded, and Beijing called the earlier version of these claims groundless. Representative Ted Lieu pointed out the obvious tension in the administration's position: it is threatening sanctions over stolen AI capability while still approving sales of high-performance AI chips to China.

None of that proves Moonshot is clean. The February trace is real, and the export-control allegation about GB300s is a separate claim that could stand on its own evidence. But the specific charge, that K3 is what it is because of Fable, is the one the timeline strains hardest to support.

🎧 OpenAI Wants to Run Your Support Desk

While Washington argued about stolen models, OpenAI quietly made a move that says more about where the money is going.

On Wednesday it launched Presence, an enterprise product for deploying voice and chat agents into real workflows like customer support, outbound sales, and internal IT requests. Each deployment is scoped to one job, the agent gets only the access that job needs, and the company sets the rules for approvals and human handoff. OpenAI's own English-language support line runs on it and resolves 75 percent of inbound issues without a human.

The real story is the business model. Presence is not self-serve: deployments are run by OpenAI engineers and approved integrators, which means OpenAI is now selling implementation services, not just tokens. That puts it in the same room as Accenture and Deloitte. Two caveats worth keeping, though. None of those performance numbers come with published methodology or independent audits, and a product built to keep agents inside their boundaries shipped the same week OpenAI disclosed that its own models broke out of theirs.

Strip out the politics and there is a real question underneath: what is the legal status of training on another model's outputs? Nobody has answered it. If Washington succeeds in establishing that distillation of American models counts as IP theft subject to sanctions, that reshapes what every lab on earth can legally train on, including US startups that quietly benchmark against and learn from frontier APIs. The precedent would not stop at Chinese companies.

For your own work, the practical read is narrower. Sanctions would hit Moonshot's access to American chips and cloud, not your ability to download weights that are already public, and K3's weights land Monday regardless of what Washington says this week. The thing worth watching is not whether Moonshot gets punished, it is whether "trained on outputs of a US model" becomes a legal category. That would touch far more than one Chinese lab.

The stories moving fast beyond today's headlines:

White House science and technology director Michael Kratsios accused Moonshot AI on Wednesday of covertly distilling Anthropic's Fable to build Kimi K3, alleging the company ran a purpose-built internal platform for large-scale extraction against US models while rotating access methods to avoid detection, and separately that it obtained banned Nvidia GB300 servers partly through Thailand. Treasury Secretary Scott Bessent threatened sanctions and Entity List designation, the instrument used against Huawei, saying open source is not open season on American IP. It is the first time a senior US official has named a specific Chinese lab for copying a specific American model. But the evidence is not public and the timeline is contested: Fable 5 only returned July 1 and K3 shipped July 16, a fifteen-day window that Prime Intellect researcher Elie Bakouch says cannot technically account for K3's performance, while OpenAI's head of strategic futures Dean Ball said he does not believe distillation explains the model either. Anthropic has not claimed evidence tying K3 to Fable, though it traced 3.4 million Claude exchanges to Moonshot back in February, and Representative Ted Lieu noted the administration is threatening sanctions while still approving advanced chip sales to China. Separately, OpenAI launched Presence, an enterprise product for deploying scoped voice and chat agents with company-set policies and a Codex-driven improvement loop, delivered as a managed service rather than self-serve, which moves OpenAI into implementation consulting alongside the model business. Kimi K3's full weights go public Monday regardless of how Washington proceeds, and the question that actually matters long term is whether training on another model's outputs becomes a legal category, because that precedent would not stop at Chinese labs.

Stay building. 🤖