In December 1996 the Commerce Department decided that encryption source code printed in a book could leave the United States freely. The same code on a floppy disk needed an export license. Phil Zimmermann had already found the gap. He'd been under criminal investigation in 1994 because his encryption program, PGP, had spread overseas, and in 1995 he published its source code as a book. His preface laid out the plan: "So, we're putting the PGP source code in a book, which may be scanned in with OCR (optical character recognition) software." Commerce also kept the right to control printed code if it turned out to be easy to scan. Somewhere, presumably, an official was holding a hardback and judging how scannable it was.

On Monday, the Bureau of Industry and Security, which is what Commerce's export office is called these days, closed a real pipe. Chinese firms had been renting American chips in data centers across Southeast Asia, and now they mostly can't. Good. But on the same day, Reuters published a review of more than 300 Chinese papers, patents, corporate filings and government documents. It shows that output from American AI models has become ordinary training material for Chinese ones. A question typed into an American chatbot has quietly become an export, and the export rules still only put a price on hardware.
My view: export controls can ration the machines that make intelligence, but they can't ration the intelligence itself. The most a leader can buy is time, plus a rival's habit of coming back to learn from it. The rules should protect that habit, not force the rival to grow out of it.
On silicon, the wall holds up. Epoch AI estimates that Huawei will make about 25 times less AI compute than Nvidia in 2026 and will stay about four years behind on chips through 2030. Their summary is fair: "While not stopping Huawei's progress entirely, export controls have made it far harder, slower, and more expensive for Huawei to compete." When DeepSeek tried to train its next model on Huawei's Ascend chips, it couldn't finish the run and went back to Nvidia. That's a working control.
Monday's rule extends it. Indonesia, Malaysia, Singapore, Thailand and Brazil are now named as an "unacceptable risk of diversion," so shipping advanced chips there needs a license. Overseas data centers selling a significant volume of compute to China-headquartered companies have 90 days to get a license or face future restrictions. The rule doesn't cover systems used by online retailers and social media companies, which the department says don't pose the same threat. (The people who brought us the recommendation algorithm may have opinions about that.) Singapore responded politely that it is "exploring how to facilitate trusted access for Singapore companies," which is diplomatic language for asking to be let back in. I don't think closing this route was a mistake. It was a real route, and the government says one Chinese lab used it and the other route at the same time.
We have information that Moonshot AI distilled Anthropic’s Fable for the development of its K3 model. To do this they developed 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. Moonshot AI has also acquired GB300-equipped servers and has accessed GB300s in Thailand, likely to train its AI models. The United States strongly supports the free and fair development of AI, including a thriving competitive ecosystem that spans frontier models, specialized systems, open-source frameworks, and open-weight models. 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.
View on XThe other route is the one no license reaches. A former Alibaba engineer told Reuters that the attraction for compute-starved Chinese developers is "making one expensive query do repeated work." The American model solves a hard problem once, and the Chinese team keeps the solution and trains on it as often as it likes. The hawks at CNAS say it more bluntly: distillation "spares Chinese developers' own limited compute for other uses." American data centers do the expensive thinking, and Chinese chips get spent on something else.
None of this is secret or new. In March 2023, Stanford researchers spent under $500 on OpenAI's API for 52,000 answers, trained a small model on them for under $100, and published a write-up.
Reuters' lead example is a paper from Shanghai Jiao Tong University, published in November 2024. In one experiment, the researchers took about 5,000 answers from OpenAI's o1, used GPT-4o mini to turn them into 10,750 Chinese-English training pairs, and fine-tuned Alibaba's Qwen2.5-72B on them. In a separate experiment, they trained a 72-billion-parameter math model on tens of thousands of o1-distilled samples and gave it the 2024 AIME, a 30-problem American math competition:
| Model | AIME 2024 (out of 30) |
|---|---|
| The distilled student | 13 |
| o1-preview | 12 |
| o1-mini | 21 |
The student beat one teacher by a single problem and finished eight behind the other. A footnote notes that, per OpenAI's terms of use, all of this was for research only. But the same paper says something more interesting. Its authors warn that distilled models are "naturally bounded by the capabilities of their teacher model," which creates a "ceiling effect." Reuters adds that the researchers worried heavy reliance on distillation would leave Chinese developers dependent on breakthroughs from more advanced providers.
So one Chinese paper shows both the leak and the leash. The answers get out, and the student stays tied to the teacher.
Washington's own documents already half-agree. The White House memo on distillation from April admits that distilled models don't match the full performance of the original. When it gets to threats, it threatens dependence, not walls:
"Foreign entities who build their AI capabilities on such fragile foundations should have little confidence in the integrity and reliability of the models they produce." Read that again. It doesn't say we will stop you. It says we can mess with what you're building on. That's leverage, and it only works if the student keeps showing up.
The September advisory from NSA, CISA and the FBI describes how students keep showing up. Chinese firms allegedly use a gray market of proxies called "transfer stations" to get around geographic restrictions. In practice that's password sharing at a national scale. The cure on offer isn't a license. Companies are advised to quietly give suspected distillers worse answers without telling them. Export control is now the waiter spitting in the soup. CNAS reports that when Anthropic released a new Claude model, MiniMax shifted its traffic to it within 24 hours. That's what a dependent student looks like: every new release sends them back to class.
The U.S. has evidence that foreign entities, primarily in China, are running industrial-scale distillation campaigns to steal American AI. We will be taking action to protect American innovation. These foreign entities are using tens of thousands of proxies and jailbreaking techniques in coordinated campaigns to systematically extract American breakthroughs. Foreign entities who build on such fragile foundations should have little confidence in the integrity and reliability of the models they produce. The U.S. government is committed to the free and fair development of AI technologies across a competitive ecosystem, from open-source to proprietary models. Read the memo:
View on XThe volumes are large. Anthropic reported 16 million exchanges, which Nathan Lambert puts at somewhere between 150 and 400 billion tokens. Epoch estimates DeepSeek-R1's entire supervised fine-tuning dataset at about 6.4 billion. A US official told Reuters that industrial-scale distillation is "the main reason" Chinese models narrowed the gap. Two people familiar with Chinese developers' work said it was just one factor alongside engineering, algorithms and data. Nobody has published a share.
Suppose the official is right. Then distillation is central to Chinese AI, and if it's central, China depends on it. That makes the leash stronger, not weaker. The real limit on my argument is a different one: the leash only exists where America controls the tap. The advisory's list of allegedly distilled models includes gpt-oss-20b, an open-weight model anyone can download. Once the weights are out, there's no tap to turn.
This is also where I part ways with the old sell-them-chips crowd, who'd like to claim the dependence argument for themselves. A GPU sold to China is gone for good, and it computes for whoever owns it. An answer from a frontier model is a one-off, and the next model's answers have to be fetched all over again.
Selling chips gives China the means to make its own answers. Selling answers keeps it a customer.
For years, US chip policy ran on what CSIS describes as a "sliding scale": keep China a fixed distance behind. In September 2022, Jake Sullivan dropped that for a goal to "maintain as large of a lead as possible." That's siege thinking, and for silicon I can see the case. Dario Amodei made the strongest version of it in January 2025, arguing that the frontier belongs to whoever has millions of chips. Notably, he set distillation aside: "I'm not taking any position on reports of distillation." He may have been right about who sets the frontier. He didn't settle who gets to use it.
For models, I'd go back to the sliding scale. Keep the gap, keep the tap American, and judge every rule by how much time it buys and how much dependence it keeps. If a rule pushes Chinese labs to stop coming back, it has given up the one lever that works on answers. (A year of not enforcing the old diffusion rule doesn't buy time. It spends it.)
The 1996 rule did delay strong encryption abroad. By January 2000 the controls on publicly available source code had mostly been relaxed, because text that people want will eventually leave the country. The difference this time is that Zimmermann's readers only had to scan his book once. Chinese labs have to keep buying the next edition.







Comments