America's best AI models are quietly training their own rivals. A Reuters investigation published on 17 November 2026 found that Chinese universities, tech companies and industrial researchers have built a systematic pipeline for turning the outputs of leading American AI models into training data for smaller, cheaper Chinese models — and Beijing is encouraging it. A day earlier, Washington answered with its toughest chip rule yet. Here is what Reuters found, and why the two stories are one race.
What Reuters found
Reuters reviewed more than 300 Chinese academic papers, patents, corporate disclosures, recruitment notices and government documents. The picture it draws is not a single secret programme — it is industrial distillation, spreading from labs into factories and internet platforms.
- What distillation is. You train a smaller "student" model on examples produced by a larger "teacher" — the teacher might be GPT-4o, OpenAI o1 or Claude; the student might be Qwen, Kimi, or a specialised factory model. The student never becomes as capable as the teacher, but it keeps selected skills at a fraction of the cost.
- November 2024: researchers at Shanghai Jiao Tong University described collecting about 5,000 answers from OpenAI's o1 model, translating them with GPT-4o mini, and assembling 10,750 bilingual training examples to further train Alibaba's Qwen2.5-72B.
- April 2025: Alibaba researchers documented using GPT-4 and GPT-4o to generate multilingual training material, then refining, scoring and folding it into smaller Qwen systems.
- February 2026: researchers from Shanghai Jiao Tong University and Meituan described using GPT-4o to generate plans for coordinating AI agents on programming problems, helping train a smaller model to organise such tasks.
- October 2026: Reuters reported, citing The Information, that DeepSeek and Moonshot AI had relied heavily on outputs from OpenAI's GPT-6 Astra to train newer models — Moonshot's Kimi K3.1 Pro and DeepSeek's V4.1-Flash 1017.
- Beijing is encouraging it. A national AI+Manufacturing policy package published by eight Chinese government agencies in January includes guidance supporting model training, fine-tuning and distillation for industrial applications — part of making AI cheaper to deploy across China's manufacturing sector.
Is it illegal? No — and that is the point
Distillation is not illegal, and it is common all over the world. Reuters is careful on this: the process cannot reproduce a leading model's full capabilities, and it does not remove the need for advanced computing power. But it can preserve selected skills in smaller systems that cost less to run — and that reduces reliance on American technology. That is exactly why it undercuts US chip curbs: if you cannot buy the best chips, you can still borrow the best models' knowledge, one answer at a time.
Washington has noticed. US authorities publicly accused Chinese companies in September of aggressive, targeted distillation of American AI models on an industrial scale — naming six companies, including DeepSeek, Moonshot AI and Alibaba. And in July, US Commerce Department officials told Chinese regulators that collecting frontier model outputs could cross into intellectual-property theft. OpenAI's terms restrict using its outputs to develop competing models; Anthropic restricts training competing systems without permission.
Washington's answer: close the remote-compute loophole
On Monday 16 November, the US Commerce Department's Bureau of Industry and Security announced an interim final rule aimed at the other half of the problem — Chinese firms renting advanced chips abroad instead of buying them:
- Remote access now needs a licence. Data centres anywhere in the world that provide compute to entities headquartered in countries of concern, such as China and Macau, must now obtain a licence.
- Five countries flagged. Indonesia, Malaysia, Singapore, Thailand and Brazil are now classified as destinations at heightened risk of chip diversion — separate export licences are required to ship advanced chips there.
- Existing customers get 90 days. Data centres already hosting Chinese clients must obtain a licence within 90 days. The rule takes effect on publication in the Federal Register, with a 60-day public comment period.
Why this week matters
The two stories landed days before Presidents Trump and Xi are expected to meet at the APEC summit in Shenzhen on 18–19 November — and the Reuters findings threaten to complicate those talks. Step back and the pattern is clear: Washington's strategy has two levers, chips and model access, and Beijing is working around both. Chips are being rented overseas instead of bought; model capabilities are being distilled instead of built from scratch. Each new American rule closes one route, and each Reuters document shows another one already in use.
Bottom line
The AI race is no longer just about who builds the biggest model or the fastest chip. It is about whether the knowledge inside those models can be contained at all — and this week, both sides showed their hand. China is industrialising distillation, and America is tightening the rules around every chip its rivals can reach. Neither move ends the race; both raise its stakes.
Sources: Reuters investigation "China taps US AI models to train homegrown rivals as chip curbs bite", 17 November 2026, by Eduardo Baptista and Laurie Chen (review of 300+ documents), summarised in this explainer of the Reuters findings; US Commerce Department interim final rule, 16 November 2026, as reported in the 17 November daily briefing. Hero photo: the surface of a 12-inch silicon wafer, by 2x910, CC BY-SA 4.0, via Wikimedia Commons.
