China expands use of U.S. AI model distillation as chip curbs tighten
Chinese entities are increasingly using output from leading U.S. artificial intelligence models to train domestic systems, broadening a technique that helps offset tighter limits on access to advanced chips. The practice is spreading from research into mainstream industrial use and is becoming part of Beijing's wider push to make AI cheaper and easier to deploy across manufacturing.
Highlights
- Chinese companies are expanding commercial use of U.S. AI model distillation to train local models, leveraging outputs from OpenAI and other American systems.
- Broader distillation adoption could reduce the impact of tightened U.S. controls on advanced chips by enabling China to build AI capabilities at lower cost and compute requirements.
- Escalating U.S.-China friction includes public U.S. accusations of 'industrial-scale' distillation and regulatory discussions ahead of the November 18-19 APEC summit in Shenzhen.
How distillation is spreading across Chinese AI development
As first reported by Reuters, a review of more than 300 Chinese academic papers, patents, corporate disclosures, recruitment notices and government documents shows distillation is moving beyond academic testing into wider commercial application in China. The records describe researchers using U.S. models to generate answers, write code and solve difficult problems, while Chinese companies develop software to automate parts of the process and specialist workers prepare training material for smaller local models.Distillation involves a stronger "teacher" model generating answers or demonstrating tasks that developers use to train a smaller "student" model. For Chinese developers working under computing constraints, the method lets one costly query produce reusable examples that can then support local training at lower operating cost.
Documents reviewed in the report include a November 2024 paper from Shanghai Jiao Tong University researchers describing how about 5,000 answers from an OpenAI model were selected, translated and turned into bilingual training examples for further work on an Alibaba model. An April 2025 paper by Alibaba researchers describes using OpenAI models to generate multilingual training material before refining responses for smaller models, while a February 2026 paper from researchers at Shanghai Jiao Tong University and Meituan describes using American models to generate plans for coordinating AI agents.
OpenAI's terms restrict the use of its output to develop competing models, and Anthropic similarly restricts training rival systems without permission. Shanghai Jiao Tong University, Alibaba and Meituan do not respond to requests for comment cited in the report.
Chip controls, policy support and rising U.S.-China friction
China's broader use of distillation could weaken the effect of U.S. curbs on advanced chip sales because developers can still build selected AI capabilities by drawing on American models. Washington says on Monday it will further restrict China's access to such chips from companies including Nvidia, adding pressure on Chinese groups to find lower-cost ways to train and deploy systems.The issue is also adding tension ahead of the November 18-19 Asia-Pacific Economic Cooperation summit in Shenzhen, where U.S. President Donald Trump and Chinese President Xi Jinping are expected to meet. U.S. authorities publicly accuse Chinese companies in September of targeted distillation on an industrial scale, and U.S. Commerce Department officials raise similar concerns with Chinese regulators during meetings in Chengdu in July, according to people familiar with the discussions.
Chinese officials defend distillation as a standard AI development technique and say they are open to acting where conduct amounts to intellectual-property theft, but they make no concrete commitments, according to a U.S. official familiar with the exchange. Chinese government documents do not direct companies to distil American models specifically, but national and local policies encourage model training, fine-tuning and techniques that reduce computing needs, including guidance issued in January by eight government agencies and separate support from Shanghai and Shenzhen.
As geopolitical scrutiny grows, Chinese AI developers in recent months ask Beijing's industry regulator for guidance on where ordinary use of foreign models ends and organised collection of output for training begins, according to one person familiar with the discussions. That suggests the sector is trying to navigate both commercial demand for cheaper AI deployment and the rising legal and political risks attached to using foreign model output.
The latest update follows our earlier article on Nvidia resilience after Jensen Huang’s resignation, as OpenAI said AI demand remains strong.