Trump talks export controls, Nvidia on historic trip to Far East

Chinese technology companies and researchers are increasingly using results from leading U.S. AI models to train smaller, less expensive domestic competitors, according to a Reuters review of more than 300 Chinese academic papers, patents and company disclosures.

The technique – known as distillation – uses a powerful “teacher” model (such as OpenAI’s GPT-4o or o1) to generate answers, code, and problem-solving steps, which are then harvested to form a smaller “student” system.

Faced with tightening U.S. export restrictions on advanced Nvidia processors, Chinese developers are automating this process to create high-performance models with far less computing power, supported by government policies encouraging lower-cost AI deployment in China’s manufacturing sector.

Papers from Alibaba, Meituan and Shanghai Jiao Tong University show that researchers systematically harvest thousands of results from OpenAI models to train domestic systems like Alibaba’s Qwen models – although OpenAI and Anthropic prohibit the use of their results to train competing AIs.

Washington sees industrial-scale distillation as the main driver of China’s AI models, narrowing the gap with U.S. leaders, creating a major flashpoint ahead of the APEC summit in Shenzhen between President Donald Trump and Chinese President Xi Jinping.

Following July bilateral negotiations in Chengdu, during which US officials warned that the practice could result in intellectual property theft, Treasury Secretary Scott Bessent warned Beijing.

“When (Chinese) companies carry out industrial-scale covert distillation attacks that cross the line into intellectual property theft, sanctions and Entity List designations will be on the table,” Bessent wrote on X in July.

Reuters contributed to this report.

Gn headline

Logo