China Scales Up AI ‘Distillation’ of U.S. Models to Offset Semiconductor Restrictions
Chinese technology firms and researchers are increasingly leveraging outputs from premier American artificial intelligence systems to train domestic competitors, employing sophisticated methods to circumvent tightening U.S. semiconductor export controls. According to a Reuters analysis of over 300 Chinese academic studies, patents, corporate filings, job listings, and state directives, the technique—known as distillation—has evolved from a laboratory concept into a widespread industrial practice across China.
Distillation operates by extracting answers, computer code, and complex problem-solving demonstrations from a powerful "teacher" model to educate a smaller, more economical "student" system. While the legal and globally common practice cannot replicate a flagship model's full capabilities or entirely replace advanced hardware, it allows developers with limited computing power to maximize the value of single, costly queries. A former Alibaba Group engineer noted that Chinese teams can capture an American model's solution to a complex prompt once, reuse the material, and systematically refine local systems. Chinese companies have even built automated software and hired specialized personnel to curate this training data, supporting Beijing’s broader push to deploy affordable AI throughout its massive manufacturing sector.
This industrial-scale adoption threatens to blunt the impact of U.S. restrictions on advanced chips from suppliers like Nvidia, which Washington moved to tighten further on Monday. Consequently, the practice has become a major diplomatic flashpoint 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 scheduled to meet. In September, U.S. officials publicly accused Chinese enterprises of aggressive, targeted distillation. Earlier, during July regulatory talks in Chengdu, U.S. Commerce Department representatives warned Chinese counterparts that systematic output harvesting could constitute intellectual-property theft, though Chinese officials defended the method as standard industry practice without offering firm commitments to curb it. Treasury Secretary Scott Bessent subsequently warned on the social platform X on July 23 that covert distillation crossing into IP theft could trigger sanctions and Entity List designations.
Academic publications illustrate how deeply embedded the technique has become, even though terms of service from U.S. developers like OpenAI and Anthropic generally prohibit using their outputs to build rival products. In November 2024, Shanghai Jiao Tong University researchers detailed how they translated roughly 5,000 responses from OpenAI’s o1 model via GPT-4o mini to create 10,750 bilingual examples for training Alibaba’s Qwen2.5-72B model, though they cautioned that overreliance on the method could stifle homegrown innovation. By April 2025, Alibaba researchers were using GPT-4 and GPT-4o to generate, score, and refine non-Chinese training datasets for smaller Qwen variants. Similarly, a February 2026 study by Shanghai Jiao Tong University and food-delivery giant Meituan outlined using GPT-4o to teach smaller models how to coordinate AI agents on coding tasks.
While a U.S. official attributed China's rapid AI progress primarily to industrial distillation, sources familiar with Chinese operations argue it is merely one component alongside improvements in algorithms, engineering, and data curation. Although Beijing stops short of explicitly ordering the distillation of American systems, a January AI+Manufacturing initiative from eight central agencies—alongside local directives in Shanghai and Shenzhen—actively promotes distillation to lower computing costs. Aware of mounting geopolitical risks, domestic developers have recently sought clarity from China’s Ministry of Industry and Information Technology regarding the legal boundaries of harvesting foreign model outputs.











