Chinese AI firms are using answers from top American models to train local rivals. A review of more than 300 research papers, patents, company records, job ads, and state files points to a sharp rise in this practice as US chip curbs make AI hardware harder to get.
The method is called distillation. A large AI model acts as a teacher, giving out text, code, or steps to solve tasks. A smaller model then learns from those results. It takes less compute power than training a large system from scratch.
Hardware costs add to the pressure. An eight-GPU NVIDIA H200 server can fetch $900,000 through China's gray market, with trade curbs and local rules making access hard.
One case came from Shanghai Jiao Tong University. Its team took about 5,000 OpenAI o1 replies, used GPT-4o mini to translate them, and made 10,750 examples in two languages. Those samples helped refine Alibaba's Qwen2.5-72B model. Yet the team warned that heavy use of this method could hold back fresh ideas.
Alibaba had another approach. Its researchers used GPT-4 and GPT-4o to rate and sort data in other languages, then trained smaller Qwen systems with the picks. In a separate effort, Shanghai Jiao Tong University and Meituan used GPT-4o to create plans for AI agents that work on code as a group.
Here's the catch. Distillation itself is legal and common across the AI field. But OpenAI and Anthropic ban the use of their systems to train rival models under their service terms. That leaves a clear gap between what the law allows and what those firms permit.
The dispute has reached officials as well. During talks in Chengdu, US trade staff warned Chinese regulators that using American AI output at scale might raise theft claims over intellectual property.
One US official linked China's gains in AI performance to large-scale distillation. Washington sees the gap with its leading models shrink, even as tighter chip curbs make access to advanced NVIDIA gear more costly.
For Beijing, access to model output offers another way to keep local AI work moving. Reports of B300 GPU smuggling add a separate concern. Whether the US will seek new laws to curb the training method remains unclear.
The method is called distillation. A large AI model acts as a teacher, giving out text, code, or steps to solve tasks. A smaller model then learns from those results. It takes less compute power than training a large system from scratch.
Hardware costs add to the pressure. An eight-GPU NVIDIA H200 server can fetch $900,000 through China's gray market, with trade curbs and local rules making access hard.
One case came from Shanghai Jiao Tong University. Its team took about 5,000 OpenAI o1 replies, used GPT-4o mini to translate them, and made 10,750 examples in two languages. Those samples helped refine Alibaba's Qwen2.5-72B model. Yet the team warned that heavy use of this method could hold back fresh ideas.
Alibaba had another approach. Its researchers used GPT-4 and GPT-4o to rate and sort data in other languages, then trained smaller Qwen systems with the picks. In a separate effort, Shanghai Jiao Tong University and Meituan used GPT-4o to create plans for AI agents that work on code as a group.
Here's the catch. Distillation itself is legal and common across the AI field. But OpenAI and Anthropic ban the use of their systems to train rival models under their service terms. That leaves a clear gap between what the law allows and what those firms permit.
The dispute has reached officials as well. During talks in Chengdu, US trade staff warned Chinese regulators that using American AI output at scale might raise theft claims over intellectual property.
One US official linked China's gains in AI performance to large-scale distillation. Washington sees the gap with its leading models shrink, even as tighter chip curbs make access to advanced NVIDIA gear more costly.
For Beijing, access to model output offers another way to keep local AI work moving. Reports of B300 GPU smuggling add a separate concern. Whether the US will seek new laws to curb the training method remains unclear.