US data companies sold Chinese AI labs $500 million worth of RL training data

An investigation by ChinaTalk found that at least nine American data companies — including Mercor, Surge AI, AfterQuery, and Turing — have sold reinforcement learning (RL) training environments and training data to Chinese AI labs worth a combined total of approximately $500 million. These are simulated environments for training models on specific tasks, such as programming, financial analysis, or office work. According to the report, Chinese labs deliberately buy the very same products that were previously sold to OpenAI or Anthropic, which allows them to verify the data's functionality and save their own compute capacity and development time — an advantage that matters all the more given that Chinese chips suffer from limited memory.

Part of the practice is distillation: Chinese labs use the outputs of GPT and Claude models as a "teacher" for training their own models, which, combined with the purchased RL environments, according to the authors creates a dual advantage that lets them catch up with American competitors at lower cost to themselves. According to a Forbes report (which ChinaTalk could not independently verify), AfterQuery is said to draw as much as around $100 million in annual recurring revenue from Chinese clients. Surge AI publicly advertises on its website that its products are "built by the same experts behind work for frontier labs." According to the report, the deals go through offshore structures; ByteDance has procurement offices in Singapore and San Jose. The business relationships were flagged by anonymous sources and by lax security on the part of subcontractors, who often outsource work on the environments to low-income workers in developing countries, particularly India — who sometimes list client names on LinkedIn even in violation of non-disclosure requirements.

According to the authors, export of training data remains essentially unregulated. They see a particular risk in dual-use areas: programming environments can also improve offensive cyber capabilities, and biology environments can improve gene synthesis with a possible impact on capabilities relevant to biological weapons. The authors therefore recommend that the US Bureau of Industry and Security bring these dual-use environments under existing export controls.

At the same time, the authors stress that the situation is not catastrophic — because they are imitating American work, Chinese labs still remain structurally several months behind their American competitors. According to the report, American frontier companies are increasingly developing sensitive environments in areas such as biology, chemistry, cybersecurity, chip design, or investment banking in-house.