U.S. data companies are selling training materials to Chinese AI labs, including simulated environments used to teach models how to perform specific tasks. A ChinaTalk investigation estimates the trade at roughly $500 million and identifies at least nine U.S. firms involved. The findings raise questions about how these products may affect AI development and whether some should face export controls.
What reinforcement learning environments provide
Reinforcement learning environments are simulated settings where an AI model practices tasks such as coding, financial analysis, or office work. They give developers a structured way to train models and assess how they respond to particular challenges.
According to the investigation, Chinese labs buy some environments that had previously been sold to OpenAI or Anthropic. Purchasing products with an established track record can help buyers avoid building comparable materials themselves and reduce the risk of spending resources on unproductive training runs.
The report says these savings matter in part because Chinese chips face acute memory bottlenecks. Buying existing training environments can therefore help labs make more of the computing resources available to them.
Data sales and model distillation
The investigation names data firms including Mercor, Surge AI, AfterQuery, and Turing. It says AfterQuery receives a significant share of its annual recurring revenue from Chinese customers, while noting that the investigators could not independently verify the reported sales. Forbes reports about $100 million in ARR.
Surge AI says on its website that its products are “built by the same experts behind our frontier-lab work.” The investigation also describes deals routed through offshore structures. ByteDance maintains procurement offices in Singapore and San Jose.
Purchased environments are one way Chinese labs can build AI capabilities. The report also points to distillation: using outputs from powerful U.S. models such as GPT or Claude as examples for training another model. Combined with ready-made environments, this can help labs develop models with less computing of their own.
Why the trade is hard to see
ChinaTalk says it identified the relationships through anonymous sources and operational security lapses among subcontractors. U.S. data firms outsource much of the environment-building and data curation to lower-paid workers in developing countries, especially India.
Some subcontractors reportedly disclosed client names on LinkedIn or in résumés, despite instructions to use code names. These details helped reveal business links that might otherwise have stayed out of view.
The investigation says reinforcement learning environments remain largely unregulated, leaving export controls with little reach over this trade. The authors argue that the lack of oversight matters most when the training material concerns fields with potential civilian and security applications.
Where the security concerns lie
Coding environments can strengthen software development skills, but the authors say they could also improve offensive cybersecurity capabilities. Biology environments may advance gene synthesis and could contribute to applications relevant to bioweapons.
ChinaTalk's authors call on the Bureau of Industry and Security to apply existing export controls to dual-use environments. Their argument focuses on the possible capabilities these materials help develop, rather than treating every training product as equally sensitive.
The report also sets limits on the concern. It says Chinese labs that build on U.S. work still trail their U.S. peers by at least a few months as a structural matter. The authors add that American frontier companies increasingly develop sensitive environments in-house in areas including biology, chemistry, cybersecurity, chip design, and investment banking.
The central policy question is how to distinguish ordinary AI training materials from environments that could support sensitive capabilities. The investigation describes a large and partly opaque market; the authors' proposed response is to bring dual-use environments within export controls while recognizing that the sales do not erase the development gap between U.S. and Chinese labs.