A $500M Industry: How AI Data Brokers Are Selling Training Environments to China

The AI race has birthed a specialized $500 million industry centered on trading reinforcement learning (RL) environments. According to an investigation by ChinaTalk, data brokers are creating simulated environments to train artificial intelligence models and selling them to leading American research labs as well as buyers in China.

Who Supplies the Training Ground

Building high-quality training simulations—ranging from complex code repositories and financial data to routine office tasks—is time-consuming and expensive. Specialized data brokers such as Mercor, Surge AI, AfterQuery, and Turing have emerged to fill this demand, supplying top-tier clients like OpenAI and Anthropic.

However, the investigation identified at least nine US-based firms selling training environments to Chinese customers. Laboratories in China intentionally seek out products already used by top US teams. This strategy allows them to offset memory limitations in domestic chips and optimize compute time while training their own models.

Security Lapses and Regulation Gaps

Much of the heavy lifting in data preparation is outsourced to low-paid workers in developing countries, particularly India. Despite non-disclosure guidelines, subcontractors frequently list real client names on LinkedIn and resumes. Because export controls currently leave RL environments largely unregulated, the authors of the report are calling on the US Bureau of Industry and Security (BIS) to introduce stricter trade restrictions.