ChinaTalk and Forbes report that U.S. data vendors supply training services to Chinese AI labs, including some reinforcement-learning environments, while the reported $500 million annual spending estimate covers a broader category.
Forbes reported that the six largest Chinese AI labs spend about $500 million a year with U.S. data-labeling firms. Two entrepreneurs told the outlet they received the estimate from executives at Tencent and ByteDance. The figure covers data-labeling services broadly; Forbes did not provide a separate amount for RL environments.
ChinaTalk’s Nov. 11 investigation named nine companies in connection with Chinese lab work: Mercor, Surge AI, Turing, AfterQuery, Abundant, Pareto, DeccanAI, Eaglepoint and Invisible. The outlet said the type of work and the evidence it reviewed varied across the companies; the list does not establish that all sold RL environments.
ChinaTalk said data it reviewed indicated that Mercor sold software-engineering RL environments to Alibaba. Forbes had earlier described communications suggesting purchase arrangements between Mercor and an Alibaba employee, without identifying what was bought.
In a Sept. 1 guide, Mercor said specialists prepared 1,928 tasks to post-train Qwen3.5-397B-A17B. The company reported that Pass@1 on a separate 480-task benchmark rose from 16.11% to 27.29%. The guide describes Mercor’s training work but does not identify an Alibaba purchase.
An RL environment gives a model tasks in a simulated workplace and feedback on its performance, according to Mercor’s guide. ChinaTalk said vendors may adapt or reuse environments, while Forbes described standardized datasets that can be offered to multiple buyers. Forbes also reported that ready-made data may reduce the time Chinese labs spend testing training approaches.
Separately, SemiAnalysis said it believed Surge AI sold RL environments to Moonshot and Z.ai and that access to them played a major role in improving Kimi K2 Thinking and GLM-4.6. ChinaTalk flagged possible dual-use risks in environments involving cybersecurity or biology, while saying its investigation did not establish that reported sales covered those areas. It also cautioned that coding work alone does not necessarily give a model cyber capabilities.
