Many MindedAmerican AI-data vendors quietly supply Chinese labs with training material
A ChinaTalk investigation reports that US data firms sell roughly $500 million of AI training data to Chinese AI labs. It identifies at least nine American companies with Chinese customers, naming Mercor, Surge AI, AfterQuery and Turing. Part of what they sell is RL environments: simulated settings used to train models on tasks such as coding, financial analysis and office work.
The appeal is practical. Chinese labs reportedly buy the same products that were earlier sold to Anthropic or OpenAI, so they know the data works and skip the cost of building it in-house and the waste of failed training runs. The report says compute savings matter especially because Chinese chips face acute memory bottlenecks. Labs also use distillation, taking outputs from US models such as GPT or Claude as teacher examples. Combined with purchased environments, that lets them narrow the gap with less compute.
For abundance, this is a case of advanced model-training know-how spreading beyond the few labs that first paid for it. Better training data bought off the shelf lowers the cost of reaching capable AI, which widens who can build it. It also touches security: the authors note the field is largely unregulated, with export controls barely applying, and urge the Bureau of Industry and Security to apply existing controls to dual-use environments such as coding and biology.
Caveats are real. The relationships surfaced through anonymous sources and subcontractor lapses, not firm disclosures, and ChinaTalk could not independently verify the AfterQuery sales or the Forbes revenue figure of about $100 million ARR. The authors stress the situation is not apocalyptic: Chinese labs still trail US peers by at least a few months structurally, and US frontier firms increasingly build sensitive environments in-house. Watch for any move on export controls.
Source: The Decoder