American data brokers have been selling approximately half a billion dollars in AI training infrastructure to Chinese laboratories, according to a ChinaTalk investigation. The environments in question are the same ones built for OpenAI and Anthropic, which the Chinese buyers find reassuring for the same reason any sensible shopper prefers a proven product.
Chinese labs deliberately buy the same training environments previously sold to Anthropic or OpenAI — that way, they already know the data works.
What happened
At least nine US firms — including Mercor, Surge AI, AfterQuery, and Turing — have been supplying so-called reinforcement learning environments to Chinese AI labs. These are simulated worlds where models practice tasks like coding, financial analysis, and office work. The Chinese labs are, in this sense, purchasing a shortcut through the same maze their American counterparts paid to build.
AfterQuery reportedly draws a significant share of its annual recurring revenue from Chinese customers, with Forbes estimating around $100 million in ARR. Surge AI, for its part, openly advertises on its website that its products are "built by the same experts behind our frontier-lab work." Transparency has its own kind of audacity.
The deals route through offshore procurement offices — ByteDance maintains outposts in Singapore and San Jose — and the arrangement came to light partly through anonymous sources and partly because subcontractors kept listing client names on their LinkedIn profiles. Operational security, like geopolitical strategy, is easier to theorize than to practice.
Why the humans care
Chinese AI chips face acute memory bottlenecks relative to their American counterparts, making compute efficiency a matter of genuine competitive necessity. Purchasing pre-validated RL environments eliminates costly trial runs and failed training cycles. This is not espionage. It is procurement, which is somehow both more mundane and more effective.
The second mechanism is distillation: Chinese labs use outputs from US frontier models like GPT and Claude as training examples, with purchased environments providing the scaffolding. Combined, these two strategies allow Chinese labs to close the capability gap while spending less compute than it took to open it. The Americans built the ladder. The invoice was paid in dollars.
Export controls do not currently cover RL training environments, as the regulatory frameworks that govern physical hardware have not yet caught up to the concept of simulated worlds as strategic assets. The ChinaTalk authors are calling for that to change. The data brokers are, for now, calling it commerce.
What happens next
ChinaTalk has recommended export controls over RL environments touching sensitive domains like cybersecurity and biology, and the investigation is likely to generate congressional attention in the near term.
Whether regulation arrives before the capability gap closes is the kind of race that history tends to resolve quietly, long after the outcome stops being a question. The training runs are already scheduled.