By Hamza Faisal

American data companies are selling roughly half a billion dollars' worth of AI training data — including the simulated worlds used to teach models new skills — to Chinese AI labs, according to an investigation published on November 11 by ChinaTalk, the newsletter covering technology in China. The trade hands some of America's most capable AI rivals the same training tools built for its own frontier labs, and Washington's export controls barely touch it.

The investigation, by Aqib Zakaria and Quinn Ennis, identifies at least nine US vendors selling to Chinese customers. The named firms include Mercor, Surge AI, Turing and AfterQuery, alongside Abundant, Pareto, DeccanAI, Eaglepoint and Invisible. The buyers cited include Alibaba, Tencent, ByteDance, Moonshot AI and Z.ai.

The same tools built for OpenAI and Anthropic

Much of the trade is in reinforcement-learning environments — simulated versions of apps and tasks, paired with grading rubrics, that labs use to post-train models in coding, finance, law, cybersecurity and biology. They are the practice grounds where a finished model learns useful work.

Chinese labs deliberately buy the same environments previously sold to OpenAI or Anthropic, the report says. Buying proven tools lets them skip building the environments themselves and saves compute during model development. According to the investigation, Mercor sold software-engineering environments to Alibaba, while Forbes has reported that 2 percent of Mercor's second-quarter 2026 revenue came from China. Surge AI has been reported by SemiAnalysis and Forbes to sell environments to Tencent, Z.ai and Moonshot AI, and advertises on its website that its products are built by the same experts behind its frontier-lab work. Turing is reported to sell data to Tencent and Alibaba, while AfterQuery is reported by Forbes to draw about half of its roughly $100 million in annual recurring revenue from Chinese business — a figure ChinaTalk says it could not independently verify.

A gap export controls barely touch

The striking part is how little of this trade is regulated. Reinforcement-learning environments fall largely outside US export controls, except where the buyer is on the Entity List. The authors argue the Bureau of Industry and Security should restrict dual-use environments and that the Pentagon should tighten its contractor rules.

The dual-use concern is direct: environments built for coding and biology can sharpen a model's software skills or its grasp of gene synthesis, but the same tools could also aid cyberattacks or bioweapon research. The investigation does not publish a breakdown of the half-billion-dollar figure, and the companies' responses were not included in the report. The broad scale of the trade was reported earlier — Forbes first put the figure at about $500 million in August — but the ChinaTalk investigation is the first to name the vendors, the buyers and the mechanics.

Not apocalyptic, but a head start at stake

The authors are careful not to overstate the damage. Chinese labs still trail their American peers by months, they write, because they must wait for US labs to commission an environment before they can buy a copy of it. The trade is not closing the gap overnight.

Even so, the findings land at an awkward moment for Washington, with scrutiny of AI exports already intensifying after the resignation of Nvidia chief executive Jensen Huang and congressional probes into chip sales to China. The question now is whether regulators treat training environments — the data and simulations behind the models, rather than the chips themselves — as the next front in the technology contest. On the evidence of this investigation, that front has been open for some time.