US Data Brokers Sell $500 Million in RL Environments to Chinese AI Labs

At least nine American data firms have sold roughly $500 million worth of AI training data to Chinese AI laboratories, according to an investigation published by ChinaTalk. The sales include reinforcement learning environments — simulated worlds where models practice tasks such as coding, financial analysis, or office work — and the buyers are deliberately purchasing the same products previously delivered to OpenAI and Anthropic.

The scale matters. Half a billion dollars is not a fringe trade in scraped text. It is a structured supply chain, and it runs in one direction: from American brokers to Chinese model developers. ChinaTalk, the outlet that conducted the investigation, names Mercor, Surge AI, AfterQuery, and Turing among the US firms involved.

Why do Chinese labs want the identical environments?

The logic is efficiency. Chinese labs, the report states, deliberately buy the same training environments that were built for American frontier companies. That gives them two guarantees: the data demonstrably works, and they can skip costly in-house development and wasted training runs.

Compute economics sharpen the incentive. Chinese chips suffer from acute memory bottlenecks, according to the report, which makes every avoided training run and every optimized environment disproportionately valuable to labs working under hardware constraints. What an American frontier lab treats as a convenience, a Chinese lab treats as a budget line.

A second mechanism compounds the advantage: distillation. Chinese labs use powerful US models such as GPT and Claude as "teachers," taking their outputs as training examples. Purchased RL environments plus distillation form what the report describes as a two-pronged advantage — one that lets Chinese labs close the capability gap with less compute of their own.

Who is selling, and how much?

ChinaTalk's authors identified at least nine US data firms selling to Chinese customers. The documented details vary in precision:

• AfterQuery reportedly draws a significant share of its annual recurring revenue from Chinese customers. ChinaTalk points to Forbes reporting of about $100 million in ARR but states it could not independently verify the sales.
• Surge AI openly advertises on its website that its products are "built by the same experts behind our frontier-lab work" — a marketing claim that doubles as an admission of reuse across customers, including Chinese ones.
• ByteDance, the Chinese buyer named in the report, maintains procurement offices in Singapore and San Jose, routing deals through offshore structures.

The trade came to light, the authors say, through a combination of anonymous sources and sloppy operational security among subcontractors. The mechanics are instructive. US data firms outsource much of the actual RL environment creation and data curation to low-paid workers in developing countries, especially India. Those subcontractors are supposed to refer to clients only by code names. Instead, they sometimes post client names directly on LinkedIn or in résumés — the operational leak that helped ChinaTalk map the relationships.

What does the regulatory picture look like?

Almost empty. RL environments remain largely unregulated, and export controls barely apply to them, according to the report. That gap is the policy stake at the center of the story: the United States has built an export control regime around advanced chips, while a parallel pipeline of training infrastructure — data, environments, curated tasks — moves across borders with essentially no oversight.

The ChinaTalk authors call for closing that gap. They argue the Bureau of Industry and Security, the Commerce Department agency that administers US export controls, should subject dual-use training environments to existing control mechanisms.

Which domains raise the sharpest concerns?

Not all RL environments carry equal risk. The report singles out domains with dual-use potential:

• Coding environments improve software development capabilities but can also sharpen offensive cyber capabilities.
• Biology environments could improve gene synthesis capabilities and potentially enhance bioweapons-relevant applications.

These are the categories where the authors see a concrete case for export controls — not a blanket restriction on training data, but targeted limits on environments whose mastery translates directly into national-security-relevant capability.

How large is the actual threat?

The authors themselves apply the brake. These relationships are "not apocalyptic," they stress. Chinese labs that piggyback on US work still trail their American peers by at least a few months as a structural matter, because buying existing environments means operating on someone else's timeline.

There is also a market response already underway. Sensitive environments in areas like biology, chemistry, cybersecurity, chip design, and investment banking are increasingly being built in-house at American frontier companies, the authors say. The most dangerous categories of training infrastructure are, in effect, being pulled off the open market by the firms that create them — not because regulators forced them, but because frontier labs now treat those environments as proprietary.

That leaves the policy question open. The $500 million figure documents what has already moved. Whether the Bureau of Industry and Security acts on dual-use environments before the next tranche ships will determine if this trade remains a regulatory blind spot or becomes the next front in US-China technology controls.