Glonce

US data firms sell about $500M of AI training data to Chinese labs, ChinaTalk reports

policyThe Glonce editorsThe Decoder🔥62Glonce rating62Impact50Novelty60Relevance85Surprise55Players35Credibility65
US data firms sell about $500M of AI training data to Chinese labs, ChinaTalk reports

According to an investigation by ChinaTalk, as reported by The Decoder, American data firms such as Mercor, Surge AI, AfterQuery and Turing are selling roughly half a billion dollars' worth of AI training data to Chinese AI labs. The investigation identifies at least nine US companies involved. Some of that data takes the form of reinforcement learning (RL) environments: simulated worlds where AI models are trained on specific tasks such as coding, financial analysis or office work.

The report says Chinese labs deliberately buy the same products that were previously sold to Anthropic or OpenAI. That way they already know the data works, and they can skip costly in-house development and wasted training runs. Because Chinese chips suffer from acute memory bottlenecks, those savings are particularly valuable to labs there. Distillation adds to the effect: Chinese labs use powerful US models such as GPT or Claude as "teachers" and train on their outputs. Combined with purchased RL environments, this gives Chinese labs a two-pronged advantage that lets them close the gap with less compute of their own.

On individual firms, AfterQuery reportedly draws a significant share of its annual recurring revenue from Chinese customers. ChinaTalk points to Forbes reports of about $100 million in ARR for the company but 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." The deals also run through offshore structures, and ByteDance maintains procurement offices in Singapore and San Jose.

The relationships came to light through a mix of anonymous sources and sloppy operational security among subcontractors, according to ChinaTalk. The US firms outsource much of the actual RL environment creation and data curation to low-paid workers in developing countries, especially India. Those subcontractors sometimes post client names on LinkedIn or in résumés, even though they are supposed to use only code names. Export controls barely apply, because RL environments remain largely unregulated.

The authors single out dual-use domains. Coding environments improve software development but can also sharpen offensive cyber capabilities. Biology environments could improve gene synthesis capabilities and potentially enhance bioweapons-relevant applications. The ChinaTalk authors argue that the Bureau of Industry and Security should subject these dual-use environments to existing export controls.

The authors also stress that the relationships are not apocalyptic. By piggybacking on US work, Chinese labs still trail their US peers by at least a few months as a structural matter. And sensitive environments in areas such as biology, chemistry, cybersecurity, chip design and investment banking are increasingly being built in-house at American frontier companies, the authors say.

Key facts

  • A ChinaTalk investigation identifies at least nine US data firms, among them Mercor, Surge AI, AfterQuery and Turing, selling roughly $500 million worth of AI training data to Chinese labs; some of it is RL environments.
  • Chinese labs reportedly buy the same environments previously sold to OpenAI or Anthropic, which cuts their development time and compute costs; this matters more because Chinese chips have acute memory bottlenecks.
  • AfterQuery reportedly draws a significant share of its annual recurring revenue from Chinese customers; ChinaTalk cites Forbes reports of about $100 million in ARR but could not verify the sales.
  • The authors argue the Bureau of Industry and Security should apply existing export controls to dual-use environments such as coding and biology, since RL environments are largely unregulated.
  • The authors say the situation is not apocalyptic: Chinese labs still trail US peers by at least a few months, and the most sensitive environments are increasingly built in-house at American frontier companies.

Why it matters

The report describes a trade that sits outside current export controls: RL environments remain largely unregulated, so the sales barely trigger them. It also names a mechanism by which Chinese labs save compute. By buying environments already proven with Anthropic or OpenAI, they skip costly in-house development and wasted training runs, which the report says is especially valuable given the acute memory bottlenecks of Chinese chips. Paired with distillation from US models, the authors describe a two-pronged advantage that lets Chinese labs close the gap with less compute of their own.

Who it affects

US data firms named in the report (Mercor, Surge AI, AfterQuery and Turing, among at least nine) are the sellers. Chinese AI labs are the buyers, though the report as relayed does not say which labs, beyond noting that ByteDance maintains procurement offices in Singapore and San Jose. US frontier companies such as OpenAI and Anthropic are affected as the original customers whose products are reportedly resold to Chinese buyers. The Bureau of Industry and Security is the body the authors want to act. Low-paid subcontractors in developing countries, especially India, do much of the actual environment building and data curation.

How to use it

This is a report, not a product, so there is nothing to adopt. The practical takeaway is the authors' proposal: the Bureau of Industry and Security should subject dual-use environments, particularly in coding and biology, to existing export controls. Readers following export policy can use the report's picture of how the trade works (offshore procurement offices, subcontractors working under code names, outsourced environment creation) as a map of where any control would have to bite.

How solid is it

This is a secondhand account: The Decoder summarises a ChinaTalk investigation, and the original is not reproduced here. The relationships are said to have surfaced through anonymous sources and subcontractors' sloppy operational security, such as client names posted on LinkedIn or in résumés. ChinaTalk could not independently verify the sales behind the roughly $100 million ARR figure for AfterQuery, which comes from Forbes reports. The $500 million figure is not broken down by period, and only some of the data is described as RL environments. No reaction from the named firms, the US government, OpenAI or Anthropic is reported.

Risks and caveats

The authors flag dual-use risk: coding environments can sharpen offensive cyber capabilities, and biology environments could improve gene synthesis and potentially enhance bioweapons-relevant applications. They also temper the alarm. The relationships are not apocalyptic, Chinese labs still trail US peers by at least a few months as a structural matter, and sensitive environments in biology, chemistry, cybersecurity, chip design and investment banking are increasingly built in-house at American frontier companies. The report says the trade is largely unregulated; it does not say the sales are illegal.

“built by the same experts behind our frontier-lab work.”

— Surge AI, on its website, as quoted by The Decoder