US Data Brokers Sell Chinese AI Labs $500M Worth of Data and RL Environments

If you thought chips were the main bottleneck in the AI race, this story has a second half. A ChinaTalk investigation, retold by The Decoder, shows that American data suppliers are quietly selling Chinese labs the material that OpenAI and Anthropic models are trained on. The total volume is about $500 million.

What exactly is being sold

This is not just about labeled text. Part of the product is RL environments (reinforcement learning environments). Think of a training simulator: the model is given a task such as "fix a bug in a repository" or "consolidate a quarterly report," and a script checks the result and issues a reward. The model tries thousands of times and gradually gets better at such tasks. Building a good simulator is expensive and slow: it requires experts, verification and debugging.

That is why the arrangement benefits both sides. According to the investigation, the list of sellers includes at least nine American companies, among them Mercor, Surge AI, AfterQuery and Turing. Surge AI states directly on its website that its products are made by "the same experts behind the work for frontier labs."

Why China wants it

Buyers deliberately choose the same products that OpenAI or Anthropic bought earlier. The logic is simple: since the American leaders have already proven that an environment works, there is no need to test it yourself. This saves months of development and eliminates "empty" training runs, which are very expensive. For Chinese labs, compute is doubly valuable: as the authors write, local chips suffer from an acute memory shortage.

Distillation adds to this. Chinese teams use strong American models such as GPT or Claude as "teachers" and train their own models on their outputs. Combined with purchased environments, this yields a double gain: catching up with fewer resources of their own.

The deals themselves run through offshore structures. ByteDance, for example, has procurement offices in Singapore and San Jose. It all came out through anonymous sources and contractor carelessness. American firms hand the bulk of the work of building environments and preparing data to low-paid contractors in developing countries, primarily India. These contractors sometimes list client names on LinkedIn or in résumés, even though they are supposed to use only code names.

Where the risk is and where it is overstated

The main concern is dual-use environments. Coding simulators improve software development but can also strengthen offensive cyber capabilities. Biology environments can improve work with gene synthesis and could potentially help with tasks related to bioweapons. Today export controls barely apply to such goods, because there is no separate regulatory category for RL environments. The authors propose that the US Bureau of Industry and Security apply existing rules to them.

At the same time, they themselves ask readers not to slide into apocalyptic thinking. As long as China follows in the footsteps of American solutions, it structurally trails American labs by at least several months. Sensitive environments in biology, chemistry, cybersecurity, chip design and investment banking are increasingly being built inside American companies themselves.

The limitations are also worth remembering. The $500 million figure is the investigation's estimate. The figure of roughly $100 million in annual revenue for AfterQuery comes from Forbes publications, and ChinaTalk could not verify it independently.

What this means in practice

For founders of data startups, the takeaway is blunt: regulation of this market will most likely tighten. It is worth finding out in advance who your end buyers are, how the contracts are structured and who receives environments in sensitive domains. This is my conclusion, not a claim made by the investigation's authors.

For those building products on top of models, the point is different. Training simulators and data alone are a weak moat: if they can be bought, competitors will buy them too. The real advantage remains in proprietary data, distribution and iteration speed. And teams that trust contractors with data collection or labeling must watch the operational security of the chain: a client name leaking from a contractor's résumé is a perfectly realistic scenario.