AivexaNewsSearch
AI news for builders and product teamsChecked every hour

US data brokers are selling half a billion dollars in RL training environments to Chinese AI labs

Collected Nov 11, 2026

A ChinaTalk investigation has identified at least nine US data companies selling AI training data to Chinese AI labs, with the total value of the trade estimated at roughly half a billion dollars. The firms named in the report include Mercor, Surge AI, AfterQuery and Turing. The findings were published by The Decoder, which summarizes the ChinaTalk work. A central element of the trade is so-called reinforcement learning environments: simulated worlds in which a model is trained and rewarded on specific tasks such as coding, financial analysis or office work.

According to the report, Chinese labs deliberately purchase the same RL environments that were previously built and sold to Anthropic or OpenAI. Buying a product already validated by a frontier lab lets a buyer skip the costly process of designing environments in-house and avoid wasted training runs on data that turns out not to work. The report says those savings are especially valuable for Chinese labs because Chinese chips suffer from acute memory bottlenecks, which makes compute efficiency a binding constraint rather than a nice-to-have.

The investigation describes a second mechanism working alongside the purchased environments. Chinese labs use powerful US models such as GPT or Claude as teachers, feeding their outputs back in as training examples, a practice known as distillation. The report frames purchased RL environments and distillation as a two-pronged advantage that lets Chinese labs narrow the gap while spending less of their own compute.

On the scale of the business, ChinaTalk reports that AfterQuery draws a significant share of its annual recurring revenue from Chinese customers. The report points to Forbes figures of about $100 million in ARR but states that ChinaTalk could not independently verify those sales numbers. Surge AI, meanwhile, advertises on its own website that its products are built by the same experts behind its frontier-lab work. The report also describes deals running through offshore structures, noting that ByteDance maintains procurement offices in Singapore and San Jose.

How the relationships became visible is itself part of the story. ChinaTalk attributes the disclosures to a mix of anonymous sources and loose operational security among subcontractors. The US data firms outsource much of the actual RL environment construction and data curation to low-paid workers in developing countries, particularly India. Those subcontractors are supposed to refer to clients only by code names, but sometimes post real client names on LinkedIn or in résumés. Because RL environments remain largely unregulated, existing export controls barely reach this trade, according to the report.

For a developer, the useful way to read this is through what an RL environment actually is. Training a model to perform a task well increasingly means giving it a reward signal inside a controlled simulation rather than only feeding it text. Building a good environment is labor-intensive: someone has to encode the task, define success, construct realistic scenarios and verify that the rewards are not gameable. That work is a product in its own right, and a validated environment carries information about what a frontier lab found worth training on. This suggests the commercial value sits less in raw tokens than in the tested scaffolding around them, which is also why the same artifact can be resold to a different buyer without modification.

Dual-use domains are where the report places the sharpest concern. Coding environments improve software development but can also sharpen offensive cyber capability. Biology environments could improve gene synthesis and, in the authors' framing, potentially enhance applications relevant to bioweapons. The ChinaTalk authors argue that the Bureau of Industry and Security should bring these dual-use environments under existing export controls.

At the same time, the report is careful not to overstate the consequences. It stresses that the situation is not apocalyptic. Because Chinese labs are piggybacking on US work, they still trail their US peers by at least a few months as a structural matter, according to the authors. The report also says that sensitive environments in areas such as biology, chemistry, cybersecurity, chip design and investment banking are increasingly being built in-house at American frontier companies. A likely trade-off follows from that: as the most sensitive work moves inside the labs, the commercial market shifts toward more generic coding and office-task environments, which are exactly the ones easiest to resell and hardest to classify as controlled technology.

Why it matters: Data vendors, frontier labs and compliance teams are all implicated, since the report indicates that a meaningful revenue stream for US data firms flows to Chinese buyers and that current export rules do not clearly cover RL environments. For product teams, the practical question is provenance and contractual control over training artifacts, not just model weights. My inference is that if regulators follow the authors' recommendation, vendors selling task environments will face classification questions similar to those that already apply to chips, and buyers will need documentation showing where an environment was built and for whom.

Read at The Decoder

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

US data brokers like Mercor, Surge AI, and AfterQuery are selling roughly half a billion dollars' worth of AI training data to Chinese labs, according to a ChinaTalk investigation. That includes the same RL environments previously built for OpenAI or Anthropic. At least nine firms are involved, and export controls barely apply. The article US data brokers are selling half a billion dollars in RL training environments to Chinese AI labs appeared first on The Decoder .