A recent ChinaTalk investigation shows that several U.S. data‑broker companies are selling simulated training worlds – known as reinforcement‑learning (RL) environments – to Chinese artificial‑intelligence labs. The deals are worth roughly half a billion dollars. The Chinese firms buy the same environments that were previously sold to U.S. labs such as Anthropic and OpenAI, so they can skip the time‑consuming step of building the data themselves.
The practice matters because China’s chips often hit memory limits, making compute expensive. By reusing proven RL environments, Chinese labs can train models with far fewer GPU hours. They also use powerful U.S. models like GPT or Claude as “teachers,” feeding the outputs into the purchased environments to accelerate learning.
Who is involved and how the trade works
The report names Mercor, Surge AI, AfterQuery and Turing as among the sellers. At least nine firms in total were identified. AfterQuery reportedly gets a large share of its yearly recurring revenue from Chinese customers, though the exact figure could not be verified. Surge AI even advertises on its website that its products are built by the same experts who work on its frontier‑lab projects.
Most of the actual environment creation is outsourced to low‑paid workers in countries like India. These subcontractors sometimes reveal client names on professional profiles, which helped investigators trace the deals. The transactions often pass through offshore structures, with companies such as ByteDance running procurement offices in Singapore and San Jose.
Regulatory gaps and dual‑use worries
RL environments are largely unregulated, so export controls barely apply. The investigation flags coding and biology environments as “dual‑use” – they can improve software development or gene‑synthesis work, but the same tools could aid cyber attacks or bioweapon research. The authors suggest the U.S. Bureau of Industry and Security should extend existing export rules to cover these kinds of data.
Despite the flow of data, the report says Chinese labs still lag U.S. companies by several months on overall capability. Sensitive domains such as chip design, chemistry and investment banking are increasingly being built in‑house by American frontier firms.
Why it matters
For Chinese AI researchers, buying ready‑made training worlds reduces the amount of expensive compute they need, speeding up model development. For U.S. policymakers, the finding highlights a regulatory blind spot that could let advanced AI tools be repurposed for harmful applications without oversight.
