US data brokers quietly sell half a billion dollars in AI training worlds to China

A new investigation reveals that US data companies are selling sophisticated reinforcement learning environments to Chinese AI labs, creating a shortcut that helps China's AI development despite US export restrictions. At least nine firms—including Mercor, Surge AI, AfterQuery, and Turing—are reportedly selling these simulated training worlds worth approximately $500 million annually to Chinese customers, with some companies deriving significant portions of revenue from these deals.

Chinese labs intentionally purchase the same RL environments previously sold to top US AI companies like Anthropic and OpenAI, allowing them to skip costly in-house development. Since Chinese AI chips suffer from memory bottlenecks, these ready-made environments provide crucial efficiency gains. Combined with 'distillation' techniques where Chinese models learn from outputs of US models like GPT and Claude, this creates a two-pronged advantage that helps China close the AI gap without matching US computing power—all while export controls remain largely ineffective against these unregulated data products.