What's happening
According to a ChinaTalk investigation, US data brokers are selling roughly half a billion dollars' worth of training data to Chinese AI labs. The firms named include Mercor, Surge AI and AfterQuery.
More than raw data
The trade is not limited to raw data. Reinforcement learning (RL) environments previously built for US labs such as OpenAI or Anthropic are also reaching Chinese labs. RL environments are simulated task settings in which a model develops behavior through trial and error.
Controls fall short
At least nine firms are involved, and current export controls barely cover this kind of data sale. In other words, restrictions on chip exports leave a large gap on the data side of model training.
Why it matters
Data and human-built training environments have become as critical as chips in the AI race. The largely unregulated nature of this trade could reignite debate over how effective US technology restrictions really are.
