American brokers are selling AI training data to Chinese labs

American data broker companies, including Mercor, Surge AI, and AfterQuery, are supplying Chinese AI labs with reinforcement learning (RL) training datasets. According to a ChinaTalk investigation, the total value of these deals is estimated at approximately $500 million.

The materials being transferred include specialized training environments that were previously developed for the needs of companies such as OpenAI and Anthropic. The scheme involves at least nine different firms specializing in the collection and labeling of data for neural network training.

Current US export control regulations barely cover such operations, allowing high-tech training environments to be transferred without significant restrictions. This sets a precedent where critical components for the development of advanced models become available to foreign developers through intermediaries.

For the industry, this means increased difficulty in controlling the dissemination of intellectual property and AI training methods. Companies investing in the creation of unique environments for RL face the risk of their developments leaking, which are subsequently used to accelerate the development of competing models outside of US jurisdiction.