US Data Brokers Selling Millions in AI Training Data to Chinese Labs

Key Takeaways:

• US data brokers are selling an estimated $500 million worth of AI training environments to Chinese labs.

• The data includes environments previously used for developing frontier models like those from OpenAI and Anthropic.

• Experts warn of dual-use risks in sensitive fields like cybersecurity and biology, urging stricter export controls.A report from ChinaTalk has identified at least nine US firms supplying training environments to Chinese AI laboratories. These simulated worlds are crucial for training models on specialized tasks, and by purchasing them, Chinese labs are reportedly saving significant compute resources and accelerating their development timelines by leveraging US-built datasets.The trade has sparked concerns regarding national security and the lack of regulatory oversight. Because these environments can be used for dual-use applications—such as improving offensive cyber tools or gene synthesis—policymakers are facing calls to bring this data trade under existing export control regulations. While the US maintains a structural lead, the leakage of these refined training methodologies could allow foreign competitors to close the gap more quickly than previously expected.