U.S. Data Startups Sell Millions in AI Training Sets to China
American data-labeling startups are selling hundreds of millions of dollars in AI training datasets to Chinese technology firms despite U.S. restrictions on AI chip exports.
American big data-labeling startups are selling an estimated $500 million in training datasets annually to the top six Chinese AI labs. This trade allows Chinese firms to bypass trial-and-error processes and close the performance gap with American rivals by purchasing human-curated data packages.
Major Chinese buyers include Tencent, ByteDance, Alibaba, and Ant Group. U.S. vendors such as Surge AI, Mercor, AfterQuery, and Turing provide these datasets. Some of these providers maintain simultaneous contracts with U.S. federal entities, including the Army and Air Force. AfterQuery alone generates at least $50 million in recurring revenue from Chinese labs.
While the U.S. government has restricted the export of advanced AI chips, it has not implemented similar precautions for the data used to train models. This gap has led some industry figures to label the practice a national security risk, while others argue that restricting data exports would hinder global open-source innovation.