Chinese AI Labs Buy $500M in U.S. Data Despite an Export Gap

By Rebecca Stone

U.S. AI-data vendors are selling roughly $500 million worth of training data to Chinese AI laboratories, according to a Nov. 11 ChinaTalk investigation naming nine suppliers and five cited buyers.

The investigation by Aqib Zakaria and Quinn Ennis, published by the China-tech newsletter, traces sales to Alibaba, Tencent, ByteDance, Moonshot AI and Z.ai.

The estimate has limits. ChinaTalk published no revenue breakdown for the $500 million, and the authors included no responses from the companies.

What are Chinese labs buying?

An undisclosed chunk of the reported total consists of reinforcement-learning environments: simulated versions of apps and tasks paired with grading rubrics that labs use to post-train models in coding, finance, law, cyber and biology.

Surge AI markets those environments as “built by the same experts behind our frontier-lab work,” according to its website.

Who is selling into China?

Zakaria and Ennis assembled their findings from data they accessed, anonymous sources and subcontractors, often in India, whose LinkedIn profiles or résumés list client names despite their work using code names.

The report named:

• Mercor: Data reviewed by ChinaTalk shows software-engineering RL-environment sales to Alibaba. Forbes separately reported that China supplied 2% of Mercor’s second-quarter 2026 revenue.
• Surge AI: SemiAnalysis and Forbes have reported sales of RL environments to Tencent, Z.ai and Moonshot AI.
• Turing: Sells data to Tencent and Alibaba, according to ChinaTalk.
• AfterQuery: Forbes reported that Chinese business accounts for half of its $100 million in annual recurring revenue; ChinaTalk said it could not independently verify that figure.
• Abundant and Eaglepoint: Each sells to ByteDance, according to the report.
• Pareto: Performs human-data work for ByteDance.
• DeccanAI: Serves Tencent and ByteDance.
• Invisible: An anonymous source at a Chinese lab said it buys from the vendor.

ChinaTalk said Scale, micro1 and Fleet were the only companies it knew of that publicly said they forgo business in China. In May 2026, Mercor changed its customer policy from an ownership test to a location test.

Why buy environments already used by U.S. labs?

The authors say Chinese labs deliberately buy environments already sold to Anthropic or OpenAI, allowing them to rely on work known to function and skip expensive failed training runs.

That reuse has particular value because rollouts, the inference-heavy part of RL, are memory-bound; ChinaTalk says Chinese chips are short of high-bandwidth memory, which magnifies the savings.

The authors also say distillation from U.S. models lets Chinese labs narrow the performance gap while using less compute of their own.

Mercor, described as one of the market’s largest vendors, announced this week that it raised $750 million at a $20 billion valuation.

Where do export controls leave a gap?

Current U.S. export controls cover advanced chips and some model weights. RL environments, however, remain largely free from export controls apart from Entity-Listed buyers such as Zhipu, ChinaTalk says.

An unlisted affiliate outside China can sidestep that restriction, according to the report. Deals often run through offshore desks; the authors place ByteDance’s data-buying teams in Singapore and San Jose.

ChinaTalk wants the Bureau of Industry and Security to classify dual-use environments under existing rules governing chip design, cyber and bioweapons, then send warning letters to vendors. The authors also want the Pentagon to tighten contractor rules through the Department of War’s procurement authority, barring sales to Chinese labs they consider concerning.

They temper the alarm. Chinese buyers must wait for a U.S. lab to commission an environment first, leaving them months behind.

The report came a day after Treasury Secretary Scott Bessent said the U.S. and China should “harbor no illusions” as they cooperate and compete on AI.

The next commercial variable is access to proven environments: Chinese labs can still reuse U.S.-commissioned work for less compute, while tighter U.S. rules could put part of the roughly $500 million vendor channel at risk.