According to an investigation led by Aqib Zakaria and Quinn Ennis, published on November 11 by ChinaTalk, a prominent newsletter focusing on technology in China, several American artificial intelligence data enterprises, including Mercor, Surge AI, AfterQuery, and Turing, are instrumental in selling nearly half a billion dollars’ worth of data to Chinese AI laboratories.
A significant portion of this trade comprises reinforcement learning (RL) environments—simulated applications and tasks paired with grading metrics that labs utilize for post-training AI models across diverse sectors such as coding, finance, law, cybersecurity, and biology.
The investigative authors discovered at least nine American vendors supplying these services to clients in China, often utilizing anonymous sources as well as subcontractors, predominantly based in India.
These intermediaries, working under aliases, have made the names of their clients visible on platforms like LinkedIn or within their professional portfolios.
However, ChinaTalk did not disclose a detailed breakdown of the reported half-billion dollars, nor did it include responses from the companies involved.
Identified Companies
ChinaTalk indicates that its findings reveal Mercor has sold software engineering RL environments to Alibaba, referencing a Forbes article that notes 2% of Mercor’s revenue from the second quarter of 2026 originated from China.
Surge AI has been reported by SemiAnalysis and Forbes to have engaged in sales of RL environments to companies including Tencent, Z.ai, and Moonshot AI. Turing has also been vocal about its data transactions with Tencent and Alibaba.
Additionally, AfterQuery is claimed to derive half of its $100 million in annual recurring revenue from Chinese operations, although ChinaTalk mentioned it could not independently confirm this assertion.
Other smaller vendors are also implicated: Abundant and Eaglepoint reportedly sell to ByteDance, Pareto provides human data services for ByteDance, while DeccanAI has dealings with both Tencent and ByteDance.
Moreover, an anonymous source from a Chinese laboratory disclosed that it procures data from Invisible.
On the other hand, the only companies known to publicly forgo business with China are Scale, micro1, and Fleet. In May 2026, Mercor adjusted its customer policy from an ownership-based assessment to a geographical one.
Significance of Uniform Environments
In the analysis, the authors articulate that the primary advantage for Chinese labs lies in “piggybacking”, with these entities acquiring the same environments previously purchased by organizations such as Anthropic or OpenAI.
This strategy allows them to utilize validated data, thereby circumventing costly failed attempts. Furthermore, developments in reinforcement learning, particularly the inference-heavy phases, are confined by memory capacities; thus, the scarcity of high-bandwidth memory in Chinese chips renders computational savings exceptionally advantageous.
Surge AI promotes its RL environments as crafted by experts integral to their pioneering laboratory efforts.
This, when combined with knowledge distilled from American models, purportedly enables Chinese labs to close the development gap while expending less of their own resources.
Mercor, recently announcing a substantial fundraising of $750 million at a staggering $20 billion valuation, is recognized as one of the largest entities within this niche market.
Shortcomings of Export Controls
Although U.S. export controls are stringent regarding advanced chips and specific model weights, RL environments remain “largely free from export controls, save for Entity-Listed purchasers like Zhipu.”
Additionally, external affiliates outside of China can easily circumvent such barriers, as noted by ChinaTalk. Transactions frequently navigate through offshore financial frameworks; ByteDance’s data procurement teams reportedly operate from locations in Singapore and San Jose.
Recommendations include that the Bureau of Industry and Security classify dual-use environments under current regulations related to chip design, cybersecurity, and bioweapons, and issue “is-informed” letters to various vendors.
The Department of Defense is urged to employ procurement protocols to prevent its contractors from engaging with concerning Chinese laboratories.
It is important to note, however, that this trade is “not apocalyptic”; Chinese clients must await U.S. labs to first initiate an environment, consequently keeping them several months behind.
This report coincides with comments made by Treasury Secretary Scott Bessent, emphasizing the complex dynamics of collaboration and competition between the U.S. and China in the AI sector.
Source link: Mlq.ai.
