American data brokers are supplying roughly half a billion dollars worth of reinforcement learning training environments to Chinese artificial intelligence laboratories, according to an investigation by ChinaTalk.reported This trade involves simulated worlds used for training models on specific tasks like coding and financial analysis.This trade helps Chinese labs save money and time by using training data already proven to work for top US AI companies.Chinese labs are leveraging purchased RL environments and model distillation to mitigate hardware memory bottlenecks and reduce compute costs.
Evidence: reported
US data brokers are selling half a billion dollars in RL training environments to Chinese AI labsaifleet-decoder
The figures
The investigation identifies at least nine US firms participating in this data trade. The total value of the training data sold to Chinese labs is estimated at $500 million.reported AfterQuery reportedly derives a significant portion of its annual recurring revenue from Chinese customers, with Forbes previously reporting approximately $100 million in ARR, though ChinaTalk could not independently verify those specific sales figures.reported
Evidence: reported
US data brokers are selling half a billion dollars in RL training environments to Chinese AI labsaifleet-decoder
Evidence: reported
US data brokers are selling half a billion dollars in RL training environments to Chinese AI labsaifleet-decoder
Behind the figures
ChinaTalk compiled this information using anonymous sources and by analyzing operational security failures among subcontractors. The report notes that Surge AI openly advertises on its website that its products are "built by the same experts behind our frontier-lab work." The deals also run through offshore structures. ByteDance maintains procurement offices in Singapore and San Jose.
What it means for your stack
Chinese labs are deliberately purchasing the same training environments previously sold to Anthropic and OpenAI. This strategy allows them to skip costly in-house development and avoid wasted training runs. Because Chinese hardware suffers from acute memory bottlenecks, these pre-built environments provide significant compute savings. Additionally, labs use powerful US models like GPT or Claude as teachers for distillation, creating a two-pronged advantage that helps them close the performance gap with less domestic compute.
Building on this
Developers should recognize that the RL environments being sold are standardized simulated worlds for tasks like office work and coding. Since Chinese labs are buying the exact same products used by US frontier models, they can replicate training pipelines without rebuilding the environments. This allows foreign labs to leverage existing, validated data structures rather than creating new ones from scratch, saving compute during model development.
Why it matters
- Inference engineers must account for the possibility that foreign models are trained on identical RL environments, potentially narrowing the performance gap between US and Chinese AI systems.
- The use of distillation with US models as teachers, combined with purchased RL data, suggests Chinese labs can achieve high performance with significantly less domestic compute power.
- Export controls barely apply to these data trades, leaving a regulatory gap for sensitive training data in areas like cybersecurity and biology.
- The involvement of at least nine US firms indicates a widespread commercial infrastructure for exporting AI training capabilities, rather than isolated incidents.
Topics: AI training, data brokers, export controls, reinforcement learning, Chinese AI
Updated
Then / now / next
US data brokers primarily sold RL training environments to domestic frontier labs like OpenAI and Anthropic.
At least nine US firms are exporting these same environments to Chinese labs for an estimated $500 million.
The authors of the investigation call for new export controls to address potential security risks in cybersecurity and biology.
Story timeline
- Then US data brokers primarily sold RL training environments to domestic frontier labs like OpenAI and Anthropic.
- 11 Nov 2026 US data brokers are selling half a billion dollars in RL training environments to Chinese AI labs
- Now At least nine US firms are exporting these same environments to Chinese labs for an estimated $500 million.
- Next The authors of the investigation call for new export controls to address potential security risks in cybersecurity and biology.
The strongest doubt
Counterpoint from the sources
ChinaTalk could not independently verify the specific sales figures for AfterQuery, which were based on earlier Forbes reports.
For readers
Developers
- Chinese labs are using identical RL environments to those used by US frontier models, potentially replicating training pipelines.
- Distillation with US models allows foreign labs to train with less domestic compute power.
Policy
- Current export controls barely apply to the sale of AI training data and RL environments.
- The investigation calls for new regulations to address security risks in sensitive fields like biology.
Investors
- At least nine US firms are involved in a roughly $500 million trade of AI training data with Chinese labs.
Ask the article
Local keyword search of this article; no generated answers.Did this article give you the evidence you needed?
Your choice stays on this device. No public tally.Questions readers ask
Which US companies are selling training data to Chinese labs?
The investigation names Mercor, Surge AI, AfterQuery, and Turing as firms involved in selling training data to Chinese AI laboratories.
What is the total value of this data trade?
According to the ChinaTalk investigation, US data brokers are selling roughly half a billion dollars worth of AI training data to Chinese labs.
How do Chinese labs use this data to save compute?
They buy the same RL environments used by US labs to skip in-house development and use US models for distillation, reducing the need for expensive domestic training runs.
Connected stories
Reading receipt
- At least nine US data brokers are selling training environments to Chinese labs.
- The total value of these transactions is estimated at roughly $500 million.
- Buyers use identical environments previously sold to OpenAI and Anthropic.