American data companies are reportedly selling approximately $500 million worth of artificial intelligence training data and reinforcement learning environments to Chinese AI laboratories, according to an investigation by ChinaTalk. The companies identified include Mercor, Surge AI, AfterQuery and Turing, which supply specialised datasets and simulated environments used to improve the capabilities of advanced AI models. The transactions are attracting attention because some of these products were originally developed for leading American AI companies, including OpenAI and Anthropic.
The investigation, published on November 11, 2026, identified at least nine US data vendors with reported business relationships involving Chinese AI customers. A portion of the products being sold consists of reinforcement learning environments that help AI models improve at tasks such as coding, financial analysis, office work and cybersecurity. The report argues that these transactions could help Chinese AI developers save computing resources and accelerate model development, while exposing gaps in US export-control regulations that primarily focus on advanced semiconductor technology.
What Are Reinforcement Learning Environments?
Reinforcement learning (RL) environments are simulated settings in which AI models practise completing specific tasks and receive feedback about their performance.
Instead of simply learning from large collections of text, an AI model interacts with an environment, attempts an action and receives a reward or penalty depending on the result.
Over repeated attempts, the model can improve its ability to complete the task.
For example, an AI developer could purchase an environment designed to teach a model how to create professional presentations.
The environment might contain a simulated presentation application, sample assignments and a grading system that evaluates the quality of the finished slides.
The model attempts the assignment, receives feedback and repeats the process until its performance improves.
Other environments can simulate:
- Software development and debugging.
- Financial analysis and spreadsheet operations.
- Legal research and document review.
- Website navigation and office applications.
- Cybersecurity testing.
- Scientific and biological research tasks.
The value of an RL environment depends on the quality of its tasks, the diversity of its data and the accuracy of its evaluation system.
These environments are becoming important because they can teach AI models practical capabilities that are difficult to develop through conventional pre-training alone. <Cite refs={[“turn134050search3″,”turn134050search0”]}/>
Nine US Companies Identified in the Investigation
ChinaTalk’s investigation identified at least nine American data vendors with reported sales or business relationships involving Chinese AI companies.
The companies include some of the largest providers of specialised AI training data, as well as smaller businesses focusing on particular technical fields.
| US data vendor | Reported Chinese business connections |
|---|---|
| Mercor | Alibaba |
| Surge AI | Tencent, Z.ai and Moonshot AI |
| Turing | Tencent and Alibaba |
| AfterQuery | Significant reported Chinese customer exposure |
| Abundant | ByteDance |
| Pareto | ByteDance |
| DeccanAI | Tencent and ByteDance |
| Eaglepoint | ByteDance |
| Invisible | Chinese AI laboratory customers |
The investigation says Mercor sold software-engineering reinforcement learning environments to Alibaba.
Surge AI has reportedly supplied environments to Tencent, Z.ai and Moonshot AI, while Turing has been associated with data sales to Tencent and Alibaba.
AfterQuery is another notable example. Forbes previously reported that the company generated approximately $100 million in annual recurring revenue, with a significant portion attributed to Chinese customers. ChinaTalk said it could not independently verify the reported revenue figures.
The investigation relied on anonymous sources, contractor information, business records and other evidence. Not every commercial relationship was independently confirmed, and the companies’ involvement should not be interpreted as evidence of wrongdoing. <Cite refs={[“turn134050search3″,”turn134050search4”]}/>
Why Chinese AI Labs Are Buying American Training Environments
Developing high-quality reinforcement learning environments requires considerable time, money and specialised expertise.
AI companies need suitable tasks, realistic simulations, high-quality datasets and reliable evaluation systems.
For specialised applications, they may also need professionals with experience in fields such as software engineering, finance, law and scientific research.
Purchasing an environment that has already been developed can save a laboratory from building the entire system internally.
ChinaTalk argues that Chinese AI companies are deliberately buying some of the same environments previously sold to American frontier AI laboratories.
This provides two potential advantages.
1. Lower Development Costs
Building sophisticated training environments from scratch can be expensive.
A Chinese AI laboratory purchasing an existing product can avoid some of the development costs and engineering work required to create equivalent simulations.
The laboratory can then concentrate its resources on training and improving its own models.
2. Access to Tested Training Products
An environment previously used by companies such as OpenAI or Anthropic may already have undergone substantial development and evaluation.
Buying an established product can reduce the uncertainty involved in developing a new training system.
This is particularly valuable for specialised tasks where creating accurate evaluation criteria requires experienced professionals.
The investigation also argues that these purchases could help Chinese laboratories compensate for computing constraints, including limitations involving memory availability in domestic AI hardware. <Cite refs={[“turn134050search3″,”turn134050search0”]}/>
How Reinforcement Learning Helps AI Models Become More Capable
The importance of these transactions becomes clearer when examining how modern AI models are developed.
Pre-training teaches an AI model general patterns from large datasets. Post-training then improves its ability to follow instructions, solve problems and perform specialised tasks.
Reinforcement learning is one method used during post-training.
A model attempts a task, receives feedback and adjusts its behaviour to improve future results.
For example, an AI model learning software engineering could be given a code repository containing a bug.
It must identify the problem, produce a correction and submit the modified code.
An evaluation system checks whether the correction works. If it succeeds, the model receives a positive reward.
By repeating this process across many different tasks, the model can improve its coding capabilities.
RL environments provide the structured setting and feedback needed for this process.
ChinaTalk argues that improvements between successive versions of the same underlying model can come substantially from post-training rather than entirely new pre-training runs.
This makes access to high-quality reinforcement learning environments strategically valuable for AI developers seeking to improve model performance. <Cite refs={[“turn134050search3”]}/>
Distillation Creates Another Advantage for Chinese AI Developers
The investigation also discusses a separate technique known as model distillation.
Distillation involves using a more capable AI model as a teacher to generate examples that can help train another model.
For example, a developer could use outputs from a powerful model such as GPT or Claude to create training examples for a smaller or less capable model.
The student model learns from these examples and attempts to reproduce the teacher’s performance on specific tasks.
Combined with purchased reinforcement learning environments, this can create a more efficient development strategy.
- Distillation: Provides examples generated by stronger models.
- Reinforcement learning environments: Give models opportunities to practise tasks and receive feedback.
- Post-training: Uses these resources to improve the model’s specialised capabilities.
However, distillation and purchasing RL environments are different activities. The investigation does not establish that every Chinese laboratory purchasing training environments is also conducting unauthorised distillation.
The two techniques nevertheless illustrate how access to existing AI resources could reduce the cost of developing competitive models. <Cite refs={[“turn134050search3″,”turn134050search0”]}/>
Cybersecurity and Biological Research Raise National Security Concerns
The most sensitive part of the investigation involves the possibility that specialised training environments could improve AI capabilities in areas with both civilian and military applications.
These are commonly described as dual-use technologies.
Cybersecurity
Coding environments can teach AI models to identify bugs, understand complex software repositories and develop fixes.
These capabilities have legitimate applications in software development and defensive cybersecurity.
However, similar skills could potentially improve the ability of an AI system to identify vulnerabilities and support offensive cyber operations.
The concern is particularly significant if specialised environments train models to discover and exploit weaknesses in software systems.
Biology and Chemistry
Biological research environments can help AI systems understand scientific workflows, analyse experimental information and assist with complex laboratory-related tasks.
Such capabilities could support legitimate medical research and biotechnology.
However, more advanced biological capabilities could also create risks if used to support dangerous research involving harmful biological agents.
Semiconductor and Financial Research
Specialised environments involving chip design, financial analysis and investment banking could improve AI systems’ capabilities in strategically important industries.
The investigation argues that these categories deserve particular attention from US policymakers because they could have applications beyond ordinary commercial productivity. <Cite refs={[“turn134050search3″,”turn134050search0”]}/>
Why Existing US Export Controls May Not Cover These Sales
One of the investigation’s central arguments is that US export-control rules have not fully adapted to the changing AI training-data market.
Washington has introduced restrictions targeting advanced semiconductor chips and certain technologies that could support China’s military and AI capabilities.
However, reinforcement learning environments are not generally regulated in the same way as advanced computing hardware.
The report argues that this creates a gap through which specialised AI training products can reach Chinese customers even when access to certain processors is restricted.
It also highlights the use of offshore procurement structures.
For example, ByteDance maintains procurement operations in Singapore and San Jose, according to the investigation. Offshore arrangements can make it more complicated to identify the ultimate customer and determine which regulations apply.
The authors also point to existing restrictions involving sensitive personal information, including the US Department of Justice’s Data Security Program and the Protecting Americans’ Data from Foreign Adversaries Act.
These rules address particular categories of data, but they do not automatically cover every type of simulated training environment or AI evaluation product.
ChinaTalk argues that the US Bureau of Industry and Security should consider extending export controls to certain sensitive reinforcement learning environments.
That is a policy recommendation from the investigation, not an announcement that new restrictions have already been introduced. <Cite refs={[“turn134050search3″,”turn134050search4”]}/>
Indian Contractors Are Also Part of the AI Training Supply Chain
Another notable aspect of the investigation involves the workers who help create these training environments.
ChinaTalk reports that US data vendors outsource substantial amounts of data collection, environment development and content evaluation to contractors in lower-cost markets.
India is identified as an important source of this outsourced labour.
According to the investigation, some subcontractors have revealed customer identities through LinkedIn profiles, resumes and other public information, even when they were expected to use confidential project names.
These disclosures reportedly helped investigators identify commercial relationships between American vendors and Chinese AI laboratories.
The arrangement also highlights how the AI training industry operates through a complex international supply chain.
A US company may sell an AI training product to a Chinese customer while relying on contractors in India or other countries to create parts of the underlying environment.
This creates additional challenges involving confidentiality, intellectual property, data security and customer verification. <Cite refs={[“turn134050search3″,”turn134050search0”]}/>
The Investigation Also Identifies Limits to the Risk
ChinaTalk does not argue that every sale of AI training data to a Chinese laboratory creates an immediate national security threat.
The investigation identifies several reasons why the risks should be assessed carefully.
First, Chinese laboratories may still receive training environments after American AI companies have already purchased or developed them. This creates a time delay that could leave Chinese developers behind their American counterparts.
Second, some environments are built around Western software and business practices, limiting their usefulness in Chinese applications.
Third, highly sensitive environments involving biology, chemistry, cybersecurity and chip design are increasingly being developed internally by leading American AI companies.
Finally, not every coding or scientific training environment creates dangerous capabilities. Many products are designed for ordinary commercial tasks such as software debugging, office work and document analysis.
The authors therefore distinguish between routine commercial AI training products and specialised environments that could improve capabilities in sensitive fields.
This distinction is important because broad restrictions could also limit legitimate research and competition in less sensitive areas. <Cite refs={[“turn134050search3”]}/>
The Bigger Picture
The investigation highlights a new dimension of competition in the global AI industry. Until now, much of the debate about US-China technology restrictions has focused on advanced processors, semiconductor manufacturing equipment and computing infrastructure. However, high-quality training data and reinforcement learning environments are also becoming important strategic resources. If Chinese AI laboratories can purchase products already developed for American frontier companies, they may reduce the time and resources required to improve their own models.
The issue presents a difficult challenge for policymakers. Restricting access to sensitive training environments could help protect capabilities related to cybersecurity and advanced biology, but broad restrictions could also damage legitimate commercial activity and scientific collaboration. As AI development increasingly depends on specialised datasets, expert feedback and simulated environments, governments may need to reconsider how existing export-control frameworks apply to the global AI training supply chain.
Looking Ahead
The next stage of this debate will likely involve closer scrutiny of American AI data vendors, their overseas customers and the contractors involved in producing training environments. US policymakers may examine whether certain products should receive additional export-control restrictions, particularly when they improve capabilities in sensitive technical fields. AI companies may also face greater pressure to verify customers, monitor subcontractors and establish clearer rules governing the resale of specialised training products.
For the global AI industry, the investigation demonstrates that computing power is only one part of the competitive equation. Access to high-quality data, expert evaluation and effective reinforcement learning environments can also influence how quickly models improve. As governments attempt to protect strategically important AI capabilities, controlling the flow of training resources could become an increasingly important part of the US-China technology competition.
SEO Metadata
SEO Title: US Data Brokers Sell $500 Million in AI Training Data to Chinese Labs
Meta Description: A ChinaTalk investigation identifies nine US data vendors reportedly selling AI training data and reinforcement learning environments to Chinese AI laboratories.
URL Slug: us-data-brokers-ai-training-data-chinese-labs
Focus Keyphrase: US AI training data sales to China
Secondary Keywords: reinforcement learning environments, Chinese AI laboratories, Mercor, Surge AI, AfterQuery, Turing, US China AI competition, AI export controls, AI training data, AI national security
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
