ChinaTalk published an investigation on November 11, 2026, reporting that U.S.-linked data vendors are selling training data and reinforcement-learning environments to Chinese AI labs. Authors Aqib Zakaria and Quinn Ennis name at least nine vendors, including Mercor, Surge AI, Turing and AfterQuery, and put the broader data business at roughly $500 million.
The competitive concern is specific: Chinese labs may be buying environments already developed and refined for U.S. frontier labs, saving some of the work involved in teaching models to code, use software and complete specialist tasks. The investigation also raises dual-use concerns around cybersecurity and biology. These are ChinaTalk’s findings and analysis, supported by a mixture of sales information, public traces, previous reporting and anonymous sources.
The dollar figure needs careful handling. It covers data sales broadly, not $500 million exclusively in RL environments. The reporting does not provide a transaction-level breakdown that would let readers separate conventional data work from environments or reproduce the aggregate estimate.
What an RL Environment Actually Provides
An RL environment gives a model somewhere to attempt a task and a way to evaluate the result. The product can include software, starting conditions, task descriptions, reference material and a scoring system.
ChinaTalk illustrates this with a PowerPoint-like application. A vendor could build a clone of the app, then commission experts to define what makes a successful slide deck: readable text, clear organization, appropriate formatting and other requirements.
A simplified training loop looks like this:
- Set the task. Give the model a brief and access to the application.
- Generate an attempt. The model takes actions and produces a slide deck. This attempt is often called a rollout.
- Score the result. A rubric or evaluator assigns a reward signal.
- Update and repeat. The training process uses those signals to adjust the model’s behavior across many attempts.
The buyer therefore acquires more than a collection of example answers. It gets machinery for generating additional training experience.
Coding environments can work differently. ChinaTalk describes vendors acquiring company repositories containing bugs and the patches engineers used to fix them. An environment can expose the earlier, broken version and evaluate whether the model produces a successful repair.
The grading mechanism is central to the product’s value. ChinaTalk notes that environments can contain bugs, poor-quality data or rubrics that reward incorrect behavior. My view is that the evaluator deserves as much scrutiny as the task itself: a polished application clone is of limited use if its scoring system teaches the model to optimize the wrong thing.
This also explains the difference from distillation. Distillation uses a stronger model’s outputs as training material for another model. An RL environment supplies tasks and feedback for practice. A lab can use both, but purchasing an environment does not itself mean obtaining another lab’s model weights or copying its answers.
What the Vendor Evidence Supports
ChinaTalk names Mercor, Surge AI, Turing, AfterQuery, Abundant, Pareto, DeccanAI, Eaglepoint and Invisible. The evidence is uneven across that list.
According to the investigation and MLQ’s detailed summary, the reported relationships include:
| Vendor | Business described in the reporting | Evidence qualification |
|---|---|---|
| Mercor | Software-engineering RL environments sold to Alibaba | ChinaTalk says it accessed sales data supporting this relationship. |
| Surge AI | RL environment sales to Tencent, Z.ai and Moonshot AI | ChinaTalk cites earlier reporting; public marketing separately promotes its frontier-lab expertise. |
| Turing | Data sales to Tencent and Alibaba | Reported by ChinaTalk; the supplied summary does not identify transaction-level documentation. |
| AfterQuery | A substantial Chinese customer business | A revenue claim attributed to Forbes that ChinaTalk says it could not independently verify. |
| Abundant | Sales to ByteDance | A relationship reported in the investigation. |
| Pareto | Human-data work for ByteDance | A relationship reported in the investigation. |
| DeccanAI | Business with Tencent and ByteDance | Relationships reported in the investigation. |
| Eaglepoint | Sales to ByteDance | A relationship reported in the investigation. |
| Invisible | Purchases by a Chinese lab | Attributed to an anonymous source at that lab. |
Public contractor posts, résumés, customer testimonials and procurement materials can help establish business connections. They do not all answer the same questions. A contractor naming a client may support a relationship without revealing the contract’s value, the product delivered or whether a particular environment was reused.
Sales data described by the investigators is a different evidence category from an anonymous source’s account. Neither should be silently converted into a publicly inspectable contract.
AfterQuery illustrates a separate accounting problem. MLQ reports that ChinaTalk cited a Forbes claim linking roughly half of AfterQuery’s reported $100 million in annual recurring revenue to Chinese business. Annual recurring revenue is an annualized business metric, not necessarily cash collected or completed sales during a particular period. That figure cannot be added to other revenue numbers without checking their definitions and timeframes.
MLQ also says company responses were not included in the report. That leaves readers without a response from the named vendors in this reporting; it should not be treated as an admission or evidence that a company refused to comment.
The Decoder’s coverage summarizes the investigation and its implications. It is useful context, but it is not a second, independent verification of the underlying sales.
Why Reusing an Environment Could Save Time and Compute
ChinaTalk’s central argument is that Chinese labs can benefit from work already commissioned by companies such as OpenAI and Anthropic.
Building a useful environment involves several uncertain choices: selecting tasks, setting their difficulty, making the software reliable and deciding how to score success. A buyer that finds a productive environment can avoid some unsuccessful development and training runs.
The potential savings come from three places:
- Less environment development: The buyer does not have to build every application, task collection and evaluator from scratch.
- Fewer poor purchases: Prior use by a capable lab can provide a signal about product quality.
- Less wasted training: Better tasks and scoring may reduce attempts that consume compute without producing useful improvements.
The Decoder reports ChinaTalk’s argument that these savings matter particularly for Chinese labs facing memory bottlenecks. Rollouts require models to generate actions and outputs repeatedly, making inference resources an important part of the training budget.
That does not yield a universal savings percentage. ChinaTalk itself says compute savings vary substantially by environment.
A further distinction matters: the same experts are not necessarily the same product. Surge AI’s reported marketing of its frontier-lab expertise signals access to relevant know-how. On its own, that marketing does not prove that two customers received identical environments, private customer data or exclusive intellectual property.
Even an identical environment may help two models differently. ChinaTalk identifies task diversity, task choice and difficulty calibration for a particular model as important factors. An environment productive for one system may be too easy, too hard or poorly matched to another.
Frequently Asked Questions
3 questions
1Is the $500 Million All for RL Environments?
No. ChinaTalk’s estimate covers data sales broadly. RL environments account for an unspecified portion, and the reporting provides no transaction-level breakdown.
