Editorial illustration for US data brokers sell USD 500M in RL training environments to Chinese AI labs
US Data Brokers Sell $500M RL Training to China
Mercor, Surge AI, AfterQuery, and Turing built their businesses on a simple pitch: supply the reinforcement learning environments that teach models like GPT and Claude how to code, trade, and work through office tasks. Those same companies are now selling that expertise to Chinese AI labs, according to an investigation by ChinaTalk that puts the total trade at roughly $500 million.
At least nine US firms show up in the investigation, moving training environments that were originally built for American frontier labs straight into the hands of their Chinese competitors. The appeal is practical. Chinese labs already know these environments work because OpenAI or Anthropic used them first, which means they can skip the trial-and-error of building their own and avoid burning compute on training runs that go nowhere. That matters more in China, where memory bottlenecks on domestic chips make every wasted cycle more expensive than it would be in the US.
There's a second layer to this, too, involving how Chinese labs treat American models as teachers, feeding their outputs into their own training pipelines. No export rules currently cover any of it.
Chinese labs deliberately buy the same products previously sold to Anthropic or OpenAI, the report says. That way, they already know the data works and can skip costly in-house development and wasted training runs. Because Chinese chips suffer from acute memory bottlenecks, those savings are particularly valuable for labs there.
Why this matters
This is an export-control story masquerading as a data-market story. The chip war got all the attention while the actual training ingredients walked out the side door through nine brokers nobody outside this investigation had heard of. If Chinese labs can buy the same RL environments built for OpenAI and Anthropic, they're not just catching up on compute, they're skipping the expensive trial-and-error that compute shortage was supposed to force on them. That's the part worth sitting with: memory-bottlenecked chips made these purchases more valuable to Beijing's labs, not less, because every wasted training run costs them more than it costs a US lab with better hardware.
For founders and researchers, the lesson is about supply chains, not politics. If your company builds or sells training environments, assume your buyer list has no real border right now. Half a billion dollars moved without anyone deciding it should. Watch whether Washington treats this as seriously as it treats silicon, because right now the data side of the AI stack is regulated by nobody in particular.
Related Reading
Research & Benchmarks Microsoft AI Chief: "If It Isn't Safe We Shouldn't Build It
Research & Benchmarks Google's FACTS benchmark shows 70% factuality ceiling across four tests
Research & Benchmarks Nvidia's DLSS 4.5 beta adds 6x Multi Frame Generation for RTX 50 GPUs
From Archives Grok Build coding agent leaks confidential data, some belonging to SpaceXSI
From Archives DeepMind Reveals Gemini Robotics 2.5 and Open Safety Benchmark ASIMOV-Jailbreak
Common Questions Answered
Which US data brokers are selling reinforcement learning training environments to Chinese AI labs?
According to the ChinaTalk investigation, at least nine US firms are involved in this trade, with Mercor, Surge AI, AfterQuery, and Turing being specifically named as major players. These companies built their businesses supplying the reinforcement learning environments that train models like GPT and Claude to code, trade, and handle office tasks.
What is the total value of RL training environments being sold to Chinese AI labs?
The investigation reports that the total trade in reinforcement learning training environments between US data brokers and Chinese AI labs is approximately $500 million. This represents a significant transfer of training infrastructure that was originally developed for leading AI companies like Anthropic and OpenAI.
Why are Chinese AI labs specifically purchasing the same RL environments previously sold to Anthropic and OpenAI?
Chinese labs deliberately buy proven products to skip costly in-house development and avoid wasted training runs, since they already know the data works from previous sales to major US AI companies. This approach is particularly valuable for Chinese labs facing acute memory bottlenecks in their chips, allowing them to bypass expensive trial-and-error processes.
How does this RL training environment trade impact the chip war between the US and China?
While the chip war received significant attention, the actual training ingredients for AI models have been transferred through these nine data brokers largely unnoticed. By accessing the same reinforcement learning environments used by OpenAI and Anthropic, Chinese labs can catch up on compute capabilities and skip the expensive trial-and-error that chip shortages were supposed to force upon them.
What makes this RL training environment export significant from an export-control perspective?
This situation represents an export-control issue disguised as a data-market transaction, where critical training ingredients have essentially walked out the side door through largely unknown brokers. The transfer of proven RL training environments undermines the intended effects of US chip restrictions by allowing Chinese labs to accelerate their AI development without the computational constraints they were supposed to face.
Further Reading
- The US Is Selling RL Environments to China - ChinaTalk
- US data brokers are selling half a billion dollars in RL training environments to Chinese AI labs - The Decoder
- AI training data, RL environment & eval vendors - WhoTrainsAI
- RL Environments Market Map: Who Sells What - RL Research
- RL environments and benchmarks for coding - RL Research
