Some of the best evidence about the AI race turned up in résumés. Subcontracted workers who were supposed to refer to their clients by code names listed the real names on LinkedIn and in their work histories, and that carelessness helped ChinaTalk piece together a trade worth roughly $500 million. At least nine American data firms, the investigation found, have been selling AI training material to Chinese labs.
The firms named in the coverage include Mercor, Surge AI, AfterQuery, and Turing. According to The Decoder’s account of the findings, Chinese labs deliberately buy the same products these firms first sold to Anthropic and OpenAI, and the deals run through offshore structures. Some of what they buy takes the form of reinforcement learning environments.
Surge AI’s own website describes its ready-made products as “built by the same experts behind our frontier-lab work.” That line is a sales pitch, and it is accurate. It also poses the question this trade puts to anyone who thinks the American lead in AI sits mainly in chips. When a Chinese lab can buy the training material that shaped an American frontier model, what exactly changes hands, and why do the export rules barely notice?
What a Training Environment Sells
The Decoder describes reinforcement learning environments as “simulated worlds where AI models are trained on specific tasks like coding, financial analysis, or office work.” A model inside one tries a task, gets graded, and tries again. It might be asked to fix a bug in a codebase, reconcile a spreadsheet, or work through a mock web form, and every attempt ends with a score.
The grading is what gives an environment its value. Anyone can write down a task. Building a world where a model’s failure is visible, specific, and cheap to repeat is the hard part, because the environment has to know what a correct spreadsheet looks like, which shortcuts count as cheating, and when a fix that passes the tests still breaks something else.
Daniel Coyle spent The Talent Code explaining why that kind of setting produces skill. His argument is that deep practice “takes events that we normally strive to avoid,” meaning mistakes, and turns them into ability. A learner improves by failing in a structured place, noticing the failure, and correcting it, again and again, under a coach who designed the drills to expose the right errors.
Coyle was writing about people, but the bridge to machine training is short. A reinforcement learning environment is a drill sheet with an automatic coach attached. The model’s errors are the raw material, and the environment’s job is to make those errors easy to read and cheap to correct.
That changes what the buyer is paying for. A textbook sells knowledge, and knowledge has been cheap for a long time; the open web is full of it. An environment sells a practice regimen that has already been tested on a strong student, which is a much rarer thing, because the only way to test a regimen is to run it and watch what happens.
Why the Same Product Sells Twice
The phrase “already tested” carries most of the commercial value. According to The Decoder, Chinese labs seek out the exact environments first sold to Anthropic and OpenAI because doing so lets them skip costly in-house development and the wasted training runs that come with it. A run that teaches a model nothing still burns computing time and money.
That waste matters more when computing is scarce. The investigation notes that Chinese chips suffer from acute memory bottlenecks, which makes every saved run “particularly valuable” to those labs. A lab with fewer good chips can’t afford to explore freely. It needs to know in advance which drills produce gains, and an environment that a frontier lab already trained on comes with that knowledge built in.
Distillation fills in the rest. Chinese labs also use outputs from American models such as GPT and Claude as teaching examples, so they get model answers from one source and the practice field from another; The Decoder calls the combination a two-pronged advantage. The authors still place Chinese labs at least a few months behind their American peers, which is a short distance in an industry that releases new models as often as this one does.
Airline training offers a close comparison. A new pilot rehearses an engine failure on takeoff in a full-motion simulator many times before ever meeting one in the air. The aircraft matters, and so does the pilot’s ability, but an airline’s hard-won experience lives in the scenarios. Each one encodes an incident that someone studied and turned into a drill.
The parts line up closely. The simulator is the environment, the scenario library is the task set, the instructor’s checklist is the grader, and the pilot is the model. An airline that licensed another carrier’s scenario library would get years of lessons without living through the incidents that produced them.
What the pairing shows is that the costly part of a curriculum was never the machinery. It was the record of what went wrong and how to catch it, and that record copies at almost no cost once it exists. Simulators are expensive to build; a well-made scenario is expensive to discover and nearly free to duplicate.
Rules Drawn Around Chips
American export controls were designed around physical things that can be counted at a border: advanced chips, the machines that make them, and the equipment that fills data centers. The Decoder’s summary is plain about how little of that reaches this trade: “Export controls barely apply because RL environments remain largely unregulated.”
Tony Hey, Stewart Tansley, and Kristin Tolle saw a version of this problem coming in The Fourth Paradigm, their book on data-driven science. Writing about international research, they observed: “Technology has rendered physical boundaries irrelevant, but legal jurisdictional boundaries remain.” Their concern ran the other way from today’s, since they worried that national laws would block scientific data that ought to flow between countries.
The same gap now runs in the opposite direction. An environment is software and data, so it moves the way research data moves: as a file, under a contract signed by an offshore entity, assembled in part by workers in a third country. The legal boundary still exists on paper, but nothing in the trade has to pass through a port or a customs desk.
Rules built for crates have a hard time with files. A chip export can be counted, inspected, and traced to a serial number. An environment can be copied a thousand times, and every copy is as good as the original, which leaves the law watching a door the goods no longer use.
The ChinaTalk authors argue that the federal office in charge of export rules should bring dual-use environments, such as those for coding and biology, under the controls that already exist. Their worry is specific: an environment that teaches a model to write better code can also teach it to find weak points in other people’s code. Whatever regulators decide, the request shows where the value has moved.
The Advantage Lives in the Curriculum
The usual account of the American lead in AI rests on two assets: the best chips and the best model weights. Both are hard to copy, and both are protected, to different degrees, by export rules and corporate secrecy. This trade shows a third asset carrying a large share of the advantage, and it is the one with the least protection.
That asset is the curriculum: the tasks, the graders, and the checked solutions that tell a model what good work looks like. A frontier lab learns which curriculum works by paying for the runs that fail. Every dead end it explores becomes part of what it knows, and the environments that survive that process carry the lesson inside them.
The curriculum is also the most labor-intensive part of the work. According to the investigation, American data firms outsource much of the environment building and data curation to low-paid workers in developing countries, especially India. The same subcontracting that lowers the cost also explains how the client names escaped into résumés.
So the asset at the center of the AI race is produced by a distributed workforce, packaged by a handful of American startups, and sold to any buyer with a budget. Nothing about that is sinister; it describes how most software gets made. It does mean that the advantage American labs paid for with their own failed runs becomes a product the moment someone else can buy the drills that came out of them.
Credit is due here, too. The data firms built something useful: a way to turn expert judgment into a repeatable training regimen that any lab can run. That is a real advance in how AI gets made, and it explains why labs on both sides of the Pacific pay for it. The friction comes from what the buyers inherit, which is the lesson of every expensive mistake an earlier customer already made.
When Practice Is Licensed Like Software
Should buying practice remain cheaper than building it, the data firms could start to look less like contractors and more like publishers of curricula. Mercor, Surge AI, and their peers may come to sell environments the way software companies sell licenses: per buyer, with terms on reuse, and with each copy marked so that a leaked environment can be traced to the customer who received it.
The frontier labs could respond by splitting their curricula in two. The Decoder notes that environments for biology, chemistry, cybersecurity, chip design, and investment banking are increasingly built in-house already. That boundary may widen, with general office and coding drills sold on the open market and the most valuable domains kept inside the lab, much as an airline might license basic simulator scenarios while guarding the ones built from its own incidents.
Regulators may eventually treat some environments the way they treat encryption software or advanced chips, with licenses required for the sensitive domains. The larger change could be commercial rather than legal. In a mature market, the price of an environment may depend on how many rivals have already trained on it, and exclusivity could become the feature buyers pay the most for.
What Is Worth Guarding
The trade in environments shows that a lab’s lead is partly a record of its own mistakes. Each failed run taught a frontier lab which drills work, and the environments that came out of those runs carry that lesson to any buyer. A memory-constrained lab that can buy the regimen saves the waste, and the gap it closes is measured in months rather than years.
Chips still set the ceiling on what a model can become. Practice decides how quickly a model reaches it, and practice is the part that travels as a file. Coyle’s point about people applies here with uncomfortable precision: errors are where skill comes from, and a curriculum built from someone else’s errors is skill at a discount.
A lead built from mistakes lasts only until someone sells the lessons.
Sources
Coyle, Daniel. The Talent Code.
Hey, Tony, Stewart Tansley, and Kristin Tolle. The Fourth Paradigm.
“US data brokers are selling half a billion dollars in RL training environments to Chinese AI labs.” the-decoder.com.
Leave a Reply