Nine American vendors selling RL environments to Chinese labs is a striking number, and the half-billion-dollar figure makes it hard to wave off as a rounding error. Export controls have spent years focused on chips, yet if most of the gains between model versions come from post-training, the simulated worlds and grading rubrics may be where a lot of real capability transfer happens. A PowerPoint clone sounds harmless. A cyber or bio environment with a well-built rubric is a different story. Worth watching whether Commerce tries to draw lines by domain, since a blanket ban would be hard to enforce and easy to route around through third-country shells. Curious how many of these vendors even know who the end customer is.
A sharp piece. Technology transfer increasingly happens through services and data rather than hardware, which makes traditional export controls harder to enforce.
Excellent work, Aqib and Quinn. You explain this part of the AI trade with unusual clarity.
Your point about supplier experience is particularly important. A lab can pay for an exclusive training environment, yet the supplier may still use what it learned to do better work for the next customer. That is a difficult thing to contain through a contract.
I also appreciated how seriously you examined the costs of broad restrictions, including the possible harm to safety research and useful competition. Giving those objections a fair hearing makes your recommendations more persuasive.
One difference from chip controls is worth spelling out. An RL environment is mostly a simulator plus a grading rubric, so after one sale it can be copied at almost no cost. A GPU has to be shipped, serviced and eventually replaced. That makes per-shipment licensing, the chip model, a poor fit here. The more workable lever is probably on the vendor side, with know-your-customer rules for lab customers like the ones proposed for cloud compute. It also matters where the expert graders are contracted, because rubric quality is the scarce input, not the code.
The codename detail is what stuck with me. If the people building these environments don't know who the end buyer is, a customer policy is just a promise. Mercor moving from an ownership test to a location test in May tells me the vendors aren't sure where the line is either.
One thing I'd add. The experts writing the cyber and bio graders are exposed too. Whoever buys these environments learns what the US thinks a good answer looks like. I have no evidence anyone is targeting those contractors, but I'd want counterintelligence looking at it, not only BIS.
I agree that visibility comes first. My question is who in government is tasked with collecting it. I don't know of anyone who owns that today.
FWIW I think the codename section may have been a bit exaggerated. Most likely, a company knows its end customer, but the experts and engineers contracted to make the environment/rubric may only know them as a codename.
This helps explain a split I've been puzzling over. On OpenRouter, Chinese models carried 72% of tokens in late October 2026, up from 7% in early 2025 — developer traffic only (OpenRouter data via Dror Poleg). Epoch AI still has the US at about 72% of AI compute delivered since 2022, versus about 10% for China. That gap isn't only US export rules: Reuters reported this month that Beijing is withholding large-scale H200 import approval, after last year's CAC and NDRC push off the H20 toward domestic chips (Bloomberg, FT). If the official bet is Huawei, do these RL-environment sales help keep model quality up while that shift happens?
Which makes naming the deliverable the hard part. Chips are discrete, shippable, and countable; an RL environment may arrive as access, a dataset, or an ongoing service — and each requires a different regulatory instrument. A chip rule can be enforced perfectly and still miss the transfer at issue.
The key point is that RL environments sell judgement, not data: the rubric is the distilled expertise of whoever wrote it. Chip controls target compute, but post-training is where capability gets cheap, and a grading rubric fits in an email. Hard to export-control slide-deck taste without export-controlling people.
Finance being on the RL list next to coding and law makes me curious who writes those grading rubrics. If a handful of contracted experts decide what a good answer looks like, isn't the rubric the real export, more than the PowerPoint clone?
The contrast with distillation is what makes this so hard to police. In the NSA, CISA and FBI advisory of September 8 (AA26-251A), the named labs, StepFun and Z.AI among them, reached US models through APIs, aggregators and reseller "transfer stations", so the model providers at least had terms to enforce and traffic to watch. Buying the same RL environment under a client codename breaks no terms and shows up in no US lab's logs. That makes your visibility point the crux: API providers can correlate accounts, but nobody correlates vendor sales. Would a reporting duty for vendors selling dual-use environments get that visibility faster than new BIS rules?
The contrast with distillation is what makes this so hard to police. In the NSA, CISA and FBI advisory of September 8 (AA26-251A), the named labs, StepFun and Z.AI among them, reached US models through APIs, aggregators and reseller "transfer stations", so the model providers at least had terms to enforce and traffic to watch. Buying the same RL environment under a client codename breaks no terms and shows up in no US lab's logs. That makes your visibility point the crux: API providers can correlate accounts, but nobody correlates vendor sales. Would a reporting duty for vendors selling dual-use environments get that visibility faster than new BIS rules?
The contrast with distillation is what makes this so hard to police. In the NSA, CISA and FBI advisory of September 8 (AA26-251A), the named labs, StepFun and Z.AI among them, reached US models through APIs, aggregators and reseller "transfer stations", so the model providers at least had terms to enforce and traffic to watch. Buying the same RL environment under a client codename breaks no terms and shows up in no US lab's logs. That makes your visibility point the crux: API providers can correlate accounts, but nobody correlates vendor sales. Would a reporting duty for vendors selling dual-use environments get that visibility faster than new BIS rules?
Nine American vendors selling RL environments to Chinese labs is a striking number, and the half-billion-dollar figure makes it hard to wave off as a rounding error. Export controls have spent years focused on chips, yet if most of the gains between model versions come from post-training, the simulated worlds and grading rubrics may be where a lot of real capability transfer happens. A PowerPoint clone sounds harmless. A cyber or bio environment with a well-built rubric is a different story. Worth watching whether Commerce tries to draw lines by domain, since a blanket ban would be hard to enforce and easy to route around through third-country shells. Curious how many of these vendors even know who the end customer is.
A sharp piece. Technology transfer increasingly happens through services and data rather than hardware, which makes traditional export controls harder to enforce.
Whew! Powerful, powerfully important. Thank you for this work!
Great deep dive!
Thank you! Your work was great source material for getting us up to speed with the data industry
Excellent work, Aqib and Quinn. You explain this part of the AI trade with unusual clarity.
Your point about supplier experience is particularly important. A lab can pay for an exclusive training environment, yet the supplier may still use what it learned to do better work for the next customer. That is a difficult thing to contain through a contract.
I also appreciated how seriously you examined the costs of broad restrictions, including the possible harm to safety research and useful competition. Giving those objections a fair hearing makes your recommendations more persuasive.
One difference from chip controls is worth spelling out. An RL environment is mostly a simulator plus a grading rubric, so after one sale it can be copied at almost no cost. A GPU has to be shipped, serviced and eventually replaced. That makes per-shipment licensing, the chip model, a poor fit here. The more workable lever is probably on the vendor side, with know-your-customer rules for lab customers like the ones proposed for cloud compute. It also matters where the expert graders are contracted, because rubric quality is the scarce input, not the code.
The codename detail is what stuck with me. If the people building these environments don't know who the end buyer is, a customer policy is just a promise. Mercor moving from an ownership test to a location test in May tells me the vendors aren't sure where the line is either.
One thing I'd add. The experts writing the cyber and bio graders are exposed too. Whoever buys these environments learns what the US thinks a good answer looks like. I have no evidence anyone is targeting those contractors, but I'd want counterintelligence looking at it, not only BIS.
I agree that visibility comes first. My question is who in government is tasked with collecting it. I don't know of anyone who owns that today.
FWIW I think the codename section may have been a bit exaggerated. Most likely, a company knows its end customer, but the experts and engineers contracted to make the environment/rubric may only know them as a codename.
This helps explain a split I've been puzzling over. On OpenRouter, Chinese models carried 72% of tokens in late October 2026, up from 7% in early 2025 — developer traffic only (OpenRouter data via Dror Poleg). Epoch AI still has the US at about 72% of AI compute delivered since 2022, versus about 10% for China. That gap isn't only US export rules: Reuters reported this month that Beijing is withholding large-scale H200 import approval, after last year's CAC and NDRC push off the H20 toward domestic chips (Bloomberg, FT). If the official bet is Huawei, do these RL-environment sales help keep model quality up while that shift happens?
Which makes naming the deliverable the hard part. Chips are discrete, shippable, and countable; an RL environment may arrive as access, a dataset, or an ongoing service — and each requires a different regulatory instrument. A chip rule can be enforced perfectly and still miss the transfer at issue.
The key point is that RL environments sell judgement, not data: the rubric is the distilled expertise of whoever wrote it. Chip controls target compute, but post-training is where capability gets cheap, and a grading rubric fits in an email. Hard to export-control slide-deck taste without export-controlling people.
Finance being on the RL list next to coding and law makes me curious who writes those grading rubrics. If a handful of contracted experts decide what a good answer looks like, isn't the rubric the real export, more than the PowerPoint clone?
The contrast with distillation is what makes this so hard to police. In the NSA, CISA and FBI advisory of September 8 (AA26-251A), the named labs, StepFun and Z.AI among them, reached US models through APIs, aggregators and reseller "transfer stations", so the model providers at least had terms to enforce and traffic to watch. Buying the same RL environment under a client codename breaks no terms and shows up in no US lab's logs. That makes your visibility point the crux: API providers can correlate accounts, but nobody correlates vendor sales. Would a reporting duty for vendors selling dual-use environments get that visibility faster than new BIS rules?
The contrast with distillation is what makes this so hard to police. In the NSA, CISA and FBI advisory of September 8 (AA26-251A), the named labs, StepFun and Z.AI among them, reached US models through APIs, aggregators and reseller "transfer stations", so the model providers at least had terms to enforce and traffic to watch. Buying the same RL environment under a client codename breaks no terms and shows up in no US lab's logs. That makes your visibility point the crux: API providers can correlate accounts, but nobody correlates vendor sales. Would a reporting duty for vendors selling dual-use environments get that visibility faster than new BIS rules?
The contrast with distillation is what makes this so hard to police. In the NSA, CISA and FBI advisory of September 8 (AA26-251A), the named labs, StepFun and Z.AI among them, reached US models through APIs, aggregators and reseller "transfer stations", so the model providers at least had terms to enforce and traffic to watch. Buying the same RL environment under a client codename breaks no terms and shows up in no US lab's logs. That makes your visibility point the crux: API providers can correlate accounts, but nobody correlates vendor sales. Would a reporting duty for vendors selling dual-use environments get that visibility faster than new BIS rules?