The record
This is a market, not a found record. A leading lab commissions practice tasks from a vendor, each with instructions, a working environment and a checker that marks the answer. The lab trains on the set until its model passes nearly all of it. The set is then called saturated, and the lab stops using it. The idea examined here is a second sale. A weaker model might still learn from a set the leaders have used up, so the buyer could be an American lab behind the leaders, a lab in China, or a business tuning a small open model. Goods sold this way, to many buyers from a catalogue, are called off-the-shelf.
An illustration, not a real contract: a vendor builds 500 spreadsheet-modelling tasks for a leading lab, using its own experts and marking schemes. A year later it repackages the same task designs as a catalogue item. A lab abroad sends vendors a list of subjects it wants covered, and the vendor licenses the catalogue item to that lab and to three others. The decision that matters is the repackaging, and the vendor makes it, because the vendor holds the rights.
Why it might matter for training AI
A model one step behind the leaders can still learn from tasks the leaders have outgrown. That is the whole case.
The trade has been reported at scale. Forbes reported in August 2026 that American data vendors, among them Surge, Mercor, AfterQuery and Turing, sell training data to Chinese labs, and that Chinese buyers ask for the datasets American labs have already bought 1 2. The article was read here only in its opening and through other outlets' summaries, so its figures are left out.
Who holds it and who may share it
As a rule the dataset stays the vendor's and the lab is licensed to use it. Only one such contract is public, and it is clear on the point. Under Proximal's standard terms, what it delivers remains Proximal's own property, and the customer's licence is for internal use and cannot be transferred or sub-licensed 3. No other lab or vendor contract was read, so how typical this is has not been established.
On those terms:
- A lab cannot resell a set once it is done with it.
- A vendor can, unless it sold the lab exclusive use, and then only once the exclusive period ends.
- For a company's raw internal records, the company keeps ownership and the buyer holds a licence. The General Data Company tells sellers that its holdings may be licensed again to later buyers as time goes on 4.
- A reseller standing outside this chain owns none of it, and would need written permission from a vendor that already has its own sales staff.
On selling to China, the law is thin. Lawfare wrote in September 2026 that almost no law governs how data providers deal with Chinese AI companies 5. One rule reaches some data. The US Justice Department's Data Security Program, in force since April 2025, prohibits selling or licensing bulk sensitive personal data about Americans to persons linked to China, and its guidance says the data counts even when it has been anonymised 6 7. Tasks written by experts that contain nothing about any individual appear to fall outside it, on Lawfare's reading 5. The volume thresholds in the rule were not read. No bill or new rule extending the programme to training data was found.
Contracts are moving ahead of the law. In its bid for the records of a bankrupt airline, micro1 promised that the assets would never be sold or passed to a Chinese entity 8.
No Australian holder is relevant.
Who is already there
The vendors that built the data are the resellers: the companies Forbes named, and others. Scale AI is reported to have walked away from a Chinese contract over national-security concerns 9.
Marketplaces fill the go-between role. DataVendor, provided by HUD, matches sellers' data, code and technical expertise with AI companies that need training and evaluation supply 10. Protege pools data from rights-holders and licenses each item to more than one buyer.
Chinese vendors are emerging to supply their home market 11.
What would have to be stripped
Anything about real people, before a sale abroad: the Justice Department rule covers health, financial, biometric and location data and personal identifiers 7. A reseller also inherits every rights problem in the original set, including work files that contributors had no right to hand over.
What stands in the way
- Vendors hold the rights and are already reselling. An outsider has nothing to hold.
- A funded marketplace already takes the go-between's cut.
- Sales to Chinese labs could be closed by a single extension of an existing rule, and the proposal has been written up in public.
- Selling to Chinese labs is now covered in the press as a national-security problem.
- Leading labs pay extra for exclusivity precisely to prevent resale.
- Each resale needs fresh proof that the set still teaches the buyer's model.