AI training data pricing models

There is no list price for AI training data. Each deal is priced on the dataset, the rights granted and the buyer's alternatives. What is consistent is the set of structures used to express a price, and the factors that move it up or down.

How is AI training data priced?

AI training data is priced using five structures, and many deals combine two of them. The right structure depends on whether the data is a one-off archive, a stream of new records or a catalog sold to many buyers.

Hybrids are common. ddiy.co's October 2026 compilation lists Labelbox Prism as paying an upfront fee plus a revenue share, and many flat-fee deals are paid in stages tied to acceptance.

What units are used in per-unit pricing?

Per-unit pricing uses the unit the buyer actually consumes. AIData.Broker, as quoted by ddiy.co, lists the company, the code repository, the hour of audio or video and the task as the usual units, each multiplied by volume.

Buyers usually price a company archive as a whole rather than per document, according to ddiy.co. That is why ranges for company data are quoted per company.

What do published price ranges look like?

Published price ranges for AI training data run from a few thousand dollars for research use to over a million dollars for exclusive or custom data. Phyllo's September 2026 guide gives reported market bands, which it says should be treated as ranges rather than quotes.

For company data, buyers publish their own ranges. Mercor's listing, as compiled by ddiy.co in October 2026, puts law firms at $50,000 to $1.2 million, construction companies at $115,000 to $1 million and call centers at $20,000 to $250,000 per company. These are offer ranges, not completed prices.

Completed prices tend to sit lower. AIData.Broker reports a median of $280,000 across buyers' published offers but a median of $31,600 across the few sales made public. Fortune reported in September 2026 that payments are typically $10,000 to $100,000 per company, citing SimpleClosure, while Google bid $10 million for bankrupt Spirit Airlines' data archive in an August 2026 auction that was subject to court approval when reported.

Our AI data deals tracker lists reported transactions, and how much is your data worth to AI companies explains the same numbers from the seller's side.

One dataset is worth more than another when it is harder to replace and safer to use. Volume matters less than most sellers expect.

• Exclusivity: AIData.Broker puts exclusive deals at about 2.5 to 5 times non-exclusive ones.
• Rights clarity: a smaller dataset with clear provenance can be worth more than a larger one whose best parts must be excluded, according to Nixon Peabody.
• Human authorship: DataVendor reports that labs pay more for code written mostly by people, and for complex products.
• Language and difficulty: DataVendor names systems languages such as Rust, Go, C and C++ as attracting a premium.
• Context: production code with pull requests, issue tickets, runbooks and incident notes explains why code changed, which DataVendor says buyers value.
• Reproducibility: DataVendor notes that evaluation sets command more attention when independent reviewers can reproduce the scores.

License scope moves price as well. A license limited to one model, one field of use or a fixed term should cost less than a perpetual license covering every model your company builds.

How does exclusivity change the price?

Exclusivity raises the price because the seller gives up every other sale of the same data. AIData.Broker's estimate of 2.5 to 5 times a non-exclusive deal reflects that lost income.

You can narrow exclusivity to pay less for it. Common options are a time-limited exclusive, an exclusive within one field such as legal drafting, or an exclusive against named competitors only. Polyshares, by contrast, makes exclusive and perpetual the default in its published deals, according to ddiy.co. Our guide to exclusive vs non-exclusive data licenses covers the trade-offs.

How should a buyer structure payment?

A buyer should structure payment so that money follows evidence. Pay a first installment on signing, the main payment on acceptance of the delivered data and, for larger deals, hold back a portion until the warranty period ends.

Published seller timelines show what the market expects. ddiy.co's compilation reports that Mercor pays 2 to 4 weeks after signing, micro1 takes up to 30 days to approve and pays 30 days after invoice, and Polyshares pays on net 30 to 60 day terms. Matching your schedule to these norms makes an offer easier to accept.

Which pricing model should you choose?

The pricing model you choose should follow how the data will be used and how often it changes.

What does Sell My Data To AI charge buyers?

Sell My Data To AI charges buyers nothing to submit a requirement or review a shortlist. Any fee is agreed in writing before an introduction, and our fees page explains how this works. You negotiate the price and payment terms directly with the seller.

To get a view of what a dataset should cost before you negotiate, book a buyer consultation or submit a data requirement.