Training-data licensing: selling the archive to the model makers

How it makes money

Every publisher and platform holds an archive: years of articles, books, forum threads and answers, written for readers and earning little once the readers move on. Large language models changed what that archive is worth. They need vast amounts of well-written, well-organized text, and the companies that hold it can charge for access.

In business model terms, the payer changes. The reader or the advertiser paid for the content when it was new. Now a model maker pays for it again, for a different use, under a contract that may run for years or close in a single payment.

The platform: Reddit

Reddit says it has explored content licensing for purposes including "training generative AI models," and it is candid about what that is: "a novel business model without an established track record." The revenue line that holds it is growing. Reddit's other revenue, which is content licensing plus user products it calls not material, was $48.0 million in the third quarter of 2026, up from $35.6 million a year earlier.

Reddit books licensing revenue gradually, "generally ratably over the license period," so its contracts show the road ahead. At September 30, 2026, its long-term contracts, primarily content licensing, held $62.4 million still to be recognized: $31.9 million for the rest of 2026 and $30.5 million for 2027.

The publisher: Wiley

John Wiley & Sons, a publisher of research and learning materials, shows the other contract shape. It licenses content for AI "generally at a fixed transaction price, with revenue recognized at a point in time." The result is lumpy. In the year to April 30, 2026, its research AI licensing revenue was $33.1 million, against about $11 million a year earlier, while its learning AI licensing revenue fell to $16.0 million from about $29 million. It has also begun licensing other publishers' content to AI developers, as a service.

How it fails

Reddit lists the ways this could go wrong in a single sentence: the market for large language models "may not develop or may be limited by regulation or other factors," the value of content for AI training "may be reduced over time," and similar terms may not be available from other licensees. A deal priced once is not a deal that renews.

Questions

What is training-data licensing?

A business model in which a publisher or platform charges AI developers for the right to use its content, such as articles, books or user posts, to train and run their models.

Which companies license their content to AI developers?

Among public companies, Reddit and John Wiley & Sons both report revenue from licensing content for AI in their SEC filings.

How much do AI content licensing deals pay?

Public figures are scarce. Wiley reported $33.1 million of research AI licensing revenue and $16.0 million of learning AI licensing revenue for the year to April 30, 2026.