What the 2024 to 2026 data deals tell us about prices

Almost every training data license is private. The few numbers that reach the public come from securities filings, earnings releases and reporters who saw documents, so they are skewed toward large, listed sellers. Read together, they still say a good deal about how this market prices data, and about what a buyer should and should not take from them.

The public figures, in one table

Seller or data	Figure	What it covered	Reported by
Reddit	$203.0 million	Aggregate contract value of licensing deals from January 2024, two to three year terms	Reddit’s S-1, via TechCrunch, Feb 2024
Shutterstock	$104 million	AI licensing revenue in 2023	Bloomberg, Jun 2024
Wiley	$23M, then $40M	AI licensing revenue, fiscal 2024 and fiscal 2025	Wiley results, Jun 2025
Informa and Microsoft	About $10 million	Initial access fee, plus recurring payments over three years	Informa update, May 2024
Spirit Airlines business data	$10 million bid	Google’s bid for employee emails, messages and files; Mercor bid $7.5 million	Reuters, Aug 2026
Unused video footage	$1 to $4 a minute	Rates AI companies were paying creators	Bloomberg, Jan 2025
HVAC company records	About $150,000	Paid by data labeling startups to individual firms	The Information, 2026
A startup’s codebase	About $25,000	What one founder netted after a wind-down	MIT Technology Review, Oct 2026

Pattern one: scale and uniqueness set the price

The largest numbers belong to sellers with something nobody else has at that size. Reddit holds years of public conversation; Shutterstock holds an image and video library; Wiley and Informa hold academic publishing catalogs. None of it can be recreated by commissioning.

At the other end, a single small company’s records are worth tens of thousands of dollars, not millions. About $150,000 handed to individual HVAC firms by labeling startups (The Information’s figure), and about $25,000 for one founder’s codebase, are what operational data from one firm fetches when the buyer wants many such firms.

What this means for a buyer

If your gap is breadth, you are competing with the largest labs for the same few catalogs, and the price will reflect that. If your gap is depth in one narrow domain, a set of smaller licenses or commissioned collection may cost far less than a headline deal suggests.

Pattern two: the big deals are multi-year

Reddit’s agreements ran two to three years. Informa’s arrangement with Microsoft paired an initial fee with recurring payments over three years. Wiley’s AI licensing revenue rose from $23 million in fiscal 2024 to $40 million in fiscal 2025, which suggests deals that repeat or renew rather than one-off sales.

That structure matters because models outlive licenses. A buyer should know what happens to a model trained during the term once the term ends, and whether a refreshed version of the data comes with renewal.

Pattern three: corporate data has become an asset class

The Spirit Airlines case shows how far the market has moved. In August 2026 a Reuters report described a $10 million offer from Google for the airline’s internal records: staff email, Teams chat, spreadsheets, calendars. The records were to be stripped of identities before any sale, a judge had yet to sign off, and Mercor had put in $7.5 million.

An airline’s internal messages are not what most people picture as a data asset, yet two AI buyers bid for them. For a buyer, the lesson is about provenance: data that comes out of a bankruptcy or a wind-down needs especially careful checks on consent and personal information.

Pattern four: media is priced by the unit

Video stands apart because it is priced like a commodity. Bloomberg’s $1 to $4 a minute is the closest thing the market has to a list price. A fourfold spread inside one medium says that even commodity data is priced on what the footage shows, not just how much of it there is.

Why you will rarely get a per-unit price for text

Text and code licenses are almost always quoted as a lump sum for a corpus, because value depends on what the corpus covers rather than its word count. Per-unit rates show up again on the commissioning side, as per-object labels and per-hour expert time; the pricing models primer compares them.

What the figures do not tell you

• What anyone else will pay. Each figure reflects one negotiation, one data set and one moment.
• Price per token or per record. None of the large deals disclosed volume in a way that converts cleanly.
• What sellers would accept from a smaller buyer. Headline deals were struck with the largest labs.

Treat them as bounds on the market, not as comparables for your own quote. The AI training data companies indexed on this site mostly price by quote, and the only number that matters for your project is the one a vendor commits to in writing after scoping.

A headline figure tells you what one buyer valued; your own model’s error rate tells you what the data is worth to you.