Since mid-2023, AI developers have signed dozens of agreements to license text, images, code, music, and clinical records, and a handful of those agreements have had their prices reported. This is what the public record says, and what it implies about how a dataset gets valued.
The deals on the record.
Most AI data deals are announced without numbers. The figures below come from securities filings, earnings calls, court filings, and reporting by outlets such as Reuters, Bloomberg, and The Wall Street Journal. Where a figure is a journalist’s estimate rather than a company disclosure, the table says “reported.”
| Deal | Data | Reported terms | Year | Source |
|---|---|---|---|---|
| Reddit → Google | Data API: user posts and comments | Reported about $60M per year (Reuters) | 2024 | [2] |
| Reddit → undisclosed licensees | Same API, several buyers | $203.0M aggregate contract value, two- to three-year terms, per S-1; at least $66.4M recognized in 2024 | 2024 | [1] |
| Reddit → OpenAI | Data API plus ads partnership | Not disclosed | 2024 | [3] |
| Associated Press → OpenAI | Text archive back to 1985 | Not disclosed; AP receives technology access | 2023 | [4] |
| Axel Springer → OpenAI | Politico, Business Insider, Bild, Welt | Reported “tens of millions of euros” per year, three years (Bloomberg) | 2023 | [5] |
| News Corp → OpenAI | WSJ, Barron’s, The Times, and others, current and archive | Reported $250M over five years, cash and credits (WSJ) | 2024 | [6] |
| Dotdash Meredith → OpenAI | Lifestyle and finance titles | At least $16M per year fixed, plus a variable component, per IAC filings | 2024 | [7] |
| The New York Times → Amazon | Articles, The Athletic, NYT Cooking | Reported $20–25M per year (WSJ) | 2025 | [8] |
| News Corp → Meta | U.S. and U.K. titles, retrieval and archive training | Reported up to $50M per year, at least three years (WSJ) | 2026 | [9] |
| Shutterstock → multiple labs | Images, video, music, metadata | $104M AI licensing revenue in 2023; $250M targeted for 2027; individual deals reported at $25–50M | 2023–24 | [10][11] |
| Photobucket → in negotiation | 13 billion photos and videos | Reported asks of 5 cents to $1 per photo, more than $1 per video (Reuters) | 2024 | [11] |
| Taylor & Francis → Microsoft | About 3,000 academic journals | $10M in year one, recurring payments for three years; parent Informa expected AI revenue above $75M for 2024 | 2024 | [12] |
| Wiley → unnamed AI companies | Books and journal content | $23M earned plus $21M expected in 2024; $49M AI revenue in fiscal 2026, up 23% | 2024–26 | [13][14] |
| HarperCollins → unnamed AI company | Nonfiction backlist, author opt-in | $5,000 per title, split 50–50 with the author; output capped at 200 consecutive words or 5% of a book | 2024 | [15] |
| Stack Overflow → Google, OpenAI | OverflowAPI: 58M-plus questions and answers | Not disclosed; non-exclusive | 2024 | [16] |
| Tumblr and WordPress.com → OpenAI, Midjourney | Public posts, 2014–2023 | Not disclosed; reported from internal documents | 2024 | [17] |
| Tempus → Pathos AI (with AstraZeneca) | De-identified multimodal oncology records | $200M in data licensing and model development fees, multi-year | 2025 | [18] |
| Harvard Health Publishing → Microsoft | Consumer health content for Copilot | Licensing fee paid; amount not disclosed | 2025 | [19] |
| Universal Music ↔ Udio | Recorded music and publishing catalog | Settlement plus new licenses; amounts not disclosed; licensed platform in 2026 | 2025 | [20] |
| Warner Music ↔ Suno | Catalog, artist opt-in | Settlement; licensed models in 2026, old models deprecated; amounts not disclosed | 2025 | [21] |
| Authors’ class ↔ Anthropic | About 465,000 pirated books | $1.5B settlement, about $3,000 per work (court-approved) | 2025 | [22] |
Two things stand out. First, the largest confirmed numbers are not licenses at all: the Anthropic settlement priced a pirated book at roughly $3,000, which is in the same range as the $5,000 HarperCollins negotiated per title with author consent [15][22]. Second, the only sellers with audited, recurring AI revenue are companies that already ran licensing businesses: Shutterstock, Wiley, Informa, and Reddit, each of which reports the line in filings [1][10][12][14].
What the prices have in common.
Almost every reported figure is annual, not lump-sum. Reddit’s S-1 spreads $203M across two- to three-year terms [1]; Axel Springer’s is per year over three years [5]; News Corp’s $250M is a five-year total [6]; Dotdash Meredith books about $4M a quarter [7]. Buyers want continuing access to fresh data and the right to re-license, and sellers want a renewal lever.
Payments are mixed. The News Corp deal was reported as cash plus credits for OpenAI’s technology [6]; AP took technology access alongside whatever cash was paid [4]; Dotdash Meredith has a fixed floor plus a variable element [7]. Pure cash deals exist (Tempus, Taylor & Francis) but are the exception among news publishers.
Scale buys price, up to a point. The per-year numbers cluster between $10M and $60M for a single large source, whether that is a news group, a journal publisher, or a social platform [2][6][8][9][12]. Prices per unit are small: cents to a dollar per image, a fraction of a cent per word, $5,000 per book [11][15]. What makes a deal large is volume multiplied by renewal, not an exotic rate.
Litigation has set a floor. Courts in the United States have held that training on lawfully acquired books can be fair use, while acquiring them from pirate libraries is not [23]. That is why Anthropic paid $1.5B for the sourcing, not the training [22]. The lesson labs have drawn is that paying for clean, consented data is cheaper than defending unclean data.
How buyers value a dataset.
Uniqueness. A buyer pays for what it cannot scrape. Reddit’s S-1 presented its platform data as a key input to leading language models precisely because it is not reproducible elsewhere [1]. Tempus commands $200M because de-identified, multimodal oncology records with outcomes do not exist on the open web [18]. Public news archives, by contrast, are largely already in training sets, which is one reason their prices are measured in tens rather than hundreds of millions.
Rights clarity. The seller must be able to grant what it is selling. Photobucket changed its terms of service to secure “unrestricted rights” before shopping its archive, and a data broker told Reuters it still avoided such platforms because the rights were murky [11]. The Tumblr episode shows the failure mode: internal documents indicated private and deleted posts had been swept into a dataset meant for sale [17]. HarperCollins solved the problem by asking each author individually [15]. A buyer discounts, or walks, when consent is unclear.
Volume and density. Price scales with how much usable signal the dataset carries, not with raw bytes. Shutterstock sells images with captions, keywords, and metadata, and its AI line reached $104M in 2023 [10]. Stack Overflow sells 58 million question-and-answer pairs that are already structured and voted on [16]. Labeled, structured, deduplicated data is worth multiples of the same material in a raw dump.
Structure of the deal. Recurring access is worth more than a one-time copy. Every large reported deal runs two to five years [1][5][6][9]. Sellers with a live API, such as Reddit and Stack Overflow, can sell freshness and charge for continuity. A static archive is usually a lump sum, and the lump is smaller.
Exclusivity. Non-exclusive is the norm and the prices above reflect it: Stack Overflow said its OpenAI deal was not exclusive [16]; Taylor & Francis described its Microsoft license as non-exclusive [12]; Shutterstock sells the same library to several labs [10][11]. Exclusivity raises the price substantially, but it also forecloses the second and third sale, so most sellers decline it.
Provenance and documentation. Regulation is converting documentation into price. Since August 2, 2025, providers of general-purpose AI models in the EU must publish a summary of training content using the Commission’s template and keep a copyright-compliance policy, with earlier models given until 2027 [26]. The U.S. Copyright Office’s May 2025 report declined to declare training fair use as a category and pointed to licensing as the path for commercial, market-substituting uses [25]. A dataset that arrives with its chain of custody written down clears a buyer’s legal review faster, and speed is worth money to a lab on a training schedule.
What this means if you hold business data.
Most of the deals above involve media companies, because media companies already had licensing desks, counsel, and archives in sellable form. The data that AI developers now say they lack is different: operational records, customer-service transcripts, domain documents, expert workflows, lab and clinical data. The Tempus and Harvard deals are the early signs of that shift [18][19], and the analyst estimates follow it: Grand View Research puts the AI training dataset market at an estimated $16.3B by 2033, growing about 22.6% a year [28].
If you run a company that holds this kind of data, the record suggests four practical points. Your data is worth more if it is unusual, not if it is large. It is worth more if you can show you have the right to license it: contracts, consents, and a clear account of where personal information sits. It is worth more if it comes with structure, such as the outcome of each case, ticket, or transaction, not only the text. And it is worth more if it ships with documentation that a buyer’s lawyers can read in an afternoon.
Caveats.
Most of the figures above are reported, not confirmed. The $60M Reddit–Google figure, the $250M News Corp–OpenAI figure, the Axel Springer range, the NYT–Amazon range, and the News Corp–Meta figure all come from journalists’ sources, and the parties either declined to comment or did not confirm them [2][5][6][8][9]. The figures that are confirmed are the ones in filings and court orders: Reddit’s $203M contract value, Dotdash Meredith’s $16M floor, Shutterstock’s $104M, Wiley’s $49M, Tempus’s $200M, and the $1.5B Anthropic settlement [1][7][10][14][18][22].
Deals also differ in what is being bought. Some cover training only, some cover retrieval and display in a chatbot, and some bundle advertising or technology credits [3][6][9]. Per-year figures are not comparable across those structures. Settlements are not licenses: they price past infringement plus the cost of future permission, and they are shaped by litigation risk as much as by data value [20][21][22].
Policy is unsettled. The U.K. government dropped its preferred text-and-data-mining exception in March 2026 and said it would keep market-led licensing under review [27]. A U.K. court found in November 2025 that model weights are not copies of the works they were trained on, and that training done abroad is outside the court’s reach [24]. U.S. courts have split on how heavily to weigh market harm [23]. Prices will move with those decisions. The direction so far, across every jurisdiction with a ruling, has been toward paying for data that is clean and away from scraping data that is not.