AI Data Licensing Deals: The Deal Tracker With Parties, Dates, Values and Sources

AI data licensing deals are agreements in which an AI developer pays a data holder for access to content or records, for training, for display in AI answers, or both. This tracker lists 74 such deals announced between 2022-10-25 and 2026-10-22, as of November 9, 2026: nine carry an amount stated by a party or a filing, 15 state training use, 32 cover display or grounding without stating training, and one excludes training.

The largest stated amounts come from securities filings. Reddit's Form S-1 of 2024-02-22 reports USD 203.0 million in aggregate contract value for data licensing arrangements entered in January 2024, with counterparties it does not name, and Tempus AI's 10-K for fiscal 2025 reports USD 200 million in data license fees over three years from Pathos AI. Rows run newest first, each with its primary source, and cover the deal side of AI training data licensing without press estimates.

What the deal tracker records and what counts as a deal

The tracker records every deal whose buyer, seller and date are confirmed by a party or a filing, with the primary source that confirms them. A deal counts when a named data holder grants an AI developer access to content or records through a license, a data API agreement, a program enrollment or a court-supervised sale, even when only the seller confirms it, as with Wiley's two unnamed-buyer projects of 2024.

Each row carries six fields:

• Buyer: the AI developer or program as the source names it, or "not named".
• Seller: the data holder that grants the rights.
• Announced: the date of the party statement or filing, or the month where only a month is public.
• Data type: one of nine groups, from news to company records.
• Use: training, display, training excluded or not stated, taken from the parties' own words.
• Value as stated: the amount and the party or document that states it, or "n/d".

Values that appear only in press reports are not shown, not even as estimates, and every count on this page comes from the rows below.

The deal tracker: every recorded deal, newest first

The tracker lists 74 deals, newest first by announced date, with the value only where a party or filing states it. The table compares buyer, seller, data type, use and value as of 2026-11-09; the Source column names the site of each row's primary document.

• Use: T = training or model improvement stated by a party or filing; D = display, grounding or answers, training not stated; X = training excluded by a party; n/s = use not stated.
• Value: n/d = not disclosed, because no party or filing states an amount.
• Order: month-only dates sort first within their month, then sellers run alphabetically.
• Links: a seller with its own deal page is linked from its first row only.

Deals by year and by data type

News publishers account for 43 of the 74 deals, and 2024 is the year with the most announcements, 27. The two tables count the tracker rows by announced year and by data type as of 2026-11-09, and the 2026 figure runs only to November 9.

Announcements rose from one in 2022 to 27 in 2024, then 24 in 2025 and 20 in 2026 to November 9. News leads partly because program announcements name several publishers on one day: Microsoft's Publisher Content Marketplace accounts for seven rows dated 2026-02-03 and Copilot Daily for five dated 2024-10-01. Platform content (8 rows) covers Reddit, Stack Overflow, Automattic, Wikimedia Enterprise and Yelp, and music (7) covers the agreements of Udio, Suno, Stability AI, Klay Vision and ElevenLabs with Universal Music Group and Warner Music Group.

Which deal values are public, and who stated them

Nine of the 74 deals carry an amount stated by a party or a filing; the rest disclose no amount, or state a pricing basis without one. The table sets each stated amount beside the party and document that state it, as of 2026-11-09, and says what the amount is not.

No party states an amount for the HarperCollins or Disney agreements, and figures that appear only in press reports are not shown for any deal.

Two of the nine amounts show the limits of the public record. Wiley's own documents of 2024-09-05 disagree on how much of the second project fell into the first quarter of fiscal 2025: the earnings release gives USD 16 million and the earnings presentation USD 17 million, and both stay visible. Reddit's 10-Q for the third quarter of 2026, filed 2026-10-29, adds that "substantially all" of its licensing contract value comes from two partners whose agreements are "approaching the end of their respective current terms", without naming either. For HarperCollins, the Authors Guild, which is not a party, states a fee of USD 5,000 per title split 50-50 between the opted-in author and the publisher (2024-11-19), so the row shows no party-stated value.

The Tempus filings are the only ones in the tracker that price data per unit: USD 100 per delivered de-identified record in the neurology agreement of 2026-10-22. Other parties state a pricing basis without a number. Microsoft states that publishers in its Publisher Content Marketplace "will be paid on delivered value" (2026-02-03), Google states that it pays participating publications in its News AI pilot (2025-12-10), and Warner Music Group states that all of its AI deals are based on variable economics (letter of 2026-03-03). The termination and deferred services fees in SpaceX's 424B4 for the Cursor agreement of 2026-04-19 apply only in defined termination cases and are not a price for developer data. How such amounts surface as licensing revenue is traced in AI data licensing revenue in company filings.

Training, display or both: what each license is stated to cover

Fifteen deals state training or model improvement; 32 cover display, grounding or answers without stating training. The counts come from the parties' own words as of 2026-11-09:

• Training stated (T), 15 deals: Google and Reddit, OpenAI with Axel Springer, Dotdash Meredith, the Financial Times, Le Monde, Prisa Media, Shutterstock and Vox Media, Wiley (two projects), Kopinor, Lionsgate, Udio with Universal Music Group, Cursor, and Tempus with Pathos AI (2025).
• Display, grounding or answers (D), 32 deals: every Meta AI, Google News AI pilot, Copilot Daily and Publisher Content Marketplace row, plus ten others such as TIME, The Atlantic and both Getty Images licenses.
• Training excluded (X), 1 deal: Le Monde and Perplexity.
• Use not stated (n/s), 26 deals, among them OpenAI and News Corp, OpenAI and Reddit, and the Disney agreement.

The wording behind each code is specific. Google's post of 2024-02-22 gives it Reddit Data API access "to display, train on, and otherwise use" Reddit content, and Tempus' 10-K for fiscal 2025 says the Pathos fees are "for the sole purpose of" training a foundation model. Le Monde's article of 2025-05-14 says its Perplexity agreement "ne prévoit pas l'utilisation de nos contenus pour entraîner un LLM", that is, it does not provide for training a language model on Le Monde content. A row marked D or n/s means that training is not stated, not that the contract forbids it. How an agreement draws this line is explained under field of use in AI data licensing agreements.

Program-wide deals with many publishers

Four announcements cover many publishers at once, and 21 tracker rows record publishers that these programs name. The table lists the four programs as of 2026-11-09 with the publishers each buyer names and the payment wording it uses; the program rows are not counted among the 74 deals.

The four programs differ most in what their buyers say about money. Google states that it pays participating publications for the delivery and use of enhanced content, and Microsoft lets PCM publishers set licensing and usage terms, while its Copilot Daily post of 2024-10-01 mentions no payment. Meta's list describes information and links in Meta AI answers and does not mention training.

Bankruptcy data sales, settlements and acquisitions outside the tracker

One data sale in the tracker comes from a bankruptcy court, Spirit Airlines, and it is a successful bid awaiting approval, not a completed sale. At the auction of 2026-08-14, Google LLC made the successful bid of USD 10,000,000 for de-identified business data of Spirit Airlines, LLC and its affiliated debtors, with Mercor.io Corporation as alternate bidder at USD 7,500,000 (Bankr. S.D.N.Y. 25-11897 (SHL), Dkt. 1463). Passenger profiles are excluded. The consumer privacy ombudsman's supplemental report of 2026-10-05 (Dkt. 1684) recommends approval from a privacy perspective, but as of 2026-11-09 the court has not approved the sale, and the hearing is adjourned to 2026-11-17 (Dkt. 1779).

Eight other transactions involve data or data companies but are not licenses for AI training, so they sit in the table below rather than in the tracker: a class settlement, three 23andMe transactions, three acquisitions and a purchase of brand assets. How estates and administrators sell records under court supervision is covered under data assets in insolvency.

The Bartz v. Anthropic settlement is final: the Northern District of California granted final approval and entered judgment on 2026-07-20, and the release covers only past conduct, before 2025-08-25. Getty Images, the seller in two display licenses, filed a prepackaged chapter 11 case on 2026-10-21 (Bankr. D. Del. 26-11643 (JKS)). Litigation behind these records is followed in the case tracker of AI training data lawsuits.

How the tracker is sourced and which deals are still open

Every row rests on a statement by the buyer or the seller, a securities filing or a court docket, and every row is checked as of November 9, 2026. Securities filings (Form S-1, 10-K, 10-Q, 8-K, 424B4 and RNS announcements) supply most stated amounts, court dockets supply the Spirit bid (not approved as of 2026-11-09) and the settlement record, and party newsrooms supply dates and use wording, following how we source, date and check every figure. Where a party gives only a month, the row carries the month: Yelp's shareholder letter of 2026-02-12 says only "we recently signed an agreement with OpenAI".

A reported deal enters the tracker only when the buyer, the seller or a filing confirms it, and press-only values stay out even for widely covered agreements such as Google and Reddit. A status change, such as a ruling on the Spirit sale, is entered with its docket number and date.

Four deal pages have no tracker row as of 2026-11-09, for the reasons below.

The last three pages gain a row, and a link, only once a party or a filing confirms a deal.

What the recorded deals mean for a data holder considering a license

The tracker shows what large holders have agreed to, not what a given holder will be offered. Its rows come mostly from news publishers, platforms and listed companies, and most terms are not public.

Benefits:

• Display without training: 32 deals license display, grounding or answers without stating training, and Le Monde's Perplexity agreement excludes training.
• Fixed fees on record: the Tempus filings show fixed and per-record fees for clinical data, set out under the stated values above.
• Program routes: Google, Meta and Microsoft run programs that take in many publishers under one set of terms.

Risks:

• Renewal concentration: Reddit's 10-Q for the third quarter of 2026 ties substantially all of its licensing contract value to two agreements nearing the end of their terms.
• Restructuring: Getty Images, the seller in display licenses with Perplexity (2025-10-31) and OpenAI (2026-06-21), filed for chapter 11 on 2026-10-21.
• Privacy enforcement: Italy's Garante issued GEDI a formal warning (provvedimento n. 741 of 2024-11-27, art. 58(2)(a) GDPR) that its archive deal with OpenAI could breach GDPR Articles 9, 10, 13, 14 and Chapter III.

Downsides:

• Few benchmarks: 55 of the 74 deals are marked n/d, with no amount from any party or filing.
• Listed-company bias: the stated amounts come from Reddit, Wiley, Informa, Tempus, Kopinor's collective agreement and a court docket, which leaves private holders few comparable figures.

Methods for pricing records without public comparables are set out in what company data is worth to AI companies.

Alternatives:

• Not licensing at all: no rights pass, and no field-of-use, deletion or indemnity terms arise.
• A display-only license: as in the TIME, Washington Post and Getty Images rows, where training is not stated.
• A collective license: Kopinor licenses Norwegian press publications to the National Library of Norway on behalf of its member organizations.
• A program route: Microsoft's Publisher Content Marketplace lets publishers define licensing and usage terms.

Each route, including stopping before any contract, is weighed in the guide to sell or license company data to AI companies.

Where to read the buyers, prices, cases and sellers behind these deals

The tracker answers who licensed what to whom; the pages below answer who buys, what data sells for and which sellers do it.

• Who buys AI training data: every buyer and intermediary, alphabetically.
• The AI training data market: size estimates and history.
• AI training data prices: per work, hour, image, token and record.
• AI data partnership payouts and program terms: dated program offers.
• Who sells data to AI companies: every type of seller.
• Human data for AI: contributor work, pay and platforms.

Seller-side detail follows for news publishers licensing content to AI companies and for musicians, labels and publishers licensing music. The path from scraped corpora to these agreements is traced in the history of AI training data.

Frequently asked questions

How many AI data licensing deals are public?

This tracker lists 74 deals with a party statement or filing behind each, as of 2026-11-09. Four program-wide announcements and eight related transactions are listed separately and not counted.

Why do most deals in the tracker show no value?

The parties have not disclosed an amount. Only nine of the 74 deals carry an amount stated by a party or a filing, most of them from listed companies' securities filings. Figures that appear only in press reports are left out because no party confirms them.

Did Disney license its characters to OpenAI?

Disney announced a three-year license on 2025-12-11 for more than 200 characters, with costumes, props, vehicles and environments, for Sora and ChatGPT Images. The license excludes talent likenesses and voices, is subject to definitive agreements, and states neither a license fee nor training use. Disney's release of 2025-12-11 separately states that Disney will make a USD 1 billion equity investment in OpenAI and receive warrants; closing is not confirmed.

Which deals let the buyer train on the data?

Fifteen deals state training or model improvement, among them Google and Reddit, Tempus and Pathos AI, and Kopinor with the National Library of Norway. The other rows license display and grounding, exclude training or leave use unstated, which does not mean training is forbidden. Whether training needs a license at all depends on copyright and AI training across jurisdictions.

How do I get deal tracker updates?

New deals, bids and status changes go by email to subscribers of the deal tracker newsletter. The tracker itself is revised when a party, a filing or a court docket adds a deal, an amount or a status change.