The AI Training Data Market: Size, Prices, Payouts and History

The AI training data market is the trade in data acquired to train or evaluate AI models: licenses, sales and bids in which data holders are paid by AI developers or by intermediaries. Reddit's S-1 of 2024-02-22 reports an aggregate contract value of USD 203.0 million for its January 2024 data licensing arrangements, with counterparties it does not name. In the Spirit Airlines bankruptcy, Google LLC placed the successful bid of USD 10,000,000 for de-identified company data on 2026-08-14 (Bankr. S.D.N.Y. 25-11897, Dkt. 1463), and the sale is not approved as of 2026-11-09.

Research publishers disagree on the market's size, and buyers' payout claims run larger than the amounts on the record. This section of AI training data licensing measures sizes, prices, payouts, filings, funding and history.

Every figure names its stating party and check date, as described under how we source, date and check every figure.

What is the AI training data market?

The market has no exchange and no single price: it is visible only through published estimates, disclosed contracts, buyers' own claims, company filings and court dockets.

1. Published estimates: research publishers size an "AI training dataset market" with their own definitions and base years.
2. Contract prices: a few contracts disclose a unit, such as Tempus's USD 100 per delivered record in its 2026-10-22 order form with Pathos AI.
3. Payout claims: buyers state ranges and totals on their own pages, for example micro1's tiers and Mercor's feed totals, both checked 2026-11-09.
4. Filings and dockets: public sellers report license income, and bankruptcy courts record bids for company data.

Not every agreement between an AI developer and a content owner is a training deal. Google states that its Reddit agreement of 2024-02-22 covers the right to "display, train on, and otherwise use" Reddit content, while Le Monde states that its 2025-05-14 agreement with Perplexity excludes training. Several publisher agreements cover display only or do not state training, so a deal counts as a training deal here only where a party states training.

The market pages at a glance

The market section has eight pages, one for each measurable part of the market, listed here as of 2026-11-09 in the order of the page title.

How large is the market? Each publisher's estimate, never averaged

Four research publishers put the AI training dataset market between USD 2.81 billion (Straits Research) and USD 3.59 billion (Fortune Business Insights) for 2025, and a fifth, Global Market Insights, estimates USD 3.2 billion for 2024. Each publisher defines the market in its own way, uses its own base year and dates its report differently, so the figures are separate claims rather than measurements of one quantity.

The table lists each estimate as the publisher states it, alphabetically by publisher, as of 2026-11-09 (each value the publisher's claim, checked that day). It has no average row and no total.

Two estimates contradict themselves. Global Market Insights states a 20.5 % growth rate from USD 3.2 billion in 2024 to USD 16.3 billion in 2034, but those two figures imply about 17.7 % a year (report as of 2025-05). The Business Research Company gives 21.6 % and 21.5 % for the same forecast on one page (report as of 2026-09). The last row measures a different market, data collection and labeling, and is not set against the dataset rows.

• Base year: the year the estimate describes, not the date of the report.
• Report as of: the date on the publisher's page when it was checked.
• Forecast: the publisher's projection, which is not market data.

Price units in signed contracts: per title, per record, per year

Disclosed contracts price training data per title, per delivered record or as a fixed annual fee, and each price belongs to one contract, not to a market rate. The Authors Guild states that the HarperCollins AI license of 2024-11-19 pays USD 5,000 per title, split 50-50 between the publisher and the opted-in author, for a buyer that no party names. Tempus states in its Q3 2026 10-Q that its Neuro Order Form with Pathos AI, signed 2026-10-22, is worth up to USD 55.0 million: USD 50.0 million at USD 100 per delivered record for 500,000 records over five years, plus USD 5.0 million for updates. Kopinor's agreement with the National Library of Norway, signed 2025-12-19, sets NOK 45,000,000 for 2026, index-adjusted from 2027 (s.17).

Fixed totals without a unit are more common: Tempus's 10-K for fiscal 2025 reports USD 200 million in data license fees from Pathos AI over three years. Values that only the press reports are not used. The full table of AI training data prices per work, record and year gathers every public price point by unit; compute costs for training are a different market and are not training data prices.

Payout claims by buyers versus amounts on the record

Most payout figures in this market are buyers' own claims about what they pay or have paid; amounts confirmed by a filing, a docket or the seller are rarer.

Buyer claims, among them (alphabetical):

• Handshake states "Earn $100K to $4M+" in its headline but up to USD 100,000 in the estimator text on the same page (checked 2026-11-09).
• Labelbox states that Prism has distributed USD 100M+ (checked 2026-11-09).
• Mercor's data feed states USD 50,000,000 "paid to companies" in one block and USD 17,000,000 in another (checked 2026-11-09).
• micro1 states payout tiers of USD 100K to 500K, 500K to 3M and 3M+ (checked 2026-11-09), and on 2026-10-09 announced a USD 1 billion data acquisition commitment over 12 months.
• Troveo states "$50M+ paid to creators globally" on its content-owners page and "$20M+ Paid to Licensors" on its business-data page (checked 2026-11-10).

On the record:

• Tempus's 10-K for fiscal 2025 states that Pathos AI paid USD 50 million upfront as of April 2025.
• Reddit's S-1 aggregate contract value, quoted above, is a contract value for unnamed counterparties, not a payment.
• In the Spirit Airlines auction of 2026-08-14, Google LLC's USD 10,000,000 is the successful bid and Mercor.io Corporation's USD 7,500,000 the alternate bid (Dkt. 1463); neither is a payment, and the sale is not approved as of 2026-11-09.
• Keelan Miskell, a founder quoted by MIT Technology Review on 2026-10-13, says he received about USD 25,000 for a codebase.

A buyer's range describes what a program offers, not what a typical seller receives. Each program's figures and conditions are set side by side under payouts and program terms compared, dated.

License income that sellers disclose in their filings

Public sellers report AI licensing income in their filings, and those filings are the most reliable evidence of what changes hands. Shutterstock's Data, Distribution, and Services revenue rose from USD 15,945 thousand in 2021 to USD 38,520 thousand in 2022, USD 137,323 thousand in 2023, USD 175,251 thousand in 2024 and USD 203,264 thousand in 2025, then fell to USD 15,823 thousand in Q3 2026 against USD 65,668 thousand a year earlier (10-K FY2025, 10-Q Q3 2026). The segment also includes Giphy and Studios, so not all of it is AI income, and as of 2026-09-30 Shutterstock held USD 30.5 million of equity received as revenue consideration.

Wiley disclosed a USD 23 million GenAI content rights project on 2024-06-13 and a USD 21 million project on 2024-09-05, with buyers unnamed. For the second project in 2024, Wiley's release gives USD 16 million for Q1 FY2025 and its presentation USD 17 million. Wiley states AI revenue of USD 49 million in fiscal 2026, and AI licensing revenue of USD 14 million in Q1 FY2027 against USD 29 million a year earlier. Informa states non-recurring AI-related revenues of USD 75m+ for Taylor & Francis in 2024.

Reddit's 10-Q for Q3 2026 states that substantially all of its licensing contract value comes from two partners whose agreements are "approaching the end of their respective current terms". Each named agreement is tracked under AI data licensing deals with parties and sources.

Capital raised by data acquirers, as each company states it

Several companies that acquire training data have raised large rounds; every figure below is the company's own statement, listed alphabetically.

• Invisible Technologies states a USD 100 million round on 2025-09-16, total raised USD 144 million, valuation not disclosed.
• Labelbox states a Series E of more than USD 100 million, including equity and a credit facility, at USD 1.25 billion pre-money on 2026-10-14.
• micro1 states a Series B at a USD 4 billion valuation on 2026-10-12.
• Scale AI states a valuation of over USD 29 billion in Meta's minority investment on 2025-06-12.
• Snorkel AI states a USD 350 million Series E at a USD 3.5 billion valuation on 2026-09-22.
• Turing states a USD 111 million Series E at a USD 2.2 billion valuation on 2025-06-10.

A valuation measures the acquirer, not any dataset's price.

Market milestones from Common Crawl to a court-recorded bid

The market moved from crawling public web data to signed licenses in 2023 and 2024, then to standing programs for publishers and companies, license income in filings and court-supervised sales. The table lists the milestones chronologically, as of 2026-11-09, each with the party that states it.

Litigation and licensing run in parallel. The Bartz settlement, final since 2026-07-20, is the largest court-approved sum in the table, while the Spirit auction still awaits its sale hearing on 2026-11-17 (Dkt. 1779). Rulings and settlements are followed in detail under AI copyright lawsuits over training data.

Supply signals: how much public data remains and who restricts it

Supply-side studies suggest that public text is finite and increasingly restricted, which is one reason developers pay for licensed and private data.

1. Epoch AI estimates an effective stock of about 300 trillion tokens of public human-generated text (90 % interval 100 trillion to 1,000 trillion), fully used between 2026 and 2032 (published 2024-06-06).
2. Longpre et al. report that more than 5 % of C4 tokens, and more than 28 % of the most critical sources, became fully restricted by robots.txt between 2023 and 2024, and that terms of service restrict 45 % of C4 (2024-07-20).
3. Cloudflare reports a ClaudeBot crawl-to-refer ratio of 70,900:1 for 19 to 26 June 2025, noting that ratios may be overstated because Claude's app sends no Referer header.

Reddit states that its Q4 2026 logged-in-only tests outside the U.S. are partly meant to protect access to its data corpus. Which bots collect what, and how sites allow, block or charge them, is covered under AI crawlers.

Reading market figures: who states them and how they age

Every market figure on DataPacts names who states it and when, because estimates, claims, filings and dockets differ in reliability and age at different speeds.

• Estimate: a research publisher's model of the market, revised yearly or less often.
• Claim: a buyer's statement about its own program, which can change without notice and is rechecked monthly.
• Filing: a public seller's reported income, audited or reviewed, and updated quarterly.
• Docket: a court record, final only when the court rules, and rechecked after each hearing.

A press figure that no party confirms is not used. Warner Music Group states that "all of our deals are based on variable economics", which is why many music and publisher agreements have no fixed headline value at all. The rules for sources, dates and corrections are set out under how we source, date and check every figure.

Benefits, risks, downsides and alternatives of selling into this market, including not selling

Selling into this market can turn existing records into income, but the evidence also shows volatile revenue, concentrated buyers and approval risk.

Benefits:

• Records already held can earn license income, as the Tempus and Reddit filings show.
• Some contracts pay per delivered record, so income follows delivery.
• Repeat projects happen: Wiley disclosed two GenAI projects in 2024.

Risks:

• License income can fall sharply, as Shutterstock's Q3 2026 and Wiley's Q1 FY2027 figures show.
• Income can depend on few buyers, as Reddit's two-partner concentration shows.
• Objections can delay approval: the Spirit sale hearing, first set for 2026-08-19, is adjourned to 2026-11-17 while the debtors work with objecting parties on employee information (Dkt. 1779).

Downsides:

• Some buyer terms are perpetual: micro1's Data License Master Terms grant a perpetual, irrevocable license, and Labelbox states that trained models cannot be unlearned (checked 2026-11-09).
• Prices are specific to each contract, so a published figure is not an offer.

Alternatives:

• License non-exclusively, so the data can be licensed again.
• License for display only rather than training, as some publishers do.
• Wait until terms for the data type settle.
• Do not sell or license at all, which keeps every option open.

Each step, including the choice not to sell, is set out in the guide on how to sell or license company data to AI companies.

Where deals, buyers, valuation and law continue

The market pages measure the trade; the pages below show who trades, on which terms and under which law.

• Who buys AI training data: every buyer and program, without fees.
• Who sells data to AI companies: every type of seller and their options.
• AI training data law by jurisdiction.
• Human data for AI: the companies, work and pay.
• Tools for AI data sellers, including program comparisons and estimates.

Acquisition routes are explained under how AI companies get training data, and resellers of datasets under dataset vendors and data labs.

Frequently asked questions

How big is the AI training data market?

No single size exists: each publisher states its own estimate. For 2025, Straits Research estimates USD 2.81 billion, The Business Research Company USD 3.19 billion, Grand View Research USD 3.2 billion and Fortune Business Insights USD 3.59 billion; Global Market Insights estimates USD 3.2 billion for 2024 (all checked 2026-11-09). Definitions are compared on the AI training data market size page.

Why do market estimates disagree?

Publishers define the market, choose base years and model growth differently. Some count only datasets, others include collection and labeling. Two reports also contradict themselves: Global Market Insights' growth rate does not match its own figures, and The Business Research Company gives two rates on one page.

How much do AI companies pay for data?

No rate card exists: each contract sets its own price. Disclosed units range from a fee per book title to a fee per delivered record, and buyers' published ranges are claims. Methods and price points are explained under what company data is worth to AI companies.

What did Reddit receive for licensing its data?

Reddit's S-1 of 2024-02-22 reports an aggregate contract value of USD 203.0 million for its January 2024 arrangements, over two to three years. Reddit does not name the counterparties, and it expected at least USD 66.4 million of that value as revenue in 2024. Per-buyer amounts are not on the record.

Has a court approved a sale of company data to an AI developer?

Not in the Spirit Airlines case as of 2026-11-09. Google LLC holds the successful bid from the 2026-08-14 auction, and the sale hearing is set for 2026-11-17 (Dkt. 1779). How trustees run such sales is explained under how trustees sell data to AI buyers in insolvency.

Is the market for training data growing or shrinking?

The evidence points both ways. Shutterstock's data segment grew every year from 2021 to 2025 and dropped in Q3 2026, and Wiley's AI licensing revenue fell year on year in Q1 FY2027. Research publishers all forecast growth, but forecasts are not market data.