Frontier AI Labs as Data Buyers: How OpenAI, Google, Anthropic and Others Acquire Data

AI labs buy training data when they pay a holder for it, through licenses with platforms and publishers, catalogs from labels and image libraries, records placed by intermediaries, bids in insolvency sales and purchased books, while much of their data still comes from crawling, their own users and synthetic generation. Of the 20 developers profiled here, only xAI has an EU training content summary located.

Every figure is the stating party's claim or a filing, checked 2026-11-09. The order is alphabetical, listing never depends on referral fees, and DataPacts takes no cut of any sale. The page belongs to the independent reference on AI training data licensing.

How each figure was sourced and dated is set out in how we source, date and check every figure.

What Makes an AI Lab a Data Buyer

A lab acts as a data buyer when it pays a holder for access to or a license over data, which separates a Reddit Data API deal from a GPTBot crawl of the same posts.

Crawling, user data and synthetic generation involve no payment to the holder. OpenAI told Canada's privacy regulator that less than 1 % of its pre-training datasets came from licensed third-party datasets, which the regulators could not verify.

Platforms, publishers, labels and human-data intermediaries sit between smaller holders and labs, so the lab is often the end recipient. In the chain of who buys AI training data, labs are the last link.

The 20 AI Developers Profiled as Data Buyers

The directory lists all 20 AI developers profiled under this category, alphabetically, with what each is on record as acquiring and whether it offers holders any door. The table is as of 2026-11-09.

• "None on record" means no seller-facing program was found, not proof that none exists.
• "Page found, terms not verified" means nothing about the terms is stated.
• "Not located" means no EU training content summary was found, not that none exists.

What Labs Disclose About Their Own Training Data

Few labs say where their training data comes from, and those that do say it in filings, regulator findings or summaries the EU AI Act requires, not on a seller page. The table lists six disclosures, each value the discloser's own statement.

The regulators could not verify OpenAI's figure, and the only EU training content summary located is xAI's, which states Yes or No plus modality and never lists datasets. The EU AI Act Article 53 training data summary explains why summaries stay short.

Five Ways Labs Pay for Data

Labs pay for data in five ways on the record: platform and publisher licenses, label and image catalogs, company records placed by intermediaries, bids in insolvency sales, and books bought outright.

Licenses With Platforms and Publishers

Platform and publisher licenses are the best documented route, yet a value stated by a party is rare: Reddit's S-1 gives one aggregate for counterparties it does not name. The filing of 2024-02-22 states USD 203.0 million for licenses of January 2024, and Reddit's 10-Q for the third quarter of 2026 says substantially all of that value comes from two partners whose agreements are "approaching the end of their respective current terms". Google announced Reddit Data API access "to display, train on, and otherwise use" Reddit content on 2024-02-22, and OpenAI's access followed on 2024-05-16, both without a value.

Other deals carry dates and no price. OpenAI licensed part of the Associated Press text archive on 2023-07-13, and Shutterstock announced an expanded OpenAI partnership and its Contributor Fund on 2022-10-25. Wikimedia Enterprise named Amazon, Meta, Microsoft, Mistral AI and Perplexity as new paid API customers on 2026-01-15, with no pricing disclosed. The tracker of AI data licensing deals holds each deal with its parties and sources.

Training, Grounding or Display: What a Lab License Covers

A lab license does not always cover training: several of the best known deals license content for display or for grounding answers, and the parties' own texts say which. The table classifies deals as of 2026-11-09 from the parties' releases, with examples in alphabetical order.

For a holder the difference decides what a license permits. Microsoft states that publishers are paid "on delivered value" for grounding, not training, while Le Monde says its Perplexity agreement does not provide for training an LLM.

Music Labels and Image Libraries as Lab Suppliers

Music and image developers buy from catalog owners rather than from individual creators, and the labels describe their deals as variable economics rather than fixed prices. Warner Music Group names AI agreements with Suno, Udio, Stability AI and KLAY in an exhibit dated 2026-03-03. Udio and Universal Music Group settled and licensed music on 2025-10-29, and ElevenLabs signed a multi-year Universal agreement on 2026-09-10, with no values disclosed.

Image developers work from libraries. Bria AI claims it is trained only on licensed data from partners including Getty Images, Alamy and Envato, and states that compensation "ties to influence, not catalog size", without a rate. Stability AI claims Stable Audio 3.0 (2026-05-20) is trained on fully licensed data.

Company Records Through Intermediaries and Insolvency Bids

Operating companies' records usually reach labs through human-data companies that license the records first and pass them on, so the lab is the end recipient rather than the counterparty. The company data partnership programs show the pattern.

micro1 states that it licenses company data to AI labs and enterprises in the US and allied countries, and Mercor states that no AI lab buyer ever sees raw data and names no lab, both checked 2026-11-09.

In the Spirit Airlines chapter 11 (Bankr. S.D.N.Y. No. 25-11897 (SHL)), the notice of auction results of 2026-08-14 (Dkt. 1463) names Google LLC the successful bidder at USD 10,000,000 (sale not approved as of 2026-11-09) and Mercor.io Corporation the alternate bidder at USD 7,500,000, for a schedule that includes 100,000,000 emails and 500,000,000 Teams items. micro1 filed a notice of intent to bid (Dkt. 1556, 2026-09-03) without an amount.

The approval hearing was set for 2026-11-17 (Dkt. 1779), and the Consumer Privacy Ombudsman's report (Dkt. 1684, 2026-10-05) asked for further research into Mercor's security incident and whether it is a qualified buyer. Under 11 U.S.C. § 363(b)(1), personal data covered by a privacy policy is sold only if consistent with it or after an ombudsman, notice, hearing and court approval, so administrators can sell data to AI buyers in insolvency only on those terms.

Books Bought and Scanned: What the Bartz Ruling Drew

In Bartz v. Anthropic the court treated buying print books and scanning them for training as fair use, and refused that protection to the pirated copies Anthropic also kept. The Order on Fair Use of 2025-06-23 (N.D. Cal. No. 3:24-cv-05417-AMO, Dkt. 231) held training use exceedingly transformative and digitizing purchased print books fair use.

The order did not excuse more than seven million pirated library copies. The case settled for a USD 1.5 billion non-reversionary fund, final approval followed on 2026-07-20 (Dkt. 680), and the first distribution is expected on or before 2026-11-15 (settlement website, checked 2026-11-09).

For holders of books, a purchase by a lab is a sale of copies, not a license of the text. Other rulings are collected in the tracker of AI copyright lawsuits over training data.

Where Labs Take Data Without Paying the Holder

Much of what labs train on is not bought: Anthropic's S-1 lists public web data and synthetic data next to licensed datasets, and xAI's Grok 4.7 summary lists crawled web data, user data and synthetic data. Three routes involve no payment to a holder:

• Crawling: OpenAI states that GPTBot crawls content that may train its models, and Anthropic that ClaudeBot collects content that may be used for training; Anthropic recognizes a no-AI-training signal from 2026-10-30 based on an individual IETF draft, not a standard. Google-Extended is a robots.txt token with no effect on Search, and Google aims to support a shared AI-training token in 2027.
• Own users: Meta trains on public posts of EU users from 2025-05-27, subject to an objection form (Irish Data Protection Commission statement, 2025-05-21).
• Synthetic data: Anthropic's S-1 lists synthetic data generated by its own models.

Holders can read the controls in the guide to AI crawlers and how to block or charge them.

Can a Company License Data Directly to a Lab?

Few labs publish a door for outside holders, and the doors on record are for publishers, creators or catalogs rather than for an operating company's internal records. Six doors are on record, in alphabetical order:

• Bria AI: a "Become a data partner" contact link, with no published process and no rate.
• ElevenLabs Voice Library: voice owners "earn rewards when others use them"; ElevenLabs claimed on 2024-10-28 to have crossed USD 1 million in payouts.
• Google News AI pilot: by invitation, for news publications; Google states that it pays participants, and training is not stated.
• Microsoft Publisher Content Marketplace: an interest form for publishers, for grounding, with no rates.
• OpenAI Data Partnerships page: its terms could not be verified, so none are stated.
• YouTube third-party training setting: off by default, the creator picks the companies, and YouTube states it is not facilitating payments.

Figure AI collects home video itself and runs no open seller program, and Adobe's and Perplexity's program terms could not be verified on 2026-11-09. A company holding internal records is usually routed through an intermediary; the guide to choosing a buyer for your data compares the routes.

Benefits, Risks, Downsides and Alternatives of Licensing to a Lab, Including Not Licensing

Licensing to a lab can bring a large counterparty and a documented deal, but values are rarely disclosed, doors are few and a license may cover display rather than training.

Benefits

• A lab is a named counterparty: Google's Reddit deal of 2024-02-22 states the parties, date and access.
• Google says it pays participating news publications in its News AI pilot.
• Label deals describe variable economics, so payment can follow use.

Risks

• Concentration: Reddit's licensing value rests on two partners whose agreements are nearing their end (10-Q, third quarter of 2026).
• Purpose: a license may cover display or grounding, not training.
• Litigation: OpenAI MDL No. 3143 (S.D.N.Y. 1:25-md-03143-SHS-OTW) has cross-motions for summary judgment briefed and no ruling located as of 2026-11-09.
• Review: a court-supervised sale can be questioned, as the ombudsman's requests in the Spirit case show.

Downsides

• No lab license value is stated by both parties, so no confirmed price exists to compare.
• OpenAI's Data Partnerships terms could not be verified.
• Through an intermediary the lab is not the counterparty.

Alternatives

• A company data program that licenses records and passes them on.
• A vendor or marketplace that lists and licenses the data.
• Blocking or charging crawlers instead of licensing.
• Not licensing at all, which is a valid outcome.

The guide on whether to sell or license your data works through which option suits a given holder.

Where Labs Sit Among Programs, Vendors, Deals and the Seller Path

Labs are the end of most data chains; the pages below cover the other buyers, the full route map and the steps a holder takes before any license. DataPacts explains how DataPacts is funded.

• Routes: how AI companies get training data, human data companies and RL environments and who buys them.
• Other buyers: AI training data marketplaces, dataset vendors and data labs, enterprises buying data and content licensing platforms for publishers.
• Sellers and data: news publishers licensing content and data types AI companies buy.
• Terms: an AI data licensing agreement and copyright and AI training.
• Before licensing: alternatives to selling data and reported payouts to data sellers.

Questions About AI Labs Buying Data

Do AI companies pay for training data?

Some of it: labs pay platforms, publishers, labels and intermediaries, but much training data is crawled, generated or taken from users. OpenAI told Canada's regulator that licensed datasets were under 1 % of its pre-training data.

How much do AI labs pay for data?

Labs and sellers rarely state a value. Reddit's S-1 of 2024-02-22 gives USD 203.0 million in aggregate for January 2024 licenses with unnamed counterparties. The valuation guide covers what company data is worth to AI companies.

Does OpenAI buy data from companies?

OpenAI licenses content from publishers and platforms, and runs a Data Partnerships page whose current terms could not be verified on the check date.

Did Google buy Spirit Airlines' data?

No sale had been approved as of November 9, 2026. Google LLC was the successful bidder at USD 10,000,000, and the approval hearing was set for November 17, 2026.

The sale is covered in the page on Spirit Airlines' data archive sale.