Data Types AI Companies Buy: What Sells, Who Buys It and What It Is Worth

AI companies buy, license or collect data of almost every kind, from Slack threads and email archives to books, medical records, robot demonstrations and agent trajectories. This index sorts 43 data types into six families, with the acquirers on record and one dated figure per type where any party has stated one, such as the USD 200 million in license fees that Tempus reports from Pathos AI over three years (10-K for fiscal 2025).

Acquirers are named alphabetically, never by fee. The page covers the data-type side of AI training data licensing: what holders own, who asks for it and what each figure on record measures.

What Data Do AI Companies Buy?

AI companies acquire six families of data: workplace records, code, published text, media and design files, regulated records, and physical-world and agent data. Each family reaches a model developer by its own route.

1. Workplace records: chat, email, tickets, documents, CRM, financial, call, meeting and transaction records.
2. Code: private repositories with their commit, pull request and review history.
3. Published text: books, news, forums, scholarly papers, course content and translation memories.
4. Media and design: video, photos, speech, music, voices, motion capture, design and CAD files, gameplay.
5. Regulated records: clinical, imaging, genomic, legal, financial market and HR data.
6. Physical-world and agent data: sensor, robotics, first-person, driving and aerial footage, agent trajectories and model outputs.

Publishers and platforms sign negotiated deals, company data programs license operating records and code, and vendors collect physical-world footage to order. This page uses "types" in the seller's sense; the machine-learning sense (labeled versus unlabeled, structured versus unstructured data) is not covered here.

Every Data Type at a Glance

The index lists 43 data types, grouped by family, with what acquirers ask for on their own pages and one dated figure where any exists. The table is as of 2026-11-10; acquirer pages behind it were checked on 2026-11-09.

• No public figure: none was stated by the check date, which does not show that no money changed hands.
• Each figure is the stating party's own, never a DataPacts estimate.
• "Sells own datasets" marks vendors that run no seller program.

What Sells in Each Data Family

Each family sells on different record traits, reaches developers through different acquirers and meets its own limit.

Workplace Records: Chat, Email, Tickets, Documents and Business Systems

Workplace records sell when they show how work was done: threads, tickets, documents and system histories that link a request to its outcome. Turing states USD 50 per ticket, spec or support thread when bundled with a codebase (newsletter of 2026-03-25), and Specific states USD 2 per accepted page of documents, both checked 2026-11-09. The Spirit Airlines asset schedule (Bankr. S.D.N.Y. 25-11897 (SHL), Dkt. 1463, Exh. A-2) lists as Included 100,000,000 emails, 500,000,000 Teams items, 667,563 IT tickets, 175,658 employee records, 3,426,618 payroll records and 516 repositories with 372,585 commits, while 97,500,000 passenger profiles are Not Included. At the 2026-08-14 auction, Google LLC made the successful bid of USD 10,000,000 and Mercor.io Corporation the alternate bid of USD 7,500,000; the sale is not approved as of 2026-11-09.

Programs that ask for workplace records, among them:

• Centific
• Mercor
• micro1
• Nyne
• Scale AI
• SuperAnnotate
• Turing (Project Lazarus)

Platform terms decide whether records can leave. Slack's API Terms (effective 2025-10-10) bar apps offered outside the organization from training a large language model on API Data or bulk exporting messages without an additional agreement, and WhatsApp Business Platform terms (last modified 2026-09-23) bar selling Platform Data, even in aggregate form. Holders license these records mostly through company data partnership programs.

Code Repositories: What a Codebase Acquirer Checks First

Codebase acquirers look for private repositories with long commit histories written by people, and price them per token or per pull request. Pangea states USD 100 per million tokens of source code and USD 40 per million tokens of git history, checked 2026-11-09. Turing states USD 1,000 per pull request for modern codebases, up to USD 10,000 per codebase, and USD 10,000 to 100,000 or more for legacy codebases (newsletter of 2026-03-25).

Mercor's software-company listing, checked 2026-11-09, asks for:

• private repositories that were never public, mirrored or open-sourced;
• three or more years of commit history from at least 10 engineers;
• 50,000 or more lines of code the company owns outright;
• code written by people, with tests that run.

Mercor rules out forks, side projects, code mostly written by AI tools, single-author repositories and client work an agency does not own. Ownership is therefore the first test. A rights audit of the code a company wants to license settles it before any listing.

Published Text: Books, News, Forums, Scholarly and Course Content

Published text is licensed mostly by publishers and platforms in negotiated deals, and many of those deals cover display in answers rather than training. Training is stated for Axel Springer, Le Monde with OpenAI, the Financial Times and Google with Reddit, among others, while the TIME, Guardian and Condé Nast deals state display only or no training term. Reddit's S-1 (filed 2024-02-22) reports USD 203.0 million in aggregate contract value for licenses entered in January 2024, with counterparties not named. Wiley's 10-K reports AI licensing revenue of USD 33.1 million in Research and USD 16.0 million in Learning for fiscal 2026.

Acquirers of published text on record, among them:

• Google: Reddit and Stack Overflow, values not disclosed.
• Meta, Microsoft and Mistral AI: paid API users of Wikimedia Enterprise (2026-01-15).
• OpenAI: Stack Overflow (2024-05-06) and news publishers.

The Authors Guild states that HarperCollins offered USD 5,000 per title, split 50-50 with the author (2024-11-19), for a buyer no party has named. Each text deal sits in the tracker of AI data licensing deals.

Media and Design: Video, Images, Speech, Music, Voice and Files

Media is licensed as archives with rights cleared by the owner, and priced per hour, per clip or as a revenue share. Claru states a typical first deal of USD 5,000 to 50,000 for a non-exclusive archive license, and SpeechData.ai sells its own speech datasets at USD 60 to 95 per hour of audio, both checked 2026-11-09.

Media acquirers on record, among them:

• Defined.ai: audio, image, video and text.
• ElevenLabs: shared Professional Voice Clones.
• Runway: a custom model on Lionsgate's catalog (2024-09-18).
• Sieve: film, TV, podcast and gameplay archives.
• Stability AI, Suno and Udio: music under label licenses.

Figma trains on Starter and Professional files by default since 2024-08-15, without payment. Disputes end in settlements: Bartz v. Anthropic (N.D. Cal. 3:24-cv-05417-AMO) received final approval of a USD 1.5 billion settlement on 2026-07-20 (Dkt. 680). It is tracked among AI copyright lawsuits over training data.

Regulated Records: Clinical, Genomic, Legal, Financial and HR Data

Regulated records carry the largest disclosed AI license values, and the rules that decide whether they may be licensed at all. Tempus reports USD 200 million in data license fees from Pathos AI over three years, including USD 50 million upfront, "for the sole purpose of" training a foundation model (10-K for fiscal 2025). A further order of 2026-10-22 runs to up to USD 55.0 million, mostly at USD 100 per delivered record (10-Q for the third quarter of 2026). GSK's 2023 license from 23andMe served drug target discovery, not AI training.

Programs for regulated records, among them:

• Gradient Health: imaging, EHR and pathology data.
• HistAI: pathology whole-slide images.
• Protege: healthcare data from providers.
• Shfa.ai: EHR exports and clinical notes.

HIPAA requires authorization for a sale of protected health information (45 CFR 164.502(a)(5)(ii)), while de-identified data falls outside the Privacy Rule. For legal data, the Third Circuit (No. 25-2153) affirmed on 2026-09-29 that ROSS's use of Westlaw headnotes was not fair use, calling the market for licensing headnotes to train AI "rapidly developing"; a rehearing petition is pending as of 2026-11-09. Further limits sit under company data that cannot be licensed.

Physical-World and Agent Data: Robots, First-Person Video, Driving and Trajectories

Physical-world and agent data is mostly collected to order by vendors and programs, because few holders already own it at scale. Mindrift states its egocentric projects pay USD 6 to 15 per hour, and Bee Maps lists USD 6 per AI event video, both checked 2026-11-09. For agent data, micro1 states that it licenses qualifying trajectories for USD 1 to 2 million annually, and Mercor states USD 20,000 to 480,000 to an AI-product company for session logs, both checked 2026-11-09.

Collectors and programs on record, among them:

• Asimov: first-person video of physical tasks.
• Build AI: collects its own; Egocentric-10K holds 10,000 hours from 85 factories.
• Figure AI: collects video itself (Project Go-Big, 2025-09-18).
• Magma: agent traces on a revenue share.
• Midcentury and XDOF: collect their own gripper and teleoperation data.
• Toloka: egocentric video and teleoperation from partners.

The EU Data Act (Regulation (EU) 2023/2854) governs data from connected products and applies from 2025-09-12. Agent traces also come from simulated tasks, covered under RL environments and who buys them.

Which Acquirers Want Which Data Family

Each family has its own set of acquirers: labs sign deals for published text and media, programs license workplace records and code, and vendors collect physical-world data. The table is as of 2026-11-10, with names alphabetical in each cell.

Labs appear only where a party has published the deal, so a blank cell means no deal on record, not no interest. Marketplaces and licensing platforms handle or meter a holder's data and do not acquire it. Every buyer and program is profiled under who buys AI training data.

What a Data Type Is Worth: Figures by Stating Party, Unit and Date

No data type has a market price; the figures on record are a handful of contracts, list prices and acquirers' claims, each stated by one party on one date. Contracts come from filings, such as Tempus's clinical licenses and Reddit's aggregate. Court bids come from dockets, and the Spirit bids remain bids, not approved as of 2026-11-09. Program claims come from the acquirer's own page and are written as "Pangea states" or "micro1 states" with the check date.

Units do not convert. A price per million tokens, per pull request, per accepted page, per hour of footage and per delivered record measure different things, so the index never averages them or derives a typical price. Unit prices are set side by side under AI training data prices.

Settlements are not prices either: the Bartz settlement compensates claims for past use, and dividing it by works gives no license rate.

Licensing a Data Type: Benefits, Risks, Downsides and Alternatives, Including Not Licensing

Licensing a data type turns records already held into income, but a license usually cannot be recalled once a model is trained on them.

Benefits

• Income from records often kept only for compliance.
• Several routes per family: a deal, a program, a marketplace or a court-supervised sale.
• Non-exclusive terms in some programs, such as Claru's, leave ownership with the holder.

Risks

• Platform terms: Slack's and WhatsApp's terms can bar export or training.
• Recording consent: California requires all parties' consent to record confidential communications (Cal. Penal Code 632(a)).
• Health records: a sale of protected health information needs authorization under HIPAA.

Downsides

• Training cannot be undone after the license ends.
• Staff, customers and authors may object to their messages or works being licensed.

Alternatives

• Licensing a narrower slice, such as de-identified tickets.
• Training a model for in-house use, as Lionsgate arranged with Runway.
• Joining a collective scheme, as Norwegian rightsholders do through Kopinor.
• Not licensing at all, which keeps every option open; whether to sell or license data at all is weighed on its own page.

Regulated records carry the most rules, workplace records hold the most people, and published text and media carry the most disputes.

How This Index Is Built: Stating Parties, Check Dates and No Fees

Every row is built from a docket, a filing or the acquirer's own page, and every figure names the party that stated it and the date it was checked. Acquirers are listed alphabetically, and vendors that collect or sell their own datasets, such as Build AI, David AI, HistAI, Midcentury, SpeechData.ai and XDOF, are marked as such.

Acquirer pages were checked on 2026-11-09, while filings and dockets carry their own dates. DataPacts charges a company nothing for an enquiry, takes no cut of any sale and has joined no referral program, as set out under how we source, date and check every figure.

Where Data Types Meet Systems, Industries, Sellers and the Seller Guide

The data type decides who wants the records; the system decides how they leave, the industry decides which rules bind them, and the seller guide decides the order of steps.

• How to sell or license company data: the complete seller guide.
• Who sells data to AI companies: every type of seller.
• Data licensing for AI by industry: sector limits.
• How AI companies get training data: every acquisition route.
• AI training data law by jurisdiction: the law for each data type.
• Export rights by source system: Slack, Microsoft 365, Jira and GitHub.
• How DataPacts is funded: funding and limits.

Frequently Asked Questions

What data do AI labs buy?

Frontier labs acquire crawled web content plus licensed publisher, platform, music and image content. Company records reach them mostly through programs and vendors. Each lab's routes are described under frontier AI labs as data buyers.

What kind of data is used to train AI models?

Acquired training data ranges from licensed posts and books to clinical records, footage and agent trajectories. Reddit's January 2024 licenses and the Tempus license with Pathos AI mark two ends. Machine-learning courses use "kinds" for labeled or unlabeled data instead.

How much do AI companies pay for data?

There is no rate card: each figure on record is a contract, a list price or an acquirer's claim with its own unit and date. Figures do not transfer between data types. How such figures feed a valuation is explained under what company data is worth to AI companies.

Do AI companies buy synthetic or model-generated data?

Yes, synthetic data is acquired, but reusing a model provider's outputs as training data depends on that provider's terms of use. Those terms are not compared here. Synthetic data's uses and limits are covered under synthetic data for AI training.

Can a company license chat and email written by its staff?

Only after notice and consent steps, and only where the platform's terms allow the export. Slack's API Terms require an additional agreement for bulk export by outside apps. The steps for staff are set out under employee notice and consent.