Who Buys AI Training Data: Every Buyer and Program, Compared Without Fees

AI training data buyers are the companies that acquire data to train or evaluate AI models, by crawling, licensing, buying or commissioning it. This directory lists 123 buyers and intermediaries as of 2026-11-09, from frontier labs and human-data companies to the brokers and platforms between data holders and labs, and 78 of its rows carry a seller-facing program whose terms were checked on that date.

The rows run alphabetically and state what each company acquires or handles. No row is ranked, filtered or described by referral fees, because the directory covers every buyer in AI training data licensing whether or not it pays anyone for introductions.

Who buys data for AI training?

Data for AI training is acquired by frontier labs and by a chain of intermediaries that collect, license and resell it to them. Human-data companies, data vendors and marketplaces license records in and pass them on, while brokers and shutdown services find buyers for a holder's data without acquiring it.

Five routes move data to a model developer, each with a dated example.

1. Web crawling: OpenAI describes GPTBot as its crawler for content that may be used in training (checked 2026-11-09).
2. Content licensing: Reddit's S-1 reports USD 203.0 million in aggregate contract value for data licensing arrangements entered in January 2024, with counterparties it does not name.
3. Company data programs: micro1 states a commitment of USD 1 billion over 12 months to enterprise data acquisition, announced 2026-10-09.
4. Commissioned collection: Claru ships cameras to partner businesses and pays for point-of-view capture of hands-on work.
5. Insolvency bids: at the auction of Spirit Airlines' de-identified data on 2026-08-14, Google LLC made the successful bid of USD 10,000,000 and Mercor.io Corporation the alternate bid of USD 7,500,000 (Bankr. S.D.N.Y. 25-11897 (SHL), Dkt. 1463).

The Spirit sale is not approved as of 2026-11-09, and the sale hearing is adjourned to 2026-11-17 (Dkt. 1779).

The buyer categories at a glance

The directory sorts buyers into nine categories by how they acquire or handle data, not by what they pay, and a tenth row points to dated buyer incidents. Counts are as of 2026-11-09.

The complete buyer directory

The directory lists every buyer DataPacts profiles, alphabetically, with what it acquires and whether a seller-facing program was on record when its terms were checked. Of its 123 rows as of 2026-11-09, 78 carry an own seller-facing program, 3 a referral program only, 3 an unverified program source, 2 a crawler opt-out only, 1 no company data program, 8 only their own data collection or dataset sales with no seller terms, and 28 no program on record.

• No record: none found on the check date, which does not prove that none exists.
• Open: the source was unreadable or unverified on the check date, so terms are not stated.
• Referral only: the page refers holders to another program, as Sunset does to Replay.

Broker, Shutdown and Platform rows are intermediaries that place or meter a holder's data and do not acquire it. Five rows stay unlinked until the company's identity is confirmed.

What each type of buyer acquires

Each type of buyer acquires different data through a different route, and the ten sections below follow the chain from labs to public institutions.

Frontier AI labs: crawling, licensing deals and bids

Frontier labs acquire most training data by crawling the public web and license the rest from publishers, platforms and archives. Six crawlers and tokens govern that training, as their operators describe them (checked 2026-11-09 unless dated), and site owners can allow, block or charge these AI crawlers.

• GPTBot: OpenAI's training crawler; OAI-SearchBot and ChatGPT-User are separate.
• ClaudeBot: Anthropic's crawler for content that may be used for training; from 2026-10-30 Anthropic recognizes a no-AI-training signal from an individual IETF draft, not a standard.
• Google-Extended: a robots.txt token without its own user agent; no effect on Search.
• Applebot-Extended: a non-crawling token controlling whether Applebot data trains Apple models (Apple support 119829, 2026-09-04).
• Meta-ExternalAgent: Meta's crawler for training foundation AI models.
• Amazonbot: content may train Amazon AI models; noarchive means no training use.

Lab licensing deals rarely carry a value stated by a party: Google and Reddit announced their Data API partnership on 2024-02-22 without one, and Reddit's S-1 gives only an aggregate for counterparties it does not name. xAI's Grok 4.7 summary of 2026-09-21 answers "Yes" on commercial licenses with rightsholders, for text and image only.

Grounding deals are not training deals: Microsoft's Publisher Content Marketplace (2026-02-03), Mistral AI's AFP agreement (2025-01-16) and Meta's news partner deals (2025-12-05) serve AI answers and do not state training as a use. Licenses, grounding agreements and bids are logged among the AI data licensing deals.

Company data programs of human-data companies

Human-data companies such as micro1, Mercor and Handshake AI run programs that license de-identified operating records from companies and pass them to labs. Their terms differ (checked 2026-11-09).

• micro1's Data License Master Terms of 2026-11-08 take a nonexclusive, irrevocable and perpetual license.
• Mercor states that it wires payment after delivery, typically 2 to 4 weeks after signing.
• Endless, an intermediary, places records non-exclusively and names micro1, Mercor and Handshake as the buyers it places data with.

The same human data companies also pay expert contributors for task work.

RL-environment builders

RL-environment builders acquire codebases and business-app histories to build training environments in which agents practice tasks. Six builders list such records, as of 2026-11-09.

• Collinear AI: private repositories, Slack, Gmail, Jira and Salesforce records.
• Ooak Data: Slack, Gmail, Notion, Jira and SharePoint records.
• Patronus AI: codebases, company records and released games.
• pre.dev: production codebases with git history.
• Sureform: email, chat, ERP records, contracts and code.
• The General Data Company: industrial companies' operating records.

The reference on RL environments and who buys them explains how records become tasks.

Data vendors and data labs

Data vendors collect or commission data, often speech, video, robotics demonstrations or medical images, and license finished datasets to developers. David AI builds and licenses speech datasets, Gradient Health connects hospitals and imaging centers with AI developers, and XDOF supplies teleoperation demonstrations (checked 2026-11-09). TELUS Corporation completed the take-private of TELUS Digital on 2025-10-31, and EXL completed its acquisition of iMerit on 2026-07-31. Not every vendor buys data: CloudFactory states that it has no program to buy or license third-party datasets.

Data marketplaces

Data marketplaces list datasets from many holders and match them with developers, and the holder usually stays the licensor. AWS Data Exchange, Databricks Marketplace and Snowflake Marketplace carry third-party data, while Shutterstock licenses content and metadata for AI training through its Data, Distribution, and Services offering (10-Q, third quarter of 2026). Kled, operated by Nitrility, Inc., takes an exclusive, perpetual and irrevocable license to individuals' uploads for AI uses (terms revised 2026-07-21). Cloudflare announced its acquisition of Human Native on 2026-01-15, and Getty Images Holdings filed a prepackaged chapter 11 petition on 2026-10-21 (Bankr. D. Del. 26-11643 (JKS)), with solicitation pending as of that filing.

Brokers and sell-side advisers

Brokers and sell-side advisers do not train models: they find buyers for a holder's data and are paid by the buyer, the seller or both. Regulr states that the buyer pays its fee when a license closes, while Data Licensing Advisors and DataSupply.ai take a success fee on the seller's side (checked 2026-11-09). DataSupply.ai is therefore a seller-side intermediary, not a buyer, and the same holds for all 14 broker rows.

Content licensing platforms and pay per crawl

Content licensing platforms sit between websites and AI crawlers and charge or meter access to published content. Cloudflare has blocked AI crawlers by default for new domains since 2025-07-01, and on 2026-09-15 it announced separate Search, AI training and AI agents controls with recommended defaults to replace "Block AI Bots"; Pay per crawl is a closed beta as of 2026-07-28. TollBit, ProRata (Gist) and Dappier meter or sell publisher access for AI answers and agents.

Shutdown services that sell data for closing companies

Shutdown services wind down startups and sell or license their source code and workspace data to AI buyers on the startup's behalf. SimpleClosure's AssetHub handles code and Slack, Drive, email, Notion and Figma content for a percentage on success, and Sunset refers holders to Replay (checked 2026-11-09). Founders find the steps for data from a shut-down or defunct company in the seller guides.

Enterprises training their own models

Enterprises that train their own models license domain data for their own use rather than to resell it. Pathos AI, Inc., a drug development AI company, licenses a de-identified oncology dataset from Tempus AI, Inc. solely to train a foundation model. Tempus's 10-K for fiscal 2025 states USD 200 million over three years for that license, including USD 50 million upfront.

Governments and public institutions

Public institutions acquire training data through national AI data programs and collective licenses. The National Library of Norway pays Kopinor NOK 45,000,000 for 2026, indexed from 2027, under an agreement signed on 2025-12-19 for the use of copyright-protected material in training Norwegian language models. Neither this category nor the enterprise category holds a profile yet.

How AI data buyers differ: exclusivity, ownership and program terms

Buyers differ less in what they say they pay than in what the license lets them keep: exclusivity, ownership of derivatives, deletion and payment form. The table compares six published terms as of 2026-11-09, with each program's own terms page as the source.

An exclusive grant such as Labelbox Prism's default limits who else can license the same records, while micro1's nonexclusive grant lets the seller license again but leaves models, weights and derivatives with micro1. Payouts and referral terms per program are dated in program terms compared.

How the directory is built: neutral order, check dates, funding and no fees

The directory is ordered alphabetically and includes every buyer DataPacts has found, whether or not it pays referral fees. Each row was reviewed against primary sources on 2026-11-09, and the Checked column carries that date wherever a seller program was found, following how we source, date and check every figure.

"Compared Without Fees" means that no row is ordered, filtered or described by a referral fee and that the page carries no referral links.

Contradictions on a buyer's own pages stay visible. Handshake AI's data partnerships page headlines "Earn $100K to $4M+" while its estimator text reads "up to $100,000", and Troveo's content-owners page states "$50M+" paid out to creators while its business-data page states "$20M+" paid to licensors, all checked 2026-11-09.

DataPacts charges a company nothing for an enquiry or introduction, takes no cut of any sale and has joined no referral program, as set out in how DataPacts is funded.

Benefits, risks, downsides and alternatives of licensing to any buyer, including not selling

Licensing data to an AI buyer can turn dormant records into revenue, but it also transfers rights that usually cannot be recalled once a model is trained.

Benefits:

• Revenue: micro1, Mercor and Handshake AI publish programs for records a company already holds (checked 2026-11-09).
• Covered costs: Mercor states that it covers extraction, anonymization and transfer costs.
• Retained ownership: Labelbox Prism and Turing Project Lazarus state that the seller keeps ownership.

Risks:

• Irreversibility: Labelbox Prism states that trained models cannot be unlearned, and micro1's terms set no duty to retrain.
• Perpetual grants: micro1 and Kled take perpetual, irrevocable licenses.
• Buyer incidents: Mercor reported to the California Attorney General on 2026-06-25 a breach through compromised LiteLLM versions (2026-03-24 and 2026-03-30), and McKinney v. Scale AI (San Francisco Superior Court CGC-24-620481) settled for a gross USD 12,500,000, with final approval filed 2026-10-30. Bartz v. Anthropic (N.D. Cal. 3:24-cv-05417-AMO) ended in a USD 1.5 billion settlement with final judgment on 2026-07-20 (Dkt. 680), tracked with other AI copyright lawsuits over training data.

Downsides:

• Seller liability: micro1's terms make the licensor indemnify IP, privacy and publicity claims and allow a clawback if the data manifest proves materially inaccurate.
• Limited visibility: under the EU AI Act, Regulation (EU) 2024/1689, with Art. 53 applying from 2025-08-02, a provider's training content summary shows licensed data only as Yes or No plus modality.

Alternatives:

• Not selling or licensing at all: no license, indemnity or deletion question arises.
• Blocking or charging crawlers: Google-Extended and Cloudflare's controls limit training on public content.
• Licensing for grounding only: Microsoft's Publisher Content Marketplace pays publishers on delivered value.
• Non-exclusive or time-limited licensing: Turing's sole-seller window ends after 90 days, and Endless deletes records at the end of the agreed term.

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

Where to compare programs, prices, law and the seller path next

The buyer directory answers who acquires data; the pages below answer on what terms, at what price, under which law and in which order of steps.

• Tools for AI data sellers: buyer comparison, program terms, estimates.
• Who sells data to AI companies: every seller type.
• AI training data law by jurisdiction: copyright, privacy, contracts.
• The AI training data market: size, prices, history.
• Choosing a buyer for your data: labs, vendors, marketplaces, brokers.
• License clause matrix: exclusivity, deletion, indemnity.
• RL environment companies: builders and vendors.
• AI training jobs and platforms: pay for contributors, not data holders.

Buyer comparisons

Twenty-seven comparisons set buyers that holders weigh against each other side by side, listed alphabetically as of 2026-11-09.

Frequently asked questions

Do AI companies pay for proprietary datasets?

Yes, and some contract values and bids are on the public record. Reddit's S-1 reports USD 203.0 million in aggregate contract value for data licenses entered in January 2024, and Google LLC bid USD 10,000,000 for Spirit Airlines' data on 2026-08-14, a sale not approved as of 2026-11-09.

How much do AI companies pay for data?

No public rate card exists; published figures are each buyer's own claim. micro1, for example, states tiers of USD 100K to 500K, USD 500K to 3M and USD 3M+ for company data, checked 2026-11-09. Methods and price points are in how much company data is worth to AI companies.

How do I sell data to AI companies?

Audit your rights, remove personal data, choose a buyer type from this directory and negotiate the license before delivery. The seller guide walks through each step, and declining every offer remains a valid outcome.

Are data labeling companies data buyers?

Some data labeling companies are buyers and some are not. Labelbox (Prism), SuperAnnotate and Scale AI run company data programs on record, checked 2026-11-09, while CloudFactory states that it has no program to buy third-party datasets. Vendors in both roles are compared among data labeling companies.

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

Not in the Spirit Airlines case as of 2026-11-09. Google LLC holds the successful bid (Dkt. 1463), and the hearing in the Bankruptcy Court for the Southern District of New York is adjourned to 2026-11-17 (Dkt. 1779). How trustees sell data to AI buyers in insolvency is covered separately.

Which buyers pay referral fees?

Many programs publish referral terms, among them Handshake AI, Mercor, micro1, Scale AI and Turing (checked 2026-11-09), but referral fees never order or filter this directory. The program terms table dates each offer. The referral programs of AI data buyers are compared separately.