• For buyers
  • Docs
Sign inSign up
Blog
Nov 14, 2026Guide

Which AI Labs Buy Company Data? Buyer Programs and How to Apply

Which AI labs buy company data in 2026: which labs run a public intake program, which deals are only reported, what labs require, and how to reach a buyer.

By DataVendor team

Most large AI labs pay for outside data, but few of them publish a way for a company to offer it. If you own a private codebase, a support archive, or years of internal records, you need to know which labs buy, how they buy, and how a seller your size can reach them.

At DataVendor, we connect sellers with the labs that buy this kind of data. This guide looks at what the leading labs have said and done in public. DataVendor wrote it, and the details come from public sources. It separates two kinds of evidence: programs where a lab invites submissions, and deals that were announced or reported without an open path for other sellers. The details are current as of November 2026. Programs open and close, so check each link before you apply.

See which labs buy outside data

The table below summarizes the public record for each major lab. A public intake channel means the lab itself publishes a form or program that outside companies can use to offer data.

LabPublic intake channelWhat is on the public record
OpenAIThe Data Partnerships formLicensing deals with AP, News Corp, Reddit, Stack Overflow, and other publishers
MicrosoftA marketplace pilot for publishers onlyA book-licensing deal with HarperCollins
GoogleAn interest form for its Content Offer Pilot, by invitationDeals with Reddit and AP, and the winning bid for Spirit Airlines' internal data
MetaNoneMulti-year news deals with CNN, Fox News, USA Today, News Corp, and others
AmazonNoneNews deals with The New York Times, Condé Nast, and Hearst
AppleNoneReported talks with publishers in 2023, with no announced deal
AnthropicNonePurchases of print books, and a settlement over pirated books
xAINoneReported talks to buy data from struggling startups
MistralNoneA deal to use the AFP archive in Le Chat answers

Most of these deals involve publishers and platforms that hold millions of documents. A smaller company with one codebase or one company's worth of records rarely gets a direct line to a lab. For those sellers, the practical route is a marketplace such as DataVendor, covered below.

Know which labs accept submissions

Only a few labs invite outside companies to offer data. Even these programs are narrow, so read the criteria before you prepare anything.

OpenAI Data Partnerships

OpenAI launched Data Partnerships in November 2023, with an application form on the same page. It described two tracks. One would produce open-source datasets that anyone could use. The other would cover private datasets for training OpenAI's own models.

OpenAI asked for data that "expresses human intention", such as long-form writing or conversations, rather than disconnected snippets. It accepts text, images, audio, and video in any language. It does not want datasets with sensitive or personal information, or information that belongs to a third party. The page does not state a minimum size or payment terms, and TechCrunch noted at launch that the program said little about compensation.

The page dates from 2023, so confirm that the form still accepts applications before you prepare a submission. Also check your data against the exclusions. A Slack archive full of employee and customer details would not qualify until you remove that information.

Microsoft Publisher Content Marketplace

Microsoft launched its Publisher Content Marketplace as a pilot in February 2026. AP, Condé Nast, Hearst, and Vox were among its early partners. Publishers register interest through a form linked from Microsoft's announcement, as Search Engine Land reported.

The marketplace is built for published content. It does not accept codebases or internal company records.

Google Content Offer Pilot

Google runs a Content Offer Pilot that pays owners of large, high-quality datasets in specialized segments, such as code repositories, for the delivery of offline content. Owners submit an interest form, and Google invites partners as space in the pilot becomes available. Google and each partner agree on a fee. The pilot is non-exclusive, and partners keep their IP and the right to sell their content elsewhere.

Google also describes a separate partnership program that pays for non-public content in science, mathematics, and learning. That page has no form, so it is not an open channel for new sellers.

Understand the deals labs have made

Most lab purchases happen through direct negotiation. These deals show what labs value, even when they offer no path for other sellers.

Google

Google has announced licensing deals with Reddit and AP. Its most relevant purchase for company-data sellers is Spirit Airlines' internal archive. In August 2026, Google won the airline's bankruptcy auction with a $10 million bid for roughly 100 million emails, 500 million Teams messages, and 30 million lines of code, as reported by Axios and TIME. Google said it would not receive any personal information from the dataset.

The sale still needed court approval in early November 2026. Lawmakers objected to it in October, and the hearing moved to November 6. Its outcome will show how far a lab can go in buying an operating company's internal records.

Meta

Meta announced multi-year news deals in December 2025 with CNN, Fox News, People Inc., USA Today Co., Le Monde, and others, according to CNBC. Press Gazette later reported a 2026 deal with News Corp worth up to $50 million a year. Meta has no public intake page.

Amazon

Amazon licensed New York Times content in May 2025, including for training its own foundation models, as TechCrunch reported. It also has deals with Condé Nast and Hearst. Like Meta, it has no public intake page.

Anthropic

Anthropic has no public program for buying data. Court filings show that it bought and scanned print books for training, and it agreed to a $1.5 billion settlement over pirated books. Its Development Partner Program shares Claude Code usage data with Anthropic, which is a different arrangement from selling data.

Apple, Mistral, and xAI

Apple reportedly offered publishers multi-year deals worth at least $50 million in 2023, but we found no announcement of a signed deal. Mistral signed a deal with AFP in January 2025 that lets its Le Chat assistant answer questions from the AFP archive, according to TechCrunch. Reports differ on whether Mistral also trains on that content.

In September 2026, xAI reportedly held informal talks to buy customer and operational data from struggling startups. No completed deal has been reported.

We found no public data purchases by Cohere or Nvidia. Nvidia's agreement to acquire Hugging Face is a company acquisition rather than a data license.

Reach labs as a smaller seller

Most startup codebases and internal records reach labs through a company that sits between the seller and the lab. These intermediaries differ in who sets the price, how much preparation they handle, and whether you can sell the same asset again.

Brokers sell your asset to labs on your behalf. DataVendor helps sellers prepare codebases, raw data, and tasks, then brokers their sale to AI labs. You set the price and keep ownership by default, and a strong asset can license to more than one lab, so a high-quality asset can earn more here than from a single sale. An asset that labs do not want will not sell. Sellers include several publicly traded companies.

Managed vendors set a price for your data and sell it to labs themselves. Mercor, Scale AI, Appen, and Handshake extract and de-identify workplace data, then license it to labs. The vendor sets the price and does the work of the sale. For a strong asset, that price is usually lower than it could earn through a broker. Payment terms vary: Mercor pays once, on delivery to a buyer, while Scale AI describes recurring revenue. micro1 runs a similar program for companies with 30 or more employees, and its public master terms describe a non-exclusive license. Turing's Project Lazarus licenses chat, tickets, wikis, and code from operating and closed companies, and shares the revenue with the seller.

Shutdown brokers form a third group. SimpleClosure's AssetHub sells code and workspace data from closing startups. In October 2026, MIT Technology Review reported more than 150 AssetHub deals, with foundation labs among the buyers.

Our comparison of the best places to sell data to AI companies looks at these routes in more detail.

Check what labs require

The published criteria and the reported deals point to the same requirements. A lab wants data it cannot collect from the open web, with rights it can rely on and no personal information.

Start with rights. OpenAI excludes data that belongs to a third party, and every buyer will ask whether your contracts with customers, employees, and contractors allow you to license the data. Then look at privacy. OpenAI excludes sensitive and personal data, and Google said it would not receive personal information from the Spirit archive. Plan to remove or replace that information before delivery. DataVendor's team removes personal information and committed secrets from codebases and anonymizes email, Slack, Google Drive, and document exports, with the scope agreed for each asset. You do not need to finish that cleanup yourself before you ask for an assessment.

Depth matters too. Long-form writing, complete conversations, and linked engineering records give a lab more to work with than fragments. For example, a support archive in which each ticket links to the customer's question, the agent's reply, and the final resolution is more useful than a list of separate messages.

For codebases, DataVendor deals show what moves the price. Labs pay the most for code written mostly by people, for complex products, and for five or more years of git history, issue tickets, and documentation. Systems languages such as Rust, Go, C, and C++ sell for more.

DataVendor's strict confidentiality policies prohibit us from naming buyers or sellers, or from sharing the figures from specific deals, so the numbers here are directional, current as of November 2026, and not guarantees. About $30,000 per license is the floor for a good codebase, and some deals for a single codebase have exceeded $10 million. A codebase built by several dozen engineers, with thousands of pull requests and over a million lines of mostly human-written code, recently licensed in the low six figures. Poor-quality codebases usually do not sell, and a codebase with weaker engineering can sell for about a third as much as a similar one built by top engineers. Our guide to selling source code covers the full criteria.

Start with DataVendor

You do not need to find an intake form for each lab or wait for a lab to contact you. DataVendor has relationships with all of the leading AI labs and brokers sales to them. Its team helps you assess, prepare, and list your data, and because its platform fee is a share of your earnings, its return rises with your price. You list once, and labs license from your listing. To see current demand, open the Wanted tab after you accept the vendor agreements. It lists open requests from buyers, and you can submit a listing that meets one.

Start by finding out whether your data is something labs will pay for. Get an estimate for your codebase, or book a call for a quote on other data. If the data is high quality, DataVendor can license it to several labs, so it can earn more than a single sale. If it is not, it may not sell, and a managed vendor that sets the price and handles the sale may suit you better.

If the estimate shows a strong asset, the next step depends on what you have:

Your assetHow to start on DataVendor
A private codebase with its engineering historyCreate a listing
A legacy codebase, such as COBOL, Java, or mainframe codeBook a call so the team can help assess and prepare it
A codebase that is not ready to uploadRegister it in Inventory for an initial assessment, then attach the code before you list it
Documents, logs, email, or other company recordsPackage them as a raw data ZIP and list them
Records from a company that is closingExport them before accounts close, then list them
Tasks, evals, or RL environmentsImport the tasks through HUD and submit a private listing to a matching buyer request

If your records contain personal information, ask DataVendor's team to agree the PII removal and secrets cleanup scope before you list. Leave out anything you do not have the right to share. Write a short description of the asset, the dates it covers, its size, and the rights you hold. If a buyer asks for exclusive terms, read our AI data licensing guide before you agree.

Some owners can also apply to a lab directly. Owners of large code datasets can express interest in Google's Content Offer Pilot, which is non-exclusive and does not stop you from listing on DataVendor. Publishers with a large body of news, books, or other published writing can apply to Microsoft's Publisher Content Marketplace, or to OpenAI Data Partnerships if its form is still open.

FAQ

Which AI labs buy company data?

OpenAI, Google, Meta, Amazon, Microsoft, and Anthropic have all bought or licensed outside data. Most of their deals are with large publishers and platforms. Google's winning bid for Spirit Airlines' internal emails, messages, and code is the clearest public case of a lab buying an operating company's records.

Can I submit data directly to an AI lab?

OpenAI accepts applications through its Data Partnerships form. Google accepts expressions of interest for its Content Offer Pilot, which names code repositories and admits partners by invitation. Microsoft accepts publishers through its Publisher Content Marketplace pilot. Meta, Amazon, Apple, Anthropic, xAI, and Mistral publish no open intake channel. To reach the leading labs without a separate application to each one, list your assets on DataVendor.

Does OpenAI pay for data?

OpenAI has paid for licensing deals with publishers and platforms. Its Data Partnerships page does not publish payment terms, so ask about compensation before you submit a dataset.

How do smaller companies sell data to AI labs?

Most sell through an intermediary, and the right one depends on the quality of the data. DataVendor has relationships with all of the leading AI labs and brokers sales to them, which lets a high-quality asset earn from several licenses. Managed vendors such as Mercor, micro1, Scale AI, and Appen set a price and sell the data themselves, which pays less but suits data of uncertain value. Turing runs the sale for a revenue share, and shutdown brokers such as SimpleClosure tie the sale to closing the company.

What data do AI labs refuse to buy?

Labs avoid data with unclear rights, data that is already public, and data that contains personal information. OpenAI's program states that it does not want sensitive or personal information, or information that belongs to a third party.

Find a buyer for your data

Before you close accounts or accept a set price, find out whether labs would pay for your codebase or records. Get an estimate for your codebase, or book a call for a quote on other data. If your data is high quality and you want to earn from more than one sale, DataVendor is the right place to sell it.

OlderAI Data Marketplaces: Where to Sell and How They Work
NewerBest Places to Sell Data to AI Companies in 2026

Estimator tools

  • Codebases
  • Company data

Resources

  • For buyers
  • Docs
  • Blog
  • Contact
  • Careers

Legal

  • Privacy
  • Terms
  • Trust

Social

  • GitHub
  • X
Provided by
  • For buyers
  • Docs
Sign inSign up