Who Buys Company Data for AI Training?

Key takeaways

The buyers of company data for AI are AI labs, big tech companies building their own models and, less often, companies training AI on data from their own field.
Labs want records the public web does not hold: OpenAI's Data Partnerships programme asks for large datasets "not already easily accessible online to the public today".
In August 2026 Google bid $10 million for Spirit Airlines' internal emails, Teams messages and spreadsheets to train its AI models. The sale still needs court approval.
When we review a company's data, depth matters more than size: years of history, outcomes attached and reasoning written down.

Do AI companies buy data?

Yes. Buying company data is now a normal part of building models, because the free supply is running out. Speaking at the NeurIPS conference in December 2024, OpenAI's former chief scientist Ilya Sutskever said "we've achieved peak data and there'll be no more", comparing the public internet to a fossil fuel that is finite.

What models still lack is a record of real work: decisions made across weeks of messages, approvals and fixes. That record only exists inside companies. Business Insider reported in October 2026 that ordinary workplace data is more valuable than ever as OpenAI, Anthropic, Meta and others build AI agents for businesses. Agents learn a job by seeing it done, so buyers pay for the threads, tickets and files where it was done.

The buyers also publish what they want. OpenAI launched its Data Partnerships programme in November 2023, asking for large datasets that are not already public and for data that "expresses human intention", such as long-form writing or conversations rather than disconnected snippets.

Who buys company data for AI training?

Public reporting points to three types of buyer. Most of the deals with a known price are for published content rather than company records, but they show who pays for AI training data and how much they are prepared to spend.

Alongside these buyers sit data companies that collect and prepare records and supply them to labs. On our page about how to sell company data to AI companies we describe the market the same way: AI labs and the data companies that supply them pay to license real business records.

Who buys company data for AI training?
Buyer type	Examples in public reporting	What they have bought or asked for
AI labs	OpenAI, Anthropic	OpenAI asks for large private datasets through Data Partnerships. A 2025 court ruling found Anthropic spent "many millions of dollars" buying print books to scan for training
Big tech building models	Google, Amazon, Microsoft, Meta	Google: Reddit's content at a reported $60 million a year (reported in July 2026 to be ending, with renewal talks under way), and a $10 million bid for Spirit Airlines' internal data. Amazon: New York Times content at a reported $20 million to $25 million a year. Microsoft: HarperCollins nonfiction books. Meta: AI data deals with news publishers
Companies training their own AI	Bloomberg (BloombergGPT)	Its own archive of financial documents, 363 billion tokens, combined with public data

What are AI labs buying company data looking for?

AI labs buying company data are mostly trying to teach models how work gets done, so they look for depth rather than volume. Business Insider reports that AI developers increasingly want long-horizon data that gives a full picture of how a job is carried out, and that a corpus loses value when a company holds back its chat, CRM or email records.

Demand is not even across sectors. The same report says fields AI companies are trying hardest to automate, such as law and accounting, see especially high demand.

When we review a company's data, these are the things we look at first:

Years of history: a long-running record shows how decisions played out over time, not just a single moment.
Outcomes attached: closed deals, resolved tickets and paid invoices let a model connect a decision to its result.
Reasoning written down: threads where people explain why they chose something, not only what they did.
Specialist work: decisions from finance, legal and engineering teams are among the records buyers name most often.

Which company data deals have been made public?

Few deals for company records are announced, and fewer with a price. The clearest is Spirit Airlines, which stopped flying in May 2026 and is selling its data in bankruptcy. Reuters reported in August 2026 that Google agreed to buy the airline's internal business data for $10 million, including employee emails, Microsoft Teams messages, spreadsheets and calendars, and that Google plans to use it for product development and training its AI models. The data is to be de-identified first, with no customer or personally identifiable information.

The sale is not yet final. 121 members of Congress wrote to Spirit and Google about former employees' privacy, and the bankruptcy court hearing has moved to 6 November 2026. Google has said it is "not looking to buy any personal information from Spirit".

Closed companies are part of the market too. MIT Technology Review reported in October 2026 that AI labs are buying data from startups that have shut down. For what drives the price in any of these deals, see what company data is worth.

Do companies buy data to train their own AI?

Some build their own models, but the evidence that they buy other companies' records is thin. Bloomberg built BloombergGPT in 2023 on a 363 billion token dataset drawn from its own archive of financial documents, combined with a 345 billion token public dataset. Its main source was data it already owned.

In our view, companies like this matter to a seller mainly as a sign of where demand is heading: models built for one field need records from that field.

Why do buyers care whether you own the data?

A buyer can only use records the seller has the right to license. OpenAI says it is not seeking datasets with sensitive or personal information, or information that belongs to a third party, and the Spirit sale shows how closely staff records are watched once a deal becomes public.

That is why we start with rights. Only records your company has the rights to are licensed, and data you hold on behalf of clients is out of scope. Our guide to the company records AI companies buy goes through each record type.

How do you sell company data to AI companies with Pangea.ai?

We start on the company's side. Our process has four steps, and nothing is licensed until you've agreed the terms. It's a licence, not a sale, so the data stays yours. The first step is the free check, with no commitment.

Check: tell us which tools your team uses and how many years of history they hold.
Review: we look at what you have and tell you whether it could qualify and what affects its value.
Agree terms: you decide which sources, date ranges and uses are included before you sign.
Get paid: the data is licensed for a set use and term. You're paid for the licence and keep the original.

Frequently asked questions

Who feeds data to AI?

Early models learned mostly from the public web and from licensed content, such as Reddit's posts, which Google licensed in 2024 for a reported $60 million a year. Now that public text is largely used up, AI labs building frontier models, and the data companies that supply them, pay to license real business records, because how work gets done is the part models haven't been able to learn from public text.

Can you sell data to AI companies?

Yes, if your company has the right to share it. AI companies buy data from operating businesses and from companies that have closed. The usual limits are your contracts with customers and suppliers, how personal data about staff and customers is handled, and whether the records are deep enough to interest a buyer. This isn't legal advice, so involve your own counsel before you sign.

How do I sell data to OpenAI?

OpenAI runs its own Data Partnerships programme, launched in November 2023. It asks for large datasets that are not already easily available online, particularly long-form writing and conversations, and says it is not seeking sensitive or personal information, or information that belongs to a third party. Many companies instead reach labs through data companies that supply them.

How much do AI companies pay for data?

Few prices are public, and most of those are for published content. Google's licence for Reddit content was reported at about $60 million a year, a deal reported in July 2026 to be ending, Amazon's deal with The New York Times at $20 million to $25 million a year, and Google bid $10 million for Spirit Airlines' internal data. Treat these as examples, not prices to expect. Depth and years of history move the value most.

Do AI companies buy data from small businesses?

They can. Business Insider reports that, generally, any company, dead or alive, big or small, can sell its data as long as it has enough of it. In our experience sector matters less than depth: if your team has years of records showing how work gets done, it is worth checking what you hold.

Sources

Who buys company data for AI?