Who buys company data for AI training
The four kinds of buyers, what each one wants, and which one fits a company like yours.
Updated September 30, 2026
"Who buys company data?" has a different answer depending on what kind of data you have and what state your company is in. For operating records like SOPs, tickets and project history, the market has four kinds of buyers.
| Buyer | Usually buys from | What they want | How deals work |
|---|---|---|---|
| AI labs, directly | Large platforms and publishers | Huge, unique content libraries | Custom multi-year licenses |
| Training data vendors | Operating companies with documented workflows | Records that show how work gets done | Program terms, anonymization, milestone payments |
| Shutdown and asset-sale services | Companies that are closing | Complete archives: chat, email, tickets, code | One-off sale during wind-down |
| Data marketplaces | Companies with structured, reusable datasets | Tables many buyers can use | Listings sold to many buyers, often by subscription |
AI labs
The deals that make headlines are between labs and very large content owners. Google agreed to pay a reported $60 million to train on Reddit posts (CBS News, February 23, 2024), and News Corp's deal with OpenAI was reported to be worth more than $250 million over five years (Variety, May 2024).
These deals show how much the labs value data they can't get anywhere else, but they aren't a realistic starting point for most companies. Labs negotiate directly with a handful of suppliers and buy the rest through vendors.
Training data vendors
Vendors that supply the big labs have spent years paying experts to write and grade examples. Several now also pay companies for their operating records, because a company's real history of tickets, projects and decisions is hard to recreate with contractors.
For a normal operating company with 30 or more people, this is the most likely buyer. Programs differ in how they anonymize data, whether they take exclusivity, how they pay and whether they resell to more than one lab. Those differences matter more than the headline price.
Shutdown and asset-sale services
Closing companies have become active sellers, because their records have no other use. SimpleClosure, which helps startups wind down, said it had handled about 100 such deals in a year, with payouts between $10,000 and $100,000, and launched a product to scale it (Gizmodo, reporting on Forbes, April 17, 2026).
If you're shutting down, this route is quick, but compare it with a vendor program before you sign. The same archive can be worth more to a buyer that needs your specific industry.
Data marketplaces
Marketplaces like Snowflake Marketplace, AWS Data Exchange and Datarade are built for structured datasets that many buyers want: pricing data, firmographics, location data and so on. They're a good fit if you have a clean, repeatable dataset to sell over and over. They're rarely a fit for internal workflow records, which need custom cleanup and a specific license.
How to vet a buyer
Before you share anything beyond a description, ask each buyer:
- Who will actually use the data, and can they pass it on to anyone else?
- Who removes personal information, when, and can you review samples before anything is used?
- Where is the raw data stored, who can see it, and when is it deleted?
- Do they want exclusivity, and what do they pay for it?
- How is payment structured, and what milestone triggers each payment?
- Can they name other companies that have been through the process?
Vague answers to any of these are a reason to slow down. Our licensing agreement checklist turns them into contract terms.
Common questions
Can we sell directly to OpenAI, Anthropic or Google?
Direct deals exist, but they have mostly gone to large content owners like Reddit and News Corp. An ordinary operating company is more likely to deal with a vendor that assembles training data for those labs.
Should we talk to more than one buyer?
Yes, if you can. Prices and terms vary a lot between buyers, and an exclusive deal with the first one you meet can leave money on the table.
Sources
- CBS News, February 23, 2024: Google strikes $60 million deal with Reddit for AI training
- Variety, May 2024: News Corp inks OpenAI licensing deal potentially worth more than $250 million
- Gizmodo, reporting on Forbes, April 17, 2026: Failed companies are selling old Slack chats and email archives to train AI
Find out what your records are worth to a buyer
Answer a few questions about your company and your records. It takes about five minutes, and you don't send us any files.