AI training data prices: what company data is worth to AI companies
AI companies pay companies for their working records: support tickets, team chat, email, CRM history, code, drawings and recordings. This report puts the ranges we can show in one place, by kind of data and by company size, with the public deals behind them. Figures are as of November 2026, in US dollars, and every one is an estimate, not an offer.
In short
- Published buyer ranges for the operating records of a company with 30 or more staff run from $100,000 to $2 million. The typical range reported for an operating company is $200,000 to $300,000.
- The archive of a company that has already closed is usually licensed once, for $10,000 to $100,000.
- The best-known public sale of one company's business records: a closed US airline's emails, chat and other records went for a $10 million bid in its bankruptcy in August 2026.
- Size, years on the same tools and records that link across systems move a price more than the number of files does.
Price ranges by data type
Most deals license several kinds of records together, so a single kind alone usually sits in the lower part of its range. Ranges are rounded and in US dollars, as of November 2026.
| Data type | Range (USD) | Basis | What moves it |
|---|---|---|---|
| Operating records of a company with 30+ staff (tickets, team chat, CRM, email together) | $100,000 to $2 million | Published buyer ranges from several programs; the range on our seller page | Headcount at peak, years on the same tools, records that link across systems |
| Support tickets and help desk history (software and service companies) | $50,000 to $500,000 | Published buyer ranges for software and subscription companies | Whole threads with the fix and the outcome, internal notes kept, years of history |
| Private code with its review history | $50,000 to $500,000 | Published buyer range for private repositories with 3+ years of commits from 10+ engineers | Never public, owned outright, pull requests and tests that run |
| CAD and engineering files | $45,000 to $1.3 million | Published buyer range for native CAD with feature trees and test results | Native files rather than exports, several product generations, designs the company owns |
| Construction project records, estimate to closeout | $115,000 to $500,000 | Published buyer range for complete project records | RFIs, change orders and daily logs kept with the estimate and the final cost |
| Law firm files from closed matters | $50,000 to $1.2 million | Published buyer range; client consent or full anonymization required | Whole matter files, five or more years, privilege cleared |
| De-identified clinical notes and health records | $80,000 to $900,000 | Published buyer ranges for practices, health systems and digital health companies | Free-text notes, a year or more per patient, de-identification that holds up |
| Consulting deliverables with drafts and workpapers | $150,000 to $300,000 | Published buyer range for original Word, PowerPoint and Excel files | Drafts kept next to the final, five or more years of engagements |
| Call recordings | $20,000 to $250,000 | Published buyer range for call centers with 1,000+ recorded hours | Consent that covers reuse, each speaker on its own channel, transcripts |
| Egocentric (first-person) video | $1 to $4 a minute; about $10 an approved hour when commissioned | Public reporting on footage bought from creators (2025); a published capture program | Hands and tasks in view, releases from every wearer; prices are falling as supply grows |
| Archive of a closed company | $10,000 to $100,000 | The range on our seller page; reported wind-down archive sales | Systems still live or fully exported, peak headcount, someone with authority to sign |
Every range here is an estimate, not an offer. Workflows and RL environments are left out: each one is priced on its own, and the RL environments page explains what sets that price.
Price ranges by company size
These are the starting ranges our data value calculator gives a US company with 5 to 10 years on the same tools and three or four kinds of records, by full-time staff at the company's peak. Industry, country, language, linked records and whether a buyer has already seen the data then move the range up or down.
| Full-time staff at peak | Starting range (USD) |
|---|---|
| Under 30 | Usually no figure: most buyers start at 30 |
| 30 to 49 | $100,000 to $250,000 |
| 50 to 99 | $175,000 to $440,000 |
| 100 to 249 | $300,000 to $750,000 |
| 250 to 999 | $500,000 to $1.25 million |
| 1,000 or more | $1 million to $2.5 million |
A company that has closed is held to the closed-archive range above, whatever its size. Estimates, not offers. Run the calculator with your own answers.
AI training data prices per image, word and hour
Media archives are often priced per unit, so the public reporting gives a price per image, per word or per minute of video. Company records are almost never sold that way: a buyer prices the whole set of tickets, chat and files together. Use these as reference points for media, not as a rate card for a business.
| Unit | Reported price | Who and what | When | Source |
|---|---|---|---|---|
| An image (large stock archive, licensed in bulk) | 2 to 4 cents an image | A stock image site licensing most of an archive of about 200 million images to two large technology companies | 2024 | Reuters, 5 April 2024 (via Inc.) |
| A photo (user uploads) | 5 cents to $1 a photo | A photo hosting site negotiating terms for its archive of about 13 billion photos and videos | 2024 | Reuters, 5 April 2024 (via Inc.) |
| An image (selected, rights cleared) | $1 to $2 an image | Rates a training data vendor said buyers generally pay | 2024 | Reuters, 5 April 2024 (via Inc.) |
| A short video clip | $2 to $4 a clip | Rates a training data vendor said buyers generally pay | 2024 | Reuters, 5 April 2024 (via Inc.) |
| An hour of longer film | $100 to $300 an hour | Rates a training data vendor said buyers generally pay | 2024 | Reuters, 5 April 2024 (via Inc.) |
| A word of text | About $0.001 a word | The market rate for text quoted by a training data vendor | 2024 | Reuters, 5 April 2024 (via Inc.) |
| A minute of unused creator footage | $1 to $4 a minute ($1 to $2 for ordinary clips) | Footage bought from independent video creators | 2025 | The Decoder, citing Bloomberg |
| The code base of a closed startup | About $25,000 | A founder who wound down a small startup, sold through a wind-down service | 2026 | MIT Technology Review, 13 October 2026 |
| The workspace data of the same startup (email, chat, documents, design files) | A further amount, still pending when reported | The same founder; the figure had not been settled | 2026 | MIT Technology Review, 13 October 2026 |
Reported figures, rounded as published, in US dollars. The 2024 rates are the most recent public ones for images, clips and text, and buyers now pay less for kinds of media they already hold in bulk. Estimates, not offers.
Public deals with reported figures
Most data deals are confidential, so the public record is thin and leans toward very large archives. These are the reported figures we could check, with the reporting linked.
| What was sold or licensed | Reported figure | When | Source |
|---|---|---|---|
| A closed US airline’s business records (about 100 million emails, 500 million team chat messages, payroll and time records, code), sold out of its bankruptcy to a large technology company | $10 million winning bid; a rival bid of $7.5 million | August 2026 | Axios, 17 August 2026; CNN, 18 August 2026; Reuters, via the New York Daily Record, 19 August 2026; Time, 25 August 2026 |
| A social platform licensing its users’ posts to a large technology company | About $60 million a year | 2024 | The Decoder, citing Bloomberg and Reuters |
| An academic publisher’s income from AI content licensing in one fiscal year | $40 million (up from $23 million the year before) | Fiscal year to April 2025 | The publisher’s earnings release |
| A trade book publisher licensing nonfiction backlist titles for a three-year term | $5,000 a title, split between publisher and author | 2024 | Publishers Weekly |
| Unused video footage bought from independent creators | $1 to $4 a minute ($1 to $2 for ordinary clips) | 2025 | The Decoder, citing Bloomberg |
| Asking prices quoted to large technology buyers for images, video and text | $1 to $2 an image; $100 to $300 an hour of longer film; about $0.001 a word of text | 2024 | The Decoder, citing Reuters |
The airline sale needed the bankruptcy court's approval, and unions and more than 110 members of Congress objected that removing names does not make records anonymous. That is why how personal details come out is settled before any sample leaves a company: how de-identification works.
AI data licensing deal statistics
No public register of data deals exists, so these counts come from the analysts and reporters who track them, each dated and linked.
- One market analysis counted 52 publicly known AI data licensing deals by May 2025: 5 in 2023, 42 in 2024 and 4 in the first five months of 2025. News media made up 87% of the text deals, and only five languages were covered. (Neudata, 10 June 2025)
- The largest reported news deal: a newspaper group’s five-year license, estimated at more than $250 million in cash and credits, or about $50 million a year. The terms were not disclosed by either side. (Engadget, citing The Wall Street Journal, May 2024)
- One firm that helps failed startups close said it had handled more than 150 wind-downs since it began selling data in April 2026, and 47% of them now include a sale of code or data. (MIT Technology Review, 13 October 2026)
How we built these ranges
| Period | As of November 2026. Buyer ranges change often, sometimes week to week. |
| Published buyer ranges | The ranges buyers of company data state on their own public program pages, read in October and November 2026 and rounded. We give the range, not the buyer. Where only one buyer publishes a range for a kind of data, the basis says so in the singular. |
| Public deals | Reported figures from news and company filings, linked above. Companies are described rather than named. |
| Prices per unit and statistics | Reported in news and analyst work from 2024 to 2026, checked against the linked source in November 2026. Figures we could not confirm in a source, such as a price per hour of audio, are left out. |
| Our own ranges | The ranges on our seller page and the starting ranges of the calculator. |
| What is left out | Private quotes and offers from deals we work on. Those belong to the companies involved and are never published. |
What raises and lowers a number is explained in how much is my company data worth to AI companies. To compare an offer you already have, read how to check an offer.
Questions about these figures
Is any figure on this page an offer?
No. Every figure is a range and an estimate. A real number comes only from a buyer, in writing, after it has reviewed a cleaned sample of your data.
Why are the public deals so much bigger than the company ranges?
The public deals are a few very large archives: a national airline, a social platform, publishers with millions of pages. Most companies we talk to have tens or hundreds of staff, and the published buyer ranges for that size are in the hundreds of thousands.
Why does company size move the number so much?
More people at the peak means more tickets, threads, files and decisions on record. Buyers count staff because it is the quickest stand-in for how much work was written down.
How much is AI training data worth per image, per word or per hour of video?
Reported prices from 2024 and 2025 run from 2 cents an image for a bulk stock archive to $1 or $2 for a selected image, about a tenth of a cent a word of text, $1 to $4 a minute of unused creator footage and $100 to $300 an hour of longer film. Company records are rarely priced per unit: buyers price the whole set, so the ranges by data type above are the better guide for a business.
How many AI data licensing deals have been made public?
One count found 52 by May 2025, most of them news publishers in 2024, and most without a published price. Deals for the records of ordinary operating companies are almost always confidential, which is why this page leans on published buyer ranges as well as reported deals.
Can I use these figures in an article?
Yes. Please link to this page and keep the words "estimate, not an offer" with any figure you quote.
Didn't find your question? Ask us directly.