Valuation
How much is your company's data worth to AI labs? (2026 guide)
Why AI labs pay for company data, the publicly reported deals, the six value drivers and illustrative ranges for mid-market companies. A 2026 guide.
Henry Jackson, Founder, Knowledge Fuel. . 6 min read
AI labs, model developers and enterprise AI teams are paying for private data because the public web is running out. Most of that money has gone to large publishers and platforms with legal teams and negotiating power. This guide covers what has been paid, what sets a price, and what a mid-market company can reasonably expect.
Why buyers pay for private data
Models learn from text, speech, images and records, and the public supply is finite. Research from Epoch AI, presented at ICML in 2024, projects that the stock of public human-generated text could be fully used between 2026 and 2032 if current trends continue (Epoch AI). What remains is the data companies hold privately: support tickets, call recordings, maintenance logs, contracts, claims files and sensor histories.
That data has two qualities public text lacks. It describes real work in the language professionals actually use. And it is paired with outcomes: the ticket that was resolved, the claim that was paid, the machine that failed three weeks after a sensor drifted. Outcomes are what let a model learn to predict and decide, and they cannot be scraped.
What has been paid in public
The reported deals give a sense of scale. These are agreements between other companies, cited from public reporting.
- Reddit licensed its content to Google in a deal reported at about $60 million per year in February 2024, according to Reuters.
- News Corp and OpenAI agreed to a deal reported at more than $250 million over five years in May 2024, according to The Wall Street Journal.
- Shutterstock reported $104 million of AI licensing revenue in 2023, per CEO Paul Hennessy in a Bloomberg News interview published in June 2024.
- Wiley recognized a $23 million content rights project with a large tech company in fiscal 2024, described as limited in duration and narrow in use, then reported $49 million of AI licensing revenue in fiscal 2026 and lifetime AI revenue above $110 million, per its fiscal 2024 results, fiscal 2026 results and Form 10-K.
- Taylor & Francis, part of Informa, gave Microsoft non-exclusive access to archive content for an initial data access fee of about $10 million, with recurring annual fees over the following three years, disclosed in May 2024. Informa expected over $75 million of AI partnership revenue in 2024, per The Bookseller and Informa's 2024 half-year report.
- Amazon and The New York Times reached a deal reported at $20 million to $25 million per year, according to The Wall Street Journal in July 2025.
Two things stand out. The money is real. And nearly all of it went to companies with scale and lawyers. A regional insurer or a 300-truck carrier holds data with the same qualities, at smaller volume, and has had no route to the same buyers. That gap is the reason Knowledge Fuel exists.
The six drivers that set a price
Buyers do not pay by the gigabyte. They pay for how hard the data is to get elsewhere and how much work it saves them.
Uniqueness
Data that exists nowhere public is the starting point: dispatch calls with regional accents, underwriting notes, PLC logs from a specific class of machine. The more specialized the data, the fewer substitutes a buyer has.
Volume
More records help, with diminishing returns. A million support tickets beats fifty thousand, but fifty thousand tickets with resolutions attached beats a million without them.
Years of history
Value starts to compound around three or more years of records. Long histories capture seasonality, product changes and rare events that short samples miss.
Cleanliness and labels
Consistent fields, timestamps and existing labels (resolution codes, failure types, quality grades) cut a buyer's preparation cost. Data that already carries outcomes is priced well above raw exports.
Rights clarity
A buyer needs to know you can license the data. Customer contracts, employee agreements, vendor terms and consent records are checked before any price is final. Clean rights raise value. Gaps lower it or remove it.
Exclusivity
An exclusive license pays more but closes the door to other buyers. A non-exclusive license pays less per deal and can be sold more than once. Over time, most sellers do better with non-exclusive terms.
Illustrative ranges for mid-market companies
The figures below are estimates for planning only. Actual value depends on uniqueness, quality, rights and buyer demand.
| Company | Data | Illustrative range |
|---|---|---|
| 60-person insurance agency | 10 years of claims notes and call recordings | $40K to $200K |
| 300-person logistics company | 8 years of dispatch records, telematics and support tickets | $100K to $600K |
| 1,200-person manufacturer | 15 years of machine sensor logs and maintenance records | $400K to $2.5M |
| 5,000-person enterprise | Millions of hours of call center audio | $1M to $10M+ |
A typical mid-market dataset falls in an illustrative range of $25K to $1M+. The same company often holds several datasets, and each can be licensed on its own terms.
How a valuation is produced
The process is built so you never expose your data to get a number.
- NDA first. Nothing is discussed in detail until a mutual NDA is signed.
- Description. You describe the data: systems, record counts, date ranges, formats and what outcomes are attached. A 10-minute form covers it. No files are needed.
- Small sample. A small, de-identified sample is reviewed to confirm quality and structure. Full datasets never move at this stage.
- Rights check. Contracts, consent records and vendor terms are reviewed alongside your counsel to confirm what can be licensed and on what terms.
The result is a range, a list of likely buyer types and a de-identification plan. You approve every buyer and every permitted use before anything is delivered. Payment follows delivery, and non-exclusive terms let you license the same data again.
Find out what yours is worth
Start with the free estimate. It takes about ten minutes, asks for no files and returns an illustrative range based on your industry, data types and years of history. If the range is worth pursuing, read how selling works or go straight to the valuation form. Either way, you keep ownership and you decide who buys.
Questions on this topic
Do I have to hand over my data to get a valuation?
No. A valuation works from a description of the data and a small, de-identified sample reviewed under NDA. Full datasets do not move until a contract is signed and you have approved the buyer and the delivery plan.
Is a smaller company's data worth anything?
Often, yes. Buyers pay for uniqueness and context, not only volume. A 60-person agency with ten years of claims notes and call recordings can hold a dataset no public source contains. Actual value depends on uniqueness, quality, rights and buyer demand.
Do I give up ownership when I license my data?
No. A license grants permission to use the data for a defined purpose, usually model training only. You keep ownership, approve every buyer, and can license the same data non-exclusively to more than one buyer.
Henry Jackson
Founder, Knowledge Fuel. Knowledge Fuel buys and licenses operational data from companies and sells it to frontier AI labs, model developers and enterprise AI teams.