A dataset is valued with three approaches: cost (what it took to create), market (what comparable data sold for) and income (what it will earn the buyer). For AI licensing, buyers weight scarcity and usability far above cost. The realistic output is a defensible range, later tested against written offers.
Why datasets are hard to value
Data does not behave like other business assets. It can be licensed to several buyers without being used up. Its value changes with the buyer and the use. Most transactions are private, so comparable prices are scarce. Accounting rules generally keep internally generated data off the balance sheet, which means most companies have never put a figure on it.
Valuation practice borrows the approaches used for other intangible assets: cost, market and income. Each answers a different question, and each has a known weakness when the buyer is an AI lab. Using all of them together gives a range and a set of arguments, which is more useful in a negotiation than a single number.
The cost approach
The cost approach asks what it would take to create or replace the dataset. Add up the labor that produced the records, the software that stored them, and the work needed to clean and document them.
Its strength is that the inputs are known. Its weakness is that cost and value are loosely related. A company may have spent heavily producing records no one else wants. Another may hold years of expert decisions that were a byproduct of normal work and cost nothing extra to capture, yet cannot be recreated at any price.
For licensing, cost is most useful as a floor on the supplier's side. The preparation work, including export, de-identification and rights review, is a real expense. An offer that does not cover it is not worth accepting, unless a partner absorbs that work.
The market approach
The market approach asks what similar data has sold for. It is the method most people reach for first, and the one with the thinnest evidence.
Some reference points exist. Neudata's market analysis of 2025-06-10 lists the Google-Reddit agreement at about $203M in total contract value and the Le Monde and Prisa agreements with OpenAI at $1-5M. Those are content deals with large publishers and platforms. For company operational data, the public evidence is what buyers advertise. On their own sites, checked on 2026-11-16, Mode advertises $100K-$5M and Polyshares advertises $100K to $2M and above.
Comparables need adjusting for volume, exclusivity, term and rights. A reported headline may cover several years or include more than data. An advertised range is not a completed sale. The market approach sets the outer bounds of a reasonable expectation. It does not locate a specific dataset within them.
The income approach
The income approach asks what the data will earn for whoever uses it, then discounts that to a present value. In theory it is the most rigorous method. In practice a supplier can rarely see the buyer's economics.
An AI lab does not earn revenue from one dataset in a traceable way. The data improves a model, and the model earns money alongside many other inputs. What the supplier can estimate is the buyer's alternative: what it would cost the lab to obtain equivalent data by commissioning experts to create it, or by finding another company that holds something similar. When no alternative exists, the buyer's willingness to pay rises.
The income approach is more direct on the supplier's side. A license with a recurring fee is an income stream, and a non-exclusive license can be repeated with other buyers.
How the approaches compare
No single approach settles the question. The table shows where each one helps.
| Approach | Question it answers | Best use | Weakness for AI licensing |
|---|---|---|---|
| Cost | What it took to create, or would take to replace | Setting a floor that covers preparation work | Cost says little about what a buyer will pay |
| Market | What comparable data has sold for | Framing a realistic range | Few public comparables, and most are advertised or bundled |
| Income | What the data will earn or save | Negotiating structure and recurring terms | The supplier cannot see the buyer's economics |
Why AI buyers price on scarcity and usability
AI labs already have the public internet. Public web data is sold as a commodity. Bright Data's public dataset price list, for example, shows up to $0.0025 per record with a $250 minimum. What labs lack is the material that never reaches the web: how work is actually done inside companies.
Scarcity is the first test. The question is whether the buyer can get this data, or something close to it, anywhere else. Private records, long histories and specialized fields score well. Anything already published does not.
Usability is the second test. Data has to be legally usable, which means clear ownership, compatible contracts and personal details removed. It has to be technically usable, which means consistent formats, intact structure and enough documentation to interpret the fields. A scarce dataset that fails either test is worth little, because the buyer cannot train on it with confidence. What the supplier spent creating it does not enter the calculation.
A worked example, without dollar figures
Consider two companies with data to license. The first is a regional insurance services firm. It holds many years of claims files in one system, with adjuster notes, correspondence and outcomes linked by claim number. Client contracts are on a standard template that the firm controls.
The second is a marketing agency of similar size. It holds campaign reports and email archives spread across several tools, with no common identifier. Much of the material was produced for clients under contracts that assign ownership to them.
On cost, the two look alike. Both paid staff for years to produce the records. On the market approach, both fall into the same broad category of company operational data. The difference appears on scarcity and usability. The insurance firm's files show expert decisions with outcomes attached, in a specialized field, with linkage and clear rights. The agency's data is fragmented, close to material available elsewhere, and largely owned by third parties.
A buyer would likely place the first well up its range and may decline the second. Neither result could be predicted from cost.
Checklist to raise value before going to market
Much of the gap between a low and a high valuation can be narrowed before any buyer is contacted. Outrich Data handles de-identification and packaging for its suppliers, but the points below hold whoever does the work.
- Inventory the systems. List each source, the record types, the years covered and rough volumes.
- Confirm ownership. Separate records the company owns outright from those created for clients or under license.
- Review contracts. Check customer agreements, NDAs and data processing terms for limits on use or disclosure.
- Preserve linkage. Keep the identifiers that connect records across systems, so they can be replaced consistently during de-identification.
- Document the data. Write down what each field means, how statuses are used and what changed over time.
- Keep outcomes. Records that show how a case ended are worth more than records that stop midway.
- Plan de-identification. Decide what personal and client-identifying details must be removed, and by what method.
- Decide on exclusivity and refresh. Know in advance whether the company will grant either, and on what terms.
- Name the signer. Identify who has authority to approve a license, so an offer does not stall.
Common questions
Which valuation method do AI buyers use?
Can a company put its data on the balance sheet?
Does a formal valuation report help in a sale?
Does de-identification reduce a dataset's value?
Sources
- Neudata, AI data licensing market analysis2025-06-10
- Mode, data licensing page2026-11-16
- Polyshares, advertised range for company data2026-11-16
- Bright Data, public dataset price list2026-11-16