How to value a dataset for AI training

In short

AI training data has no list price. Value is set by how scarce, useful and legally clean the data is for a specific buyer.
Published reference points range from about 2 cents per image to $4 per minute of video and $10 million for a large corporate archive.
Exclusivity, outcome labels and rights documentation can raise value more than extra volume.
Valuations are uncertain because buyers price on a guess about how much the data will improve their models.

Why there is no price list

Each buyer values data by how much it will improve their model on tasks they care about, which they estimate before buying. MIT Technology Review reported that buyer valuations vary widely because they rest on uncertain research theses (MIT Technology Review). The same data can be worth very different amounts to two buyers.

The seven value drivers

Driver	Raises value	Lowers value
Scarcity	Not online, specialist domain	Similar data is public
Quality	Edited, accurate, consistent	Noisy, duplicated, machine-generated
Structure	Roles, timestamps, outcomes, links	Flat text dumps
Volume	Enough to change model behavior	Too small to matter
Coverage	Many sources, years, variations	One template or narrow period
Rights	Documented, cleared, de-identified	Unclear ownership or consent
Exclusivity	Exclusive or first access	Widely licensed

Published reference points

Images: 2 to 4 cents (Freepik), 5 cents to $1 (Photobucket) and $1 to $2 (Defined.ai) per image (Reuters via The Star)
Video: $1 to $4 per minute of footage (Bloomberg via The Star)
Text: about $0.001 per word through one vendor (Reuters via The Star)
Corporate archive: $10 million for Spirit Airlines data including about 100 million emails and 500 million Teams messages (TravelPulse)
Small startup code: about $25,000 for one codebase (MIT Technology Review)

See all figures with sources on our AI data prices page.

A simple framework

Estimate volume in the unit buyers use (messages, hours, images, documents).
Find the nearest reference price for that unit.
Adjust for scarcity and quality against comparable data.
Adjust for exclusivity: exclusive deals command a premium.
Subtract the cost and risk of preparation, de-identification and legal review.
Test the range with more than one buyer.

Where wind-down platforms value data

One wind-down platform's CEO told MIT Technology Review it values assets on headcount, years in operation, data volume and how the team used its tools (MIT Technology Review).

Questions

Can you give me a price for my data?

We can give a range after a consultation, based on comparable deals and current buyer interest. Anyone quoting a firm price without seeing an inventory is guessing.

Does more data always mean more money?

No. Past the point where a buyer has enough of a type, extra volume adds little. Scarce, well-labeled data usually beats bulk.