Insights · 6 min read

AI Training Data Pricing: Five Models and How Each Is Built

Before you ask what a price is, ask how it is built.

Bearish AI · October 21, 2026

AI training data pricing looks mysterious because the numbers rarely come with their logic. In practice, most deals reuse a small set of structures. Once you know which structure you are in, you know which levers you can pull. This guide walks through five, using the public evidence we could verify.

1. Per unit

A unit is a task, a title or a case. Epoch AI’s interviews put graded tasks mostly at $200 to $2,000 each. HarperCollins priced book licenses at $5,000 per title, split between author and publisher. Per-unit pricing rewards you for building something countable and checkable, and punishes you for volume without quality: a buyer who finds that many tasks cannot be solved, or can be gamed, stops paying for them.

2. Flat package

One price for the whole archive. Reports on shut-down companies describe $10,000 to $100,000 per company, and the Spirit Airlines bankruptcy sale was a single $10 million bid. Flat pricing is simple, but the buyer is doing the work of guessing what the archive can yield. MIT Technology Review reports that buyers estimate how many tasks a body of data produces and then quote very different prices as a result. If you cannot explain your own yield, you are leaving the guess to them.

3. Term license

You license for a period and get paid for it. Reddit’s disclosed contracts ran two to three years with an aggregate value of $203 million. The New York Times and Amazon deal was reported at $20 to $25 million a year. Term pricing fits content that keeps growing, such as publications and platforms. It fits a static company archive less well, which is why archives are usually sold as packages.

4. Per model and field of use

A practitioner memo on AI data licensing says the licensee typically gets to train, fine-tune and evaluate models, often limited to one named model, with extension to derivative or successor models available for an additional fee. The same memo recommends a perpetual license for models trained during the term, paired with finite access to the dataset. In other words, price often depends not just on the data but on how many models it can touch and for how long.

5. Exclusivity premium

Two environment founders interviewed independently told Epoch that exclusive deals cost roughly four to five times more than non-exclusive ones. That is two people’s estimate, not a standard, but it frames the trade. We cover it in detail in /insights/exclusive-vs-non-exclusive-data-licensing.

The levers inside any structure

• Scarcity. Data that is not already public is harder to find. Epoch estimates the stock of public human text at about 300 trillion tokens and projects it will be fully used between 2026 and 2032, while noting that non-public data such as instant messages seems unlikely to be used at scale because of legal issues. Material that has been cleared for use is therefore uncommon.
• Difficulty. Epoch’s interviewees wanted tasks that models solve some of the time, around 2 to 3 percent or better, not never and not always.
• Provenance. Buyers increasingly expect a provenance file covering where data came from, how it was processed and any known risky subsets, according to the same practitioner memo.
• Warranties. Buyers typically ask for a warranty that you own or can license the data. If you cannot give it, expect a lower price or no deal.
• Cleanliness. Suppliers scrub personal information and exclude HR records, and buyers scrub again, MIT Technology Review reports. In one reported bankruptcy sale, an outside firm had to certify de-identification to the court before each transfer.

A worked example of how a price gets built

The numbers here are hypothetical, chosen only to show the logic. Suppose an archive could yield 300 tasks that pass a buyer’s review. At the lower end of the reported per-task range, $200 each, that is $60,000. At $600 each it is $180,000. Now suppose the buyer sees only the raw files, and guesses that 100 tasks might be buildable at a risk discount. The flat package they offer might be a fraction of the first figure. The gap between those two outcomes is not greed on either side. It is who carries the uncertainty about how many tasks the data supports.

Questions to ask any buyer about price

• What is the unit? Per task, per package, per term or per model, and what happens to units that fail review.
• Who pays for cleaning? Scrubbing, de-identification and certification have a cost. Ask who carries it.
• What is the scope? Which models, which uses, and what fee applies to successor models.
• What is the term? When do your obligations end, and what survives?
• Is it exclusive? If so, ask for the premium to be stated, not buried.

What you cannot negotiate

You cannot negotiate away the buyer’s compute cost, which Epoch cites at about $2,400 per task during training, or the quality bar their researchers apply. You can negotiate scope, term, exclusivity and who bears the cost of cleaning.

Preparation changes which structure you can use at all. Raw files usually sell as a package at the buyer’s guess. Checked, linked, graded material can sell per unit. The free estimate at /estimate shows both a raw and a prepared range for your archive so you can see the gap. As always, have a lawyer review any term you sign.