Data Marketplaces & Intermediaries

AI data deals are multiplying as per-deal prices fall — and pricing is shifting to usage and rev-share

Source: AI Industry Reviews · Oct 6, 2026

The clearest picture yet of the training-data licensing market has arrived, and it tells two stories at once. Across 312 disclosed deals signed between early 2023 and September 2026, quarterly deal volume climbed roughly thirteen-fold — from a trickle of about three a quarter to nearly forty — as buying licensed data went from a novelty to the default way frontier labs source material they can defend in court. Licensing, in other words, won the argument. But over the same window the median annual deal value fell from about $9.4M to $3.6M. The market is getting much wider and, for the typical deal, cheaper — expanding and commoditizing in the same motion.

Underneath the headline numbers, the more consequential shift is in how deals are priced. Flat, one-time fees — long the default — have dropped from 58% of deals to 35%, while usage-based pricing has roughly doubled (13%→27%) and revenue-share arrangements have tripled off a small base (4%→13%). Buyers increasingly want to pay in proportion to what a dataset actually contributes to a model, not a lump sum up front. The report also puts rare public numbers on unit economics: video around $2.65 a minute, transcribed audio near $58 an hour, still images about $1.12 each, and books near $4,300 a title — with frontier-model developers accounting for roughly two-thirds of all deals.

For a data owner, the falling median is not bad news so much as a sorting signal. When price is linked to usage and outcomes, the premium flows to data that keeps earning — corpora that are scarce, cleanly sourced, and rich enough to matter across many training runs. Generic, single-modality text is precisely what a crowded, commoditizing market drives toward the floor. A consented, multimodal interview corpus — video, audio, live problem-solving, and verifiable provenance bound to a real, credentialed person — is the opposite kind of asset: hard to assemble, hard to replace, and well-suited to usage- and revenue-share terms that reward durable value rather than a one-time dump. As the market's center of gravity moves to paying for contribution, provenance and depth stop being compliance checkboxes and become the pricing lever.

Key Points

• A new survey of 312 training-data licensing deals (Jan 2023–Sep 2026) finds quarterly deal volume up roughly 13x — from 3 to 39 — as licensing becomes the default sourcing route
• But the median annual deal value fell from $9.4M to $3.6M: the market is widening and commoditizing at the same time, not just getting bigger
• Pricing is moving off flat fees (58%→35%) toward usage-based (13%→27%) and revenue-share (4%→13%) structures — buyers want to pay for what a corpus actually contributes
• Reported unit rates: video ~$2.65/min, transcribed audio ~$58/hr, still images ~$1.12, books ~$4,300/title — with frontier-model buyers driving 67% of deals
• When pricing is usage- and outcome-linked, durable, high-provenance, multimodal corpora win; generic single-modality text is exactly what the falling median is commoditizing