What AI Training Data Is Worth: Lessons from Disclosed Deals

For a market this consequential, remarkably few prices have ever been printed. Most data licensing happens privately, under NDA, between a lab that does not want competitors reading its sourcing strategy and a seller who does not want anyone knowing they licensed. But a handful of transactions have become public, through filings, announcements, and court records. Read together, they sketch the price logic of the entire market.

This essay quotes no deal figures. The public record is easy to search, and the numbers age quickly. What matters is the shape of each deal: who paid, for which property of the data, and on what basis. That shape is what your own archive will be priced against.

The deals that printed

The clearest archetype is the major publisher archive deal: news organizations licensing decades of edited, dated, rights-clean text to AI labs, usually as recurring arrangements that renew year over year. What buyers were paying for is visible in the terms. Not raw words, which the open web supplies for free, but editorial judgment, verifiable provenance, and a continuing stream of fresh, dated material.

The second archetype is forum-corpus licensing by large platforms: years of threaded human conversation, question and answer, argument and correction, licensed on an annual basis and to more than one buyer at once. Two lessons print here. Conversation with visible human judgment is a distinct asset class, and non-exclusive licensing lets the same corpus be licensed more than once.

The third is per-title book licensing by a big-five publisher, which put a unit price on the professionally written long form: an offer per title, worth a few thousand dollars each, for a defined training use. Sellers learned that even a single well-made object has a floor, and that unit pricing is how collections scale.

The fourth is the one that should most interest ordinary companies: a bankrupt airline's operational archive, sold at auction as part of the estate. The seller was not a media house. It was an operating business whose manuals, records, and operational history cleared at a real price under the least flattering circumstances a seller can have. If a distressed estate's archive sells at auction, a healthy company's archive negotiates.

The last two archetypes are standing markets rather than single transactions. Licensed video now trades per minute, tiered by resolution, subject matter, and exclusivity. Speech datasets trade in tiered bundles priced by the hour, stepping upward with language coverage and consent documentation. Modalities with unit prices behave like commodities at the bottom of the market and like fine lots at the top, where rights and rarity separate them.

What the deals agree on

Across all six archetypes, buyers paid for the same underlying things. Rights first: every disclosed deal is a license of material the seller demonstrably owned, which is why rights-unclear corpora trade at a discount or not at all. Judgment second: edited text, threaded argument, professional production, operational reality. Continuity third: the arrangements that renew are pricing a future stream, not just an archive. And legibility last: the sellers who could describe exactly what they held, at what scale, under what consent, closed faster and closed higher.

It is equally instructive what none of the deals paid for. No disclosed transaction rewards volume by itself: an enormous corpus with murky provenance appears nowhere in the public record, because no buyer can put a contested asset in front of its own counsel. And none of them price novelty for its own sake. The premium attaches to the boring virtues, ownership, documentation, continuity, in every deal that printed. For sellers, that is encouraging news: the qualities buyers actually pay for are qualities an ordinary company can genuinely have.

Scarcity is the multiplier

The open web is no longer the counterfactual. What a lab can scrape, it will not buy; what it can buy identically elsewhere, it will not pay a premium for. The premium sits where supply is genuinely short: consented speech across accents and languages, production code with intact history, workflow records that show how decisions were actually made, structured archives linked to real outcomes.

Scarcity is why two archives of the same byte count can sit at opposite ends of a valuation band. It is also why the first question a specialist asks is not how big, but how replaceable. An archive a buyer could reassemble from public sources prices like a commodity; an archive no one else can produce prices like a lot.

Why ranges, not prices

Given all this, a point price quoted before review is a fiction. No two archives share provenance, rights posture, context, or scarcity, and the final number is set by the deal itself: which buyers are matched, what exclusivity is chosen, what the review reveals. That is why Aribase publishes bands rather than prices. The current department spans:

The low ends are deliberately conservative, set at figures we can defend before anyone has reviewed anything. The high ends are anchored to the disclosed transactions above and to what matched buyers currently pay at the top of each class. Where your asset lands inside its band is what the deal process exists to discover.

Reading your own archive against the market

Three questions position any archive. Do you own it, cleanly enough to license it? Is it replaceable, or would a buyer struggle to assemble the same corpus from any public source? And does context travel with it, the history and surroundings that turn records into evidence of how work actually happens? Yes, no, yes is the profile that reaches the top of a band. Everything else is a conversation, and the conversation is free.