How AI Data Licensing Deals Work: Structures, Terms, and What Gets Negotiated

With rare exceptions, AI data deals are licenses. The owner keeps the data and grants a defined right to use it, for defined purposes, by defined parties, for a defined period, exclusively or not. Nothing changes hands permanently except money and a copy. If you're new to the idea, our explainer on what a data license is covers the basics before the deal terms below.

That matters because a license can be shaped. The same catalog can be licensed non-exclusively to several buyers, or exclusively to one at a premium. Training rights can be granted while display rights are withheld. A license can expire while a model trained under it doesn't. Every one of those choices moves the price, which is why two deals for similar data can differ by an order of magnitude. What data is worth, and the disclosed numbers behind that range, is covered in our guide to AI training data cost. This guide is about the agreement underneath the number.

The largest disclosed deals are mostly recurring: News Corp and OpenAI at more than 250 million dollars over five years, Google paying Reddit around 60 million dollars a year, Amazon and The New York Times at a reported 20 to 25 million dollars a year. Those are living sources that keep producing, so buyers pay for continuity. Archives price differently. Spirit Airlines' operational history drew a single 10 million dollar bid from Google in bankruptcy court, a flat purchase of a finished corpus; our breakdown of the Spirit Airlines data deal covers how that one was structured and why the court is involved. Marketplace deals are usually per-unit or revenue share, because they aggregate thousands of owners who could never each negotiate a flat fee with a lab. The full set of reported figures, updated monthly, lives on our AI training data statistics page.

If you want to know what these deals pay, here's every one with a reported number. Most contracts stay confidential, so read this as the visible edge of the market rather than its average. We keep the sourced figures current on our AI training data statistics page.

Two patterns show up in that list. Buyers pay the most for sources that keep producing, and they pay for them as a recurring fee. A finished archive sells once for a flat price, and the Spirit bid is the clearest example. If you own company data rather than a publisher's catalog, the six-figure floor on our business data page is the reference point that applies to you, and our guide to how much your company's data is worth to AI labs walks through what moves the number.

The headline fee is usually the fastest part of the negotiation. The slow parts are the terms that decide what the fee buys.

• Permitted use: training, fine-tuning, evaluation, retrieval at answer time, or some combination. These are different rights with different values, and a deal that grants "AI use" without saying which is a deal that will be argued about later.
• Model coverage: which models, whether successors are included, whether the license covers internal research only or commercial products, and whether affiliates and cloud partners can touch the data.
• Term and survival: how long the license runs, whether the data must be deleted at expiry, and the clause almost every buyer insists on, that models already trained survive the license ending, because nothing can be untrained.

Exclusivity is the biggest single lever on price. An exclusive license can be worth several times a non-exclusive one, but it also stops you signing any other deal for the term, so price what you're giving up, not just the premium. Field-of-use exclusivity, exclusive for video generation but not for robotics, is the common middle ground. Our guide to exclusive vs non-exclusive data licenses has the full comparison, including when the premium is worth taking.

Then the risk terms:

• Representations and warranties: you promise you hold the rights you're licensing and that the data was acquired lawfully, and the buyer wants that promise to be specific, per asset where possible.
• Indemnification: who pays if a third party claims otherwise, usually with a cap tied to the deal value.
• Privacy obligations: the standard for removing names and personal details, and who bears the cost if personal information turns up.
• Audit and reporting rights: how you verify usage in recurring and revenue-share deals.
• Output restrictions: limits on a model reproducing licensed material verbatim, which appear increasingly in content deals.
• Confidentiality: near universal. So few deal terms are public because almost every agreement forbids disclosing them.

For buyers, those warranty and provenance terms are the whole point of paying rather than scraping. Courts have drawn a line between training and how data was acquired, and unauthorized acquisition is where liability has concentrated. Our guides to rights-cleared training data and AI data provenance explain what documented permission looks like in practice and why buyers now require it before diligence closes.

Most failed deals fail on things that could have been found in week one.

The rights chain is the classic one. A company owns its video archive, but the music in the background belongs to a label, the footage was shot by a contractor whose agreement never mentioned AI training, or user-uploaded content came in under terms that don't cover licensing to third parties. Warranties you can't stand behind are a liability.

Third-party data hides inside first-party data. Financial and market data is the sharpest example, and a frequent buyer question. Premium datasets often arrive under upstream licenses that prohibit redistribution or derivative use, personal financial information is regulated, and material non-public information can't be licensed to anyone. The data a firm generated itself may be licensable. The data it bought from a vendor usually isn't.

Scope creep on the buyer side is the mirror image. "All data, all models, perpetual, worldwide, sublicensable" is a wish list. Owners who sign it have licensed everything once and can never license it again.

And the exclusivity trap catches owners who take the premium without modeling the term. Two years of exclusivity to one lab, in a market where buyers are multiplying, can cost more than the premium paid for it.

Less formally than the contracts suggest. Very few buyers issue a specification for the data they want, and most don't know until they see it. Deals typically start with a sample (a curated slice of a dataset put in front of a research team), a reaction, then a conversation about scope and terms. That's why owners who wait to be asked rarely are, and why marketplaces exist.

There are three routes to a deal. Direct negotiation works at News Corp scale, where the owner has a legal team, a data engineering team, and a catalog famous enough that labs come to it. Court-supervised sales like Spirit's are the rare case where an archive becomes an asset only because a company is being wound down. For everyone else there's the marketplace route. A licensed marketplace pools rights holders into one counterparty, standardizes the terms above, handles packaging and privacy review, and puts curated samples in front of buyers continuously. Our guide to AI data licensing covers that market end to end.

Because the data already exists and already cost money to produce. A company that licenses operational history, support records, or a content catalog is earning a second return on work it did anyway, and a non-exclusive license can earn it more than once. It also puts the owner in control of terms, scope, and privacy, which is the alternative to that same data being scraped for nothing or auctioned in a liquidation. Our guide to data licensing as a business model makes the full case, with examples at every scale and the situations where it doesn't fit.

Value varies a lot and not every archive is worth licensing. Buyers pay for scarce, connected, rights-clean data with visible outcomes, and a serious process is selective (inventory, rights review, scope, privacy review, packaging, then terms).

Our guide to licensing company data for AI covers how that works from the owner's side, and our guide to how to sell data to AI companies covers what sells and what it pays. To get a first read on what your own company holds, take Troveo's free, five-minute data value assessment (ten questions).

Troveo is the marketplace route. More than 7,000 rights holders have licensed real-world video, audio, gaming, robotics, and now business data through us, around 95 percent of them exclusively, and we've paid out more than 20 million dollars to owners. If you own data, we handle the terms above once, at the platform level, and you're paid on every sale, not just the first. If you're buying, it's one agreement instead of thousands, with rights documented per asset and data delivered training-ready. Like every serious participant in this market, we keep individual deals confidential.

Browse the catalog in Lens, take the data value assessment if you own data, or talk to us about either side of a deal.