DataPacts

Seller track · Lesson 2 of 9 · 9:14

What buyers actually want

What you'll learn

  • Which kinds of data buyers want
  • Who the buyers really are
  • How buyer programs work
  • How to read payout claims

Free lesson

Enter your email to watch this lesson and the rest of the course.

Transcript

AI companies buy data of almost every kind, from chat threads and email archives to books, medical records and robot demonstrations. But ask what any of it is worth, and there's no price list. The figures on record are a handful of contracts, list prices and claims, each stated by one party on one date.

First, which kinds of data buyers go after.

Buyers sort data into six families. Workplace records. Code. Published text. Media and design files. Regulated records. And physical-world and agent data. Each family reaches a model developer by its own route.

Workplace records sell when they show how work was done. Threads, tickets, documents and system histories that link a request to its outcome. A request, what people did about it, and how it ended. That chain is what a model can learn from, and it's why old support logs and project histories get asked for by name.

Some programs print unit prices. One states two dollars per accepted page of documents. Another states fifty dollars per ticket, spec or support thread, when it's bundled with a codebase. Notice how different the units are already.

To see what a real workplace data set looks like, take the asset schedule from that airline's bankruptcy. It lists a hundred million emails, five hundred million chat items, and more than six hundred thousand IT tickets. It lists five hundred sixteen code repositories too. Ninety-seven and a half million passenger profiles are not included.

Code is the clearest example. One program's listing for software companies asks for private repositories that were never public. Three or more years of commit history from at least ten engineers. Fifty thousand lines of code the company owns outright. And code written by people, with tests that run.

The same listing rules out forks, side projects, code mostly written by AI tools, single-author repositories, and client work an agency doesn't own. Notice the pattern. Ownership is the first test. Before anything else, a buyer asks whether the code is really yours to license.

Published text is licensed mostly by publishers and platforms in negotiated deals. Media is licensed as archives with rights cleared by the owner. And regulated records carry the largest disclosed license values, plus the rules that decide whether they may be licensed at all.

Physical-world and agent data works differently. It's mostly collected to order by vendors and programs, because few holders already own it at scale. So if you run an ordinary company, your workplace records and code are the likelier fit.

Even wanted data can be stuck. Platform terms decide whether records can leave. A chat tool's terms can bar export or training. Recording consent rules can apply to calls. And a sale of health records needs authorization under HIPAA. Which rules reach your records is a question for a lawyer.

Now, who is actually on the other side of the table?

Company data reaches AI developers through seven kinds of counterparty. Only some of them train models themselves. The others take a license and pass it on, or find a buyer for you.

Frontier labs train the models, but their disclosed licenses are mostly with platforms and publishers. Few publish a door for an operating company's records. The most common published route is a company data program. It licenses your records to itself first, then passes them on to labs, which the program pages rarely name. So you often won't know who ends up with the data.

Then there are builders of training environments, who take codebases, tickets and business-app records. Vendors, who collect or commission data and license finished datasets. And marketplaces, where the holder usually stays the licensor.

Two more kinds. Brokers don't train models. They find buyers for your data, and they're paid by the buyer, the seller, or both. Enterprises that train their own models license domain data for their own use, not to resell it.

Next, how the programs actually work.

What is a company data partnership program? A standing offer to license your internal records. It differs from a one-off deal in four ways. The offer is standing, the page publishes the process, an intake form starts it, and a payout claim is stated before any negotiation.

Nearly every program follows the same five stages. First, intake. Then scoping under an NDA. Then de-identification and transfer. Then review and acceptance. And only after that, payment.

Eligibility is where programs differ first. One states thirty or more full-time employees. Another states twenty or more and three or more years on some pages, and no minimum on another page of the same program. One simply says any size, any industry, anywhere. So don't rule yourself out from a headline. Read the page, and ask.

Buyers split on who de-identifies your data, and who pays. One covers extraction, anonymization and transfer. One anonymizes inside your own environment before anything leaves. One says it de-identifies sensitive information itself. And one covers the cleanup costs.

Check fit before anything else. One program publishes twenty-five industry listings, each asking for different records. A buyer is worth approaching only if it takes the data type and systems you hold.

Now the part everyone asks about. What do they say they pay?

Every payout figure in an ad is the program's own claim. One program says it has paid out two hundred and fifty million dollars to partners, without saying who the partners are. And no seller has publicly confirmed an amount received under any program.

Programs also disagree with themselves. Thirteen pairs of values contradict each other on the same operator's own pages. One program says a wire arrives two to four weeks after signing, and elsewhere on the page, as soon as processing is complete.

Watch the wording too. One program states a commitment of one billion dollars to enterprise data acquisition over twelve months. That is a commitment. It is not money already paid.

Real money does show up in filings. A public medical-data company reports two hundred million dollars in license fees over three years from a drug-development company, including fifty million upfront. That license is for training a foundation model, and nothing else.

Court dockets show bids too. At an auction, the successful bid was ten million dollars, for a bankrupt US airline's de-identified data. As of early November that sale was not approved, so it's still a bid.

Don't line those figures up and average them. Units do not convert. A price per million tokens, per pull request, per accepted page, per hour of footage and per delivered record measure different things.

So how do you choose between buyers?

Six criteria separate buyers before any offer arrives. Data fit. Who holds the license. Exclusivity. Payment structure. De-identification. And the buyer's identity and track record.

Exclusivity decides whether picking one buyer closes the door on the others. The published terms run from exclusive by default, to a sole-seller window of ninety days after you sign, to no exclusivity at all.

Payment comes in three structures. A fixed payment agreed before anything is shared. An upfront fee plus a share of downstream use. Or a revenue share only, which pays on later sales.

A buyer's legal entity, incidents and corporate changes tell you who will hold the license years later. Breach litigation against a counterparty can be on the public record. And a counterparty can change owner, or enter chapter 11.

Several programs publish no license terms at all, so ask in writing, and ask more than one candidate. Who is the licensee? Who receives the data? What are the exclusivity terms? And what is the payment structure, and its trigger?

So, in one line. Know which of the six families you hold, know which kind of buyer you're facing, treat every payout as a claim, and compare written terms before you choose. And honestly, not licensing is a valid outcome too.

Next up is lesson A3, Who owns it: employees, customers, contracts.

Read and watch the full lesson free. Enter your email to unlock every lesson of the academy.

Sources on DataPacts