DataPacts

Broker track · Lesson 1 of 7 · 10:30

How the AI data market works

What you'll learn

  • Who holds the data
  • The routes data takes to AI developers
  • Who sits in the middle, and how they get paid
  • Where a good broker adds value

Transcript

One fact to know before you call yourself a data broker. On the public record, no seller has confirmed an amount received through any intermediary in the directory we keep. Sixteen intermediaries on record, and not one confirmed payout.

First, a word that confuses people. Data broker, in the privacy-law sense, means reselling personal data about individuals for marketing. That's a different trade. Here we mean a company that finds a licensee for a holder's own records and is paid a fee for it.

Chapter one. Who holds the data.

Six groups of sellers license data to AI companies. Companies, publishers, platforms and nonprofits, creators and stock libraries, collectives and public institutions, and individuals. They differ less in what they hold than in who owns the rights and who may sign.

Now the three roles. A seller holds the rights. An intermediary places the data. A buyer trains on it or resells it. If you're the broker, you're the middle one, and in the directory we keep, none of the intermediaries states that it acquires the data.

Advisers should know one more holder. When a company fails, a trustee or administrator can sell its data. In chapter eleven, the trustee or debtor signs with court approval, and in a US bankruptcy a court decides whether personal data may go with it.

Chapter two. The routes data takes to developers.

There are six routes from holder to developer. Crawling public web content. Licensing archives and platforms. Buying records outright. Commissioning contributor work. Bidding in insolvency. And generating synthetic data.

Crawling is the route the holder controls least. Licensing sits at the other end: a publisher or platform grants access for a fee and a term. That licensing end is where brokers work.

The insolvency route has a live example. A bankrupt US airline's de-identified business data went to auction on August fourteenth. A big tech company made the successful bid of ten million dollars, and an AI data company the alternate bid of seven and a half million. The court still had to approve the sale when this lesson was made.

Chapter three. The kinds of buyers and intermediaries.

On the buying side, data reaches a lab through a chain. Labs themselves get most of their training data by crawling the public web, and license the rest from publishers, platforms and archives.

Then come human-data companies. They run programs that license de-identified operating records from companies and pass them to labs. One of them states a commitment of one billion dollars over twelve months. Remember, that's a commitment it states, and the page doesn't count it as paid.

Others fill in the rest. Vendors license finished datasets, such as speech or medical images. Marketplaces match holders with developers, and the holder usually stays the licensor. And some enterprises buy domain data to train their own models, for their own use.

The directory we keep lists a hundred and twenty-three buyers and intermediaries. Brokers, sell-side advisers and shutdown services sit beside the programs but don't acquire data themselves. They find buyers for a holder and place the data with them.

Buyers differ most in what the license lets them keep. One human-data company's master terms take a nonexclusive, irrevocable and perpetual grant. One data platform is exclusive by default, with a kill clause that forces deletion within thirty days. Read those terms before you point any seller at any buyer.

The sixteen intermediaries split into four kinds by the work they do. Five advisers run the process for the holder. Six brokers place records through a mandate or a third party. Four matching services connect holders with buyers whose rules fit. And one referral service sends companies to a single buyer's program.

Chapter four. How intermediaries get paid.

Intermediaries are paid in four ways, as they state it. A success fee from the seller. A fee or referral payout from the buyer. A spread between two licenses. Or no intermediary at all, when the seller goes direct. And the payer matters more than the percentage, because it decides whose interest the intermediary serves.

Of the sixteen, four publish a percentage the seller pays, between six and twenty percent. One broker uses bands, from twelve percent of the first two hundred fifty thousand dollars down to six percent above three million. A second states twenty percent, only from an offer you accept. A third states ten percent of the license price.

Six state that the buyer pays them, or that the company pays nothing. One matching service puts it plainly: the buyer pays it a fee; the business owner pays nothing. Another says it's paid by AI labs. And six publish no percentage at all, which is where your questions start.

Buyers run referral programs too. One human-data company states fifty thousand dollars for enterprise deals up to seven hundred thousand, and seven and a half percent of deal value above that. Another program states ten thousand dollars when the deal closes. Different buyers, different scales.

Here's why the payer matters. A broker paid by the buyer is paid only if that buyer signs, so its income can favor the buyer with the highest referral fee rather than the best license for the seller. The pages show no evidence that any named intermediary steers sellers. The point is who pays, and that's why a good broker tells the seller first.

The third model hides the most. If an intermediary licenses the data to itself and licenses it on, it earns the difference between the two prices, and the seller rarely sees the second price. One data marketplace states that it splits its spread fifty-fifty with the agents who bring companies to it.

Chapter five. What deals are on record.

The deal tracker records seventy-four deals where a party or a filing confirms the buyer, the seller and the date. Twenty-seven were announced in twenty twenty-four, twenty-four in twenty twenty-five, and twenty in twenty twenty-six up to November ninth. Forty-three of the seventy-four are news publishers.

Only nine of those seventy-four carry an amount stated by a party or a filing. The rest disclose no amount, or state a pricing basis without one. The amounts come from a social platform, two publishers, a health-data company, a collective agreement and a court docket. That leaves private holders few comparable figures.

Only the filings of a health-data company price data per unit: a hundred dollars per delivered de-identified record, in a neurology agreement signed October twenty-second. The big social platform's two hundred and three million is an aggregate for unnamed buyers. And the market page puts it bluntly: no rate card exists, each contract sets its own price.

A deal isn't always a training deal, either. Of the seventy-four, fifteen state training, thirty-two cover display, grounding or answers, one excludes training, and twenty-six don't say. When you describe a deal to a seller, repeat only what the parties said.

Treat buyers' own numbers with care. One human-data platform states earn a hundred K to four million plus in its headline, and up to a hundred thousand in the estimator text on the same page. Most payout figures in this market are buyers' own claims. So a broker never promises a price.

Chapter six. Where a broker adds value for the seller.

So where does a broker earn the fee? Reach: one party searching several buyer programs for you. Preparation: some work on a separate copy, as one broker states. And control, if you handle it right. Some intermediaries never receive the data, others prepare a de-identified copy, and the seller should know which before sharing anything.

The costs sit on the other side. A seller-paid fee reduces proceeds, and published rates run from six to twenty percent. An exclusive brokered license can bar a direct program: one human-data company lists data already licensed exclusively through a broker as not a fit. And no public figure shows a broker gets a better price than going direct.

That's why mandate terms matter. Read exclusivity, notice and tail. One broker publishes a non-exclusive mandate with thirty days' notice and a twelve-month tail limited to listed buyers. In most published models the seller signs the license with the licensee, so the seller can enforce it without the intermediary.

Seven questions separate one intermediary from another. Who pays you, by buyer? Is the mandate exclusive? Who signs the license? Do you hold the data? Which buyers do you approach? Have you closed a license? Does anyone else get a fee? A seller should ask you these, so have written answers ready.

To recap. Sellers hold the rights, buyers train, and a broker places the data and gets paid by someone, so say who. No confirmed payouts, no rate card, and no promised prices. And not selling at all remains a valid outcome.

Next is lesson B2, Finding sellers: who holds valuable data. The full course is free at datapacts dot com slash academy.

Sources on DataPacts