DataPacts

Broker track · Lesson 2 of 7 · 9:05

Finding sellers: who holds valuable data

What you'll learn

  • The six kinds of sellers
  • Seller or go-between?
  • Which data types buyers ask for
  • Why bankrupt companies can sell

Free lesson

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

Transcript

Where do you look for sellers? One source index sorts them into eighteen seller types, in six groups. So the answer is wider than most people think. Anyone who holds rights in records or content that an AI developer wants can be a seller.

Chapter one. The six groups of sellers.

Here are the six. Companies, from operating firms to closed ones. Publishers of news, books and journals. Platforms and online communities. Creators, stock libraries and music rights holders. Collectives and public institutions. And individuals who sell their own photos, video and voice.

Now the distinction that matters most for you. A seller holds the rights. If you're a broker, you're on the other side, and so is a shutdown service or a pooling platform. It doesn't own the data. It places a holder's data with buyers, and the seller stays the seller.

Keep three roles straight in your head. The seller holds the rights. The intermediary places the data. And the buyer either trains on it or resells it.

There's a handy way to split your search. The seller type tells you who may sign. The data type tells you which buyers will want the records. We'll cover both, starting with who signs.

Chapter two. Companies, and who may sign.

Companies are the biggest group for a broker. They sell the same kind of asset the whole way, workplace and product records. But the person who may sign changes. An operating company signs through its officers. After closure, the wind-down party signs. After a merger, the acquirer. And in insolvency, the estate signs and the court approves.

Closed companies are a real pool. One shutdown service states that more than three hundred and fifty companies have offered assets on its marketplace since its April launch. That's a claim by the service itself, dated September 2026. It does show where closing companies turn.

What does that look like in money? A founder of a defunct crypto startup states net proceeds of roughly twenty-five thousand dollars for its code base. One founder, one claim. It isn't a price list.

Chapter three. Publishers, platforms, creators and individuals.

Publishers license archives and live content, but many of their deals are display or grounding deals, where training isn't mentioned. Read what a deal actually covers. Platforms license what their users posted, usually through a paid interface, and the users aren't parties to the deal. So ask who may license, and under which terms.

Creators and artists sell in one of three ways. Directly. Through a platform that pools their work. Or through a stock library or label that licenses on their behalf. The route decides who negotiates. And watch the payout claims, because each platform's figures are its own, and one platform can contradict itself.

Collectives sign for many rights holders at once. Universities, libraries and governments often publish data openly, or keep commercial use subject to permission. Individuals sell their own media and voice through apps whose terms decide how long the license lasts, and whether it's exclusive.

Chapter four. What buyers ask for.

Once you know who signs, ask what they hold. The data index lists forty-three data types, sorted into six families. Every figure on it is the stating party's own, never an estimate.

The families are workplace records, code, published text, media and design files, regulated records, and physical-world and agent data. Each reaches a developer by its own route.

Start with workplace records. They sell when they show how work was done. Threads, tickets, documents and system histories that link a request to its outcome. A dull archive of files is worth less than one that tells a story.

Code has its own checklist. Buyers look for private repositories with long commit histories written by people. One buyer's listing asks for three or more years of history from at least ten engineers. And ownership is the first test.

That same listing rules some code out. Forks. Side projects. Code mostly written by AI tools. Single-author repositories. And client work an agency doesn't own. Screen for these before you call a prospect a seller.

Regulated records carry the largest disclosed license values. One genomics-and-oncology company reports two hundred million dollars in license fees from a single customer over three years. But the same rules that raise the price decide whether the records may be licensed at all.

The other families route differently. Published text is licensed mostly by publishers and platforms in negotiated deals. Media is licensed as archives with rights cleared by the owner. And physical-world data is mostly collected to order, because few holders already own it at scale.

To see how big an estate's asset list can get, look at one bankrupt airline's schedule. It lists a hundred million emails as included. And ninety-seven and a half million passenger profiles as not included. The estate carved out the consumer records, and offered the business records.

Chapter five. Insolvency, where the estate is the seller.

When a company becomes insolvent, its data becomes an estate asset. A trustee, administrator or liquidator may sell or license it to AI buyers. But only with the court or creditor approval the insolvency law requires. And note who decides: the officeholder, not the board.

In the US, a court sale runs in steps. Bidding procedures. An auction. A privacy review. A sale hearing. And a sale order. Only the order lets data move.

Here's why the order matters. At one bankrupt US airline's auction in August 2026, the successful bid for its de-identified business data was ten million dollars. As of the page's November date, the sale wasn't approved. A successful bid is an auction result, not a sale. Never tell a client otherwise.

The gate differs by country. In the US it's a court hearing. In Germany it's creditor consent for significant acts. In France it's a judge's authorization. Different approval, same rule: no approval, no sale.

Personal data sets the limit of what an estate can offer. In the US, a consumer privacy ombudsman may be appointed. And the legal definition covers consumer data only, so employee and business data fall outside it on its face. That gap is a risk to flag, not a loophole to use.

Three real cases show the range. A genetic database sold to a nonprofit, with court approval, for three hundred and five million dollars. Not an AI training license. A bankrupt airline's archive, still waiting for approval. And a chatbot company's sale that leaves training data out.

If you ever help an officeholder, the order of work follows the deadlines. First secure exports before subscriptions lapse, because vendors delete data on a clock. Then place retention holds. Then carve out personal data. Then appoint an independent de-identification provider.

Chapter six. What to ask before you approach a seller.

So here's a short list to run on every prospect. Who holds the rights? Who may sign at this stage? What records exist, and which family are they? Is personal data inside? And is anyone on a clock, like a closing company or an estate?

The source index puts it plainly. No rate card exists, and every figure is the stating party's own filing or claim. So don't promise a price. Show every figure as a claim, and say who made it.

Last point. Not every holder should sell. A license to train usually can't be undone once a model is trained. So not selling is a valid outcome, and an honest broker says so.

So, to recap. Six groups of sellers, and a broker is never one of them. The seller type says who signs, the data type says who buys. Insolvent estates sell only with approval. And never promise a price. Next up is lesson B3, Qualifying a company's data.

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

Sources on DataPacts