Transcript
Everyone tells you to get more than one offer. Yet there's no public evidence that parallel processes raise prices for company data. Several buyers can still make sense. You just need to know what it costs you in promises.
First, the two ways to run several buyers.
Parallel selling ends in one of two ways. In a competitive process, several bidders chase one grant, and the losing bidders receive nothing. In non-exclusive stacking, you grant several licenses, and each buyer pays you separately. A third pattern is one buyer after another, and the same warranties apply.
The only fully documented competitive process for company data comes from a bankruptcy court. A bankrupt US airline put its data archive up for auction. A big search company was the successful bidder, at ten million dollars. Another company was the alternate bidder, at seven and a half million. As of November ninth, the court still hadn't approved the sale.
Outside a court, no operating company has published how it ran competing bids. So these steps come from programs' own intake processes, not from a market norm. Give every bidder the same data description. Sign a separate NDA with each one. Set a sample policy, with deletion if there's no deal. And ask for one written offer from each, by a common date.
Stacking is the other route, and filings show sellers doing it. One big social platform reports two hundred three million dollars in aggregate contract value for licenses signed in January 2024, with terms of two to three years. A developer Q and A site announced two licensees. A health data company licensed to more than one buyer. Neither the social platform nor the Q and A site says whether its licenses are exclusive.
Some programs do the stacking for you. One says you can license the same data to anyone else. Another says the same record can go to more than one lab. A third says you approve each buyer, can block competitors permanently, and get paid on each relicense. They're licensees that sublicense.
Next, the promise that can break a second deal.
It's called a no-prior-grant warranty. You promise each new buyer that you haven't already granted the same rights to someone else. So the order of signing decides which of your promises stay true.
Take one program's section 5.1, headed Due Authority. It has you warrant that you haven't granted a third party the same data on the same or similar terms. Sign a similar license elsewhere first, and you can't give that warranty truthfully. Sign this program first, and later grants fall outside its wording.
And each extra license repeats your ownership warranty, so one rights gap can break several promises at once. Liability caps vary. One program's cap is three times the fees. Another's is two times the fees of the prior twelve months. A third states an uncapped ownership warranty.
Windows can pause parallel selling altogether. One program is the only seller of a record for ninety days after signing, then you can renew or opt out. A health data company agreed an eighteen month side letter limiting third-party licensing of neurology records. And one program's licenses are exclusive by default. That blocks the scope it covers.
Even non-exclusive programs can slow you down. One has a twelve month non-circumvention period, so you can't deal directly with its buyers. Another lets licensees keep access for twelve months after termination, and three years for academics. And leaving a program isn't instant: thirty days' notice, then processes can run another ninety days.
Keep a register of every grant. Write down the licensee, the data covered, any window end date, the field of use and term, the warranty wording, and the payment trigger. No buyer requires it. It just lets you answer every no-prior-grant question from one record, not from memory.
Now, negotiating. What can you actually trade?
Start with where it happens. Buyers that publish license terms put the negotiable points in an order form. In one program's master terms, the price in the order form is the only payment, and exclusivity and audit point back to it. And honestly, no public record shows how often sellers change a buyer's terms.
Prepare before the first call. Set a walk-away point. Rank the levers by what they're worth to you. Name the terms you won't give. And know your fallback if no deal closes, including keeping the data.
Seven levers move a deal, and each trades against price: price structure, scope, exclusivity, retraining, deletion, audit and your protections. Walking away is a lever too.
Price itself is negotiable. A share moves the risk to you. A fixed sum moves it to the buyer. One program says nothing upfront. A health data company reported two hundred million dollars over three years, fifty million of it upfront.
Another structure is a fee per record. One health data company's filing describes an order form for five hundred thousand de-identified records at a hundred dollars each, plus five million for updates, up to fifty-five million in all.
Scope is often the easiest ask. A narrower purpose limits what your data may train without touching the price structure. One health data filing limits a license to the sole purpose of a foundation model. An author group says one publisher's license caps outputs at two hundred consecutive words, five percent of a book, or both. Compare a grant sublicensable through multiple tiers.
Exclusivity is a lever buyers say they pay for. One data labeling company says buyers pay for stricter exclusivity. A relicensing program says exclusive licenses pay more up front. But no published figure says how much more for company data. The only multiple on record is for exclusive RL environments, and it covers no company data.
After training, the buyer usually keeps the models. One data labeling company states that trained models can't be unlearned. So if you want more, ask for a stop date for new training, no distribution of models after termination, and notice of new versions. No buyer publishes whether it accepts those. Deletion deadlines run from thirty to a hundred and twenty days.
An audit right lets you check the limits you negotiated, but few programs publish one. One supplier program allows an audit once a year on fifteen business days' notice. A written deletion confirmation is the lighter substitute. And watch your protections: a liability cap, who indemnifies whom, and your exit.
Levers get traded in pairs. Exclusivity against price and term. A narrower field of use against a perpetual grant. A revenue share against an upfront fee. A shorter deletion deadline against accepted model retention. Ask only when the other side values it more than you do. No buyer publishes which trades it accepts, so each is a position to test.
Some intermediaries negotiate the license for you while the AI side pays them. So ask every one who pays them, and how.
Last big question. How long does all this take?
By the programs' own statements, a deal takes from under three weeks to about three months. One program says under twenty-one days from first call to first wire. Another says about three months from first conversation. No law, filing or seller report sets a standard.
Before signing, the published claims are short. A first response in a day or two business days. A sample review in five business days. An offer in one to two weeks. Negotiation in weeks three to six, by one broker's schedule. A license in five to ten days at one program. They come from different programs, so they don't add up to one timeline.
Why do the figures differ? They measure different spans. One program counts from first call to first wire. Another counts from the signed agreement to the wire. A third counts to the first accepted contribution, which isn't payment. And a timeline on a website is a claim. Only the signed agreement sets deadlines.
Two steps have no published duration at all: the rights and consent check, and your internal approval. Those depend on you. Exporting your data has waits too. One suite says an export typically takes seventy-two hours and can take up to fourteen days. So start those first.
After signing, the buyer's review window matters most, because payment usually waits for acceptance. One program's terms give thirty business days to reject, and silence is deemed acceptance. Payment is claimed at two to four weeks after signing by one program, and thirty to ninety days by another.
Four hundred forty-five days. That's how long a bankrupt US airline's data sale had run from the petition to the scheduled hearing, without an approval. It's the only company data sale with its clock in a primary court record.
Finally, when one buyer, or none, is the right answer.
Running several buyers is optional. You can grant one exclusive license, priced for exclusivity. You can use one program that relicenses on your behalf. You can sign a single non-exclusive license and stop there. Or you can decide not to license at all. Talking to one buyer, or to none, is a valid outcome.
So, in one line. Several offers can help, but every extra license repeats your promises, so keep a register, read each order form for windows and warranties, and trade levers in pairs. Saying no is a perfectly good answer too.
Next up is lesson A8, Getting paid: payment, tax, accounting.
Read and watch the full lesson free. Enter your email to unlock every lesson of the academy.