Selling to Labs
Non-exclusive, exclusive, first-look
Two sellers with identical corpora can transact at a 5× difference in value purely on structure. Exclusivity, term, field of use, refresh commitment and the definition of what "delivery" even means move the number more than volume ever will. This is the shape of the deals being done, and the terms worth fighting over.
The four shapes
| Structure | What the buyer gets | Relative value | Fits when |
|---|---|---|---|
| Non-exclusive licence | Right to train on the corpus; you may license it to others | Baseline | Your corpus is good but substitutable, and you want several buyers |
| Exclusive licence | Sole right to train on it for a term and field of use | 3–10× baseline | The corpus is genuinely scarce and one buyer's advantage depends on others not having it |
| First-look / right of first refusal | A window to bid on new material before anyone else sees it | Small fee, large option value | You produce continuously and want a relationship without locking the asset |
| Feed subscription | Committed ongoing delivery at a stated cadence and volume | Recurring; usually exceeds the archive within 2–3 years | Your data is generated by an ongoing operation rather than sitting in an archive |
The multiples above are directional, from deals we have seen and public reporting, not a rate card. What is robust is the ordering, and the fact that the biggest structural lever is exclusivity — which is also the one sellers give away cheapest, usually by accepting a broad definition of it without negotiating the boundaries.
Exclusivity has four dials, not one
"Exclusive" on its own is a bad term for both sides. Negotiate it as four separate things and you will usually find the buyer only needs one of them:
- Field of use. Exclusive for foundation-model training but not for evaluation, research, or a different modality. A buyer building a coding model rarely needs to block a medical imaging licensee.
- Term. Two years of exclusivity followed by non-exclusive is often worth nearly as much to the buyer as perpetual, because model cycles are short. Perpetual exclusivity should cost perpetual money.
- Territory or counterparty class. Exclusive against named competitors rather than against the world. Cheap for you to give, valuable to them.
- Scope of corpus. Exclusive on the archive to date, non-exclusive on everything generated afterwards — or the reverse.
The archive-to-feed shift
The market has repriced around continuous access. A static archive is a depreciating asset; a committed feed is a subscription, and subscriptions get valued on a multiple rather than as a one-off.
Most large disclosed deals are archive plus ongoing access, not archive alone. The reason is straightforward: an archive is a one-time capability gain that every subsequent model inherits for free, while a feed keeps supplying what the model has not seen. If your data is generated by an operation that is still running, the feed is the product and the archive is the sample.
What a feed commitment needs to be worth anything:
- A stated cadence and floor volume — "monthly, no fewer than 50,000 records" — because an uncommitted feed is priced at zero.
- Schema stability, with a notice period for changes. A feed that breaks the buyer's loader quarterly is a support cost.
- A quality band that later deliveries must stay inside, measured the same way the first delivery was.
- A remedy that is not termination. Fee abatement for a missed month keeps the deal alive; a termination right invites the buyer to walk when they have what they need.
Terms that quietly decide the value
| Term | What to watch |
|---|---|
| Derivative outputs | Can the buyer generate synthetic data from your corpus and license that onward? If unaddressed, the answer tends to be yes, and your exclusivity is worth much less than you think. |
| Model weights are not returnable | Termination cannot untrain a model. Deletion clauses cover the corpus, not its effects — so price the licence as permanent in substance whatever the term says. |
| Survival on termination | Which rights persist for models already trained? Buyers will insist on this and they are right to; charge for it rather than resisting it. |
| Audit | Your ability to verify use is realistically limited. Prefer structures whose value does not depend on policing — an upfront fee beats a usage royalty you cannot count. |
| Indemnity and its cap | Buyers want IP indemnity. It is often grantable, and it is one of the few things that meaningfully raises price — but cap it at a multiple of fees received, never uncapped. |
| Acceptance criteria | Define what a conforming delivery is before delivery, ideally as a measurable test. Otherwise "acceptance" becomes a renegotiation with your data already in their hands. |
| Attribution and publicity | Most labs will not name suppliers. If a reference matters to you commercially, it has to be negotiated in, and it is usually available in exchange for very little. |
A sequencing that works
- Paid pilot on a carved-out slice, with a measurable contribution test agreed up front. Small money, and it converts the argument from claims to numbers.
- Non-exclusive licence of the archive to the first buyer who clears the pilot, with a short exclusivity window as the sweetener rather than the core.
- Feed subscription layered on once delivery has worked twice, which is when a buyer will actually commit to a cadence.
- Exclusivity sold last, if at all, once you know what the corpus does for a model and what several buyers will pay for it.
Selling exclusivity first is the most common structural error, because it is the term you can only sell once and you will price it before you know what it is worth.
Read next: Ship it training-ready — delivery is a term too, and the one most likely to sour a signed deal.