# Data Licensing Offer: How to Review It Before Signing

> Got a data licensing offer? Check price against public deals and a $20K-per-employee median, then exclusivity, scope, payment timing and warranties.

Canonical: https://origindatapartners.com/guides/review-a-data-licensing-offer
By: Origin Data Partners
Status: Published
Published: 2026-11-10
Updated: 2026-11-10
Sources checked: 2026-11-10

## Quick answer

A data licensing offer is fair when the price sits within the range of comparable deals and the terms do not quietly take more than the price pays for. Check the price per record or per employee (a median of about $20,458 per employee in 12 deals Origin helped place), then exclusivity, scope, payment timing, holdbacks, warranties and who will actually use the data.

## Key points

- Price is only half the offer. Exclusivity, field of use, the right to keep trained models and the liability cap decide what the price actually buys.
- Public deals give anchors: per-record fees, multi-year terms, fixed payments invoiced monthly or quarterly.
- Make onward use a clause: name the end user or the class of end user, and require your consent before any sublicensing.
- De-identification should be done by a provider you retain and direct, before anything leaves your systems.
- A second offer is the strongest negotiating tool an owner has.

## What a normal data licensing offer contains

Most first offers arrive as an email, a term sheet or a short order form after a call. Whatever the format, a complete offer answers eight questions: which records, for what use, for how long, exclusive or not, at what price, paid when, with what promises from you, and what happens to the data and to trained models at the end.

If an offer answers only the price, the rest will be filled in later by the buyer's standard contract. That is where most of the value moves. A law firm that advises sellers puts it plainly: a licence that is silent on these points tends to be read in favour of whoever holds the data at the end, which is the buyer ([Latsoudis + Associates](https://www.latsoudislaw.com/data-ai-greece/data-licensing-contract-key-terms/)).

This guide walks through each clause an owner should check, with what is normal and what to push back on. For the contract itself, see [what a data licensing agreement should include](https://origindatapartners.com/guides/data-licensing-agreement).


## Is my data licensing offer fair? Price benchmarks from public deals

Public prices are scarce, and they vary widely. Quartz summarised the market this way: "The rates vary by orders of magnitude depending on the content, the publisher, and the buyer." ([Quartz](https://qz.com/ai-training-data-pricing-licensing-deals-market-052126)). Use the anchors below as reference points, not a price list.

- Large archives. News Corp's five-year deal is reported at more than $250 million, and Reddit disclosed data licensing arrangements with an aggregate contract value of $203 million and terms of two to three years ([Quartz](https://qz.com/ai-training-data-pricing-licensing-deals-market-052126)).
- A court-supervised sale. Spirit Airlines' internal data, about 100 million emails and 500 million Microsoft Teams chats, drew a winning bid of $10 million in August 2026, pending a federal judge's approval ([Axios](https://www.axios.com/2026/08/17/google-spirit-airlines-bankruptcy)).
- Per record. In an October 2026 filing, Tempus disclosed "$50 million in data license fees of $100 per delivered record, invoiced monthly" for up to 500,000 de-identified neurology records ([Tempus Form 8-K](https://www.sec.gov/Archives/edgar/data/1717115/000119312526429533/d135568d8k.htm)); its quarterly results describe the deal as $55 million over a five-year term ([Tempus Q3 2026 results](https://www.biospace.com/press-releases/tempus-reports-third-quarter-2026-results)).
- Small archives. A founder winding down a crypto startup netted roughly $25,000 for his code base, with about $15,000 more pending for workspace data ([MIT Technology Review](https://www.technologyreview.com/2026/10/13/1145755/ai-labs-buying-data-from-dead-startups-to-train-models/)).
- Per employee. In 12 deals Origin helped place where headcount is known, companies were paid a median of about $20,458 per employee, from $12,778 to $52,857. Multiply by your headcount for a first bracket, then adjust for years of history and rights. The method is on [company data value per employee](https://origindatapartners.com/guides/company-data-value-per-employee).


## Offer clause by clause: what is normal and what to push back on

The table is Origin's view as an independent data partner, built from the deals and contracts we have seen and the public sources cited on this page. Your lawyer should read the final contract.

| Offer clause | What is normal | What to push back on |
| --- | --- | --- |
| Scope of systems and records | A named list of systems, date ranges and record types, delivered as a prepared copy | "All company data", live system access, or records you do not own (client files, designer drawings) |
| Field of use | Training, fine-tuning or evaluation of named model types, stated precisely | "Any purpose", resale of raw data, or use to build a product that competes with you |
| Exclusivity | Non-exclusive, or exclusive for a limited period, field or subset | Perpetual or worldwide exclusivity without a matching premium or minimum payments |
| Term and trained models | A fixed term with deletion of raw data at the end; buyers usually keep trained models | No deletion duty, or an open right to create and resell derivative datasets |
| Price and structure | A fixed fee, a per-record or per-delivery fee, or a fee plus a smaller recurring part | A revenue share with no minimum, no reporting and no audit right |
| Payment timing | Payment on signing or on delivery, or fixed instalments invoiced monthly or quarterly | Payment only after the buyer's model ships, or open-ended acceptance testing |
| Holdbacks and clawbacks | A small holdback released on objective acceptance criteria by a set date | Large holdbacks, subjective acceptance, or clawbacks for anything beyond a breach of warranty |
| De-identification | Done before transfer by a provider the seller retains and directs | Raw data sent to the buyer to "clean on our side" |
| Warranties and indemnities | Knowledge-qualified warranties, a cap tied to fees received, a short claims period | Unlimited indemnities or absolute warranties that no records are personal |
| End user and sublicensing | The end user is named or the class of end users is stated; sublicensing needs consent | Silent sublicensing to unnamed third parties |


## Exclusivity, term and scope of systems

Exclusivity is where a low price does the most damage. One law firm calls it "the buyer's most valuable request and the seller's most expensive concession", because the same data can often be licensed to several buyers who do not compete ([Latsoudis + Associates](https://www.latsoudislaw.com/data-ai-greece/data-licensing-contract-key-terms/)). If you grant it, keep it narrow. The Tempus filing shows what narrow looks like: the company agreed not to license the neurology records to other third parties only for a period ending 18 months after delivery ([Tempus Form 8-K](https://www.sec.gov/Archives/edgar/data/1717115/000119312526429533/d135568d8k.htm)). The IRS lists exclusivity, geographic limits and duration among the terms that make two licenses comparable at all ([IRS IRM 4.48.5](https://www.irs.gov/irm/part4/irm_04-048-005)).

Field of use matters as much. "Training AI models" on its own lets the buyer do almost anything; the same law firm calls it "a door left open". Settle whether the license covers pre-training, fine-tuning or evaluation, one model or all of them, internal use or models offered to others. A partner at Davis+Gilbert has noted that AI platforms are learning which additional permitted uses to write into agreements ([Davis+Gilbert](https://www.dglaw.com/digiday-wtf-is-ai-grounding-licensing-and-why-do-publishers-say-it-matters-over-training-deals/)), so read that list line by line.

Scope of systems is the third lever. An offer for "your email and chat archive" should say which mailboxes, which years and which channels, and leave out HR, legal and board material. For how each choice moves price, see [exclusive vs non-exclusive data licenses](https://origindatapartners.com/guides/exclusive-vs-non-exclusive-data-license).


## Who is responsible for de-identification?

You are, in practice, even when the buyer offers to do it. In the Spirit Airlines sale the buyer said the data would be "rigorously scrubbed of any personally identifiable information by a third party before receipt" ([Axios](https://www.axios.com/2026/08/17/google-spirit-airlines-bankruptcy)). That order of events, cleaning before the data moves, is the one to insist on.

A de-identification provider has set out rules for these transfers: the provider should be retained by the seller, take direction from the seller and its counsel, and be paid whether or not the sale closes; and the buyer should commit in writing not to re-identify anyone before the data moves ([Tonic.ai](https://www.tonic.ai/blog/de-identification-ai-data-licensing)). California's privacy law makes the same contractual point: data counts as deidentified only if the business, among other steps, "contractually obligates any recipients of the information to comply" ([Cal. Civ. Code 1798.140](https://leginfo.legislature.ca.gov/faces/codes_displaySection.xhtml?lawCode=CIV&sectionNum=1798.140)).

Push back on any offer that asks for raw exports or a live connection so the buyer can clean the data itself. The method is covered in [how to de-identify company data before licensing](https://origindatapartners.com/guides/de-identify-company-data-before-licensing).


## Payment timing, holdbacks and clawbacks

Normal payment terms are simple: a fee on signing or delivery, or fixed instalments. Tempus invoices its per-record fees monthly and its fixed oncology payments quarterly, and the larger customer may terminate only on 12 months' notice and payment of "a substantial termination fee" ([Tempus Form 8-K](https://www.sec.gov/Archives/edgar/data/1717115/000119312526429533/d135568d8k.htm)). An upfront fee, sometimes in instalments tied to delivery milestones, shifts the risk of the model's success to the buyer ([Latsoudis + Associates](https://www.latsoudislaw.com/data-ai-greece/data-licensing-contract-key-terms/)).

A holdback is part of the price paid later, after the buyer accepts the delivery. Keep it small, tie it to objective checks such as file counts, formats and a de-identification report, and set a date after which it is released automatically. A clawback lets the buyer recover money already paid. Limit it to a proven breach of an express warranty, capped at the fees received.

Revenue shares need the most care. Their value "depends on a definition of revenue the seller cannot easily verify" ([Latsoudis + Associates](https://www.latsoudislaw.com/data-ai-greece/data-licensing-contract-key-terms/)), so ask for a minimum payment, regular reports and an audit right. The structures are compared in [data licensing pricing models](https://origindatapartners.com/guides/data-licensing-pricing-models).


## Sublicensing, warranties and indemnities: what you are promising

Many offers come from intermediaries that assemble datasets for model builders, so the sublicensing clause decides where your records end up. Reporting on startup data sales found that founders "have little to no visibility into who is buying the data and what they plan to do with it" ([MIT Technology Review](https://www.technologyreview.com/2026/10/13/1145755/ai-labs-buying-data-from-dead-startups-to-train-models/)). Fix that in the contract: the license should name the end user, or at least the class of end user, forbid onward transfer of raw records, and make any sublicense subject to your written consent.

Warranties are the promises you make about the data: that you own it, that it was lawfully collected, that personal information was removed. Sellers' counsel usually limit them to the seller's knowledge after reasonable enquiry, list known exceptions in a schedule, and cap indemnities at an amount "usually linked to the fees received over a defined period" with "a short claims period" ([Latsoudis + Associates](https://www.latsoudislaw.com/data-ai-greece/data-licensing-contract-key-terms/)). Do not promise that no record anywhere contains personal information. Promise the process you followed instead.

Checking who the buyer is, how it is funded and how it has handled data before is a separate job from reading the offer. Do it in parallel with the [data licensing due diligence checklist](https://origindatapartners.com/guides/data-licensing-due-diligence); this page stays with the terms and the price.


## Compare competing offers line by line

Headline prices mislead when the terms differ. The two offers below are a hypothetical example for the same archive, laid out the way Origin compares offers for owners. Put each real offer into the same rows before you decide.

- Ask Offer A what premium it pays for exclusivity, and whether it would accept a shorter period or a narrower field.
- Ask Offer B for a minimum number of paid update years, or a fee if it cancels the updates early.
- Use the strongest clause from each offer as the starting point for the next round with both.

| Term | Offer A | Offer B | What the difference costs you |
| --- | --- | --- | --- |
| Price | $120,000 on signing | $80,000 on delivery, plus $20,000 a year for annual updates | B pays $140,000 over three years if the updates continue |
| Exclusivity | Exclusive, worldwide, three years | Non-exclusive | A blocks a second license of the same records for three years |
| Field of use | "Training AI models" | Fine-tuning and evaluation of named model types | A lets the buyer do almost anything with the records |
| Sublicensing | Silent | Only with your written consent | Under A the records can move to parties you never approved |
| Warranties and cap | Absolute warranties, no cap | Knowledge-qualified, capped at fees received | A leaves you exposed for more than you were paid |
| End of term | No deletion duty | Raw data and copies deleted, certificate on request | A leaves copies in circulation indefinitely |


## How to negotiate a data license: competing offers and next steps

A second offer changes the conversation more than any argument. Buyers price on their own research theses and "quote wildly different numbers" ([MIT Technology Review](https://www.technologyreview.com/2026/10/13/1145755/ai-labs-buying-data-from-dead-startups-to-train-models/)). In the Spirit Airlines auction the next-highest bid was $7.5 million against the winning $10 million ([Axios](https://www.axios.com/2026/08/17/google-spirit-airlines-bankruptcy)). "Deals will depend on the negotiating strength of each party," as one data lawyer put it ([Davis+Gilbert](https://www.dglaw.com/digiday-wtf-is-ai-grounding-licensing-and-why-do-publishers-say-it-matters-over-training-deals/)).

- Do not send samples before an evaluation agreement limits their size and use and requires deletion. A generous sample can deliver most of the value without a license.
- Write down what you want before the call: price floor, exclusivity limits, excluded systems and the end user.
- Ask each buyer the same questions so offers can be compared line by line.
- Trade, do not concede: exclusivity for a higher fee or minimum payments, a longer term for annual updates at a price.
- Get the offer in writing, with every clause in the table above answered, before you start preparing data.


## What to ask before signing a data license

Ten questions settle most offers. Which records and years exactly? For which uses? Exclusive for how long and in what field? Who is the end user? Who de-identifies, and is it done before transfer? When is each payment due? What can be held back or clawed back, and on what test? What do we warrant, and what is the cap? What happens to the data and to trained models at the end? Who can audit what?

If you cannot get clear answers, the offer is not ready to sign. For the full range of what companies have been paid, see [how much AI companies pay for company data](https://origindatapartners.com/guides/company-data-value).

This page is general information to take to your lawyer and accountant. It is not legal, tax or investment advice.

Origin Data Partners is an independent data partner that helps companies get their data bought. A buyer may pay Origin when a deal closes.

Not sure whether a buyer or an offer is sound? Origin checks who the buyer is, what the license allows and how the price compares with published deals before anything is signed. [Get a free estimate and a second opinion](https://origindatapartners.com/check-fit).


## Questions and answers

### Is my data licensing offer fair?

Compare the price with public deals and with about $20,458 per employee, the median in 12 deals Origin helped place. Then check exclusivity, field of use, payment timing and warranties, because those terms decide what the price actually buys.

### How do I evaluate an offer for my company's data?

List what the offer covers: records, use, term, exclusivity, price, payment dates, warranties and the end user. Anything the offer leaves out will be decided by the buyer's contract, so ask for it in writing before you prepare any data.

### What should I ask before signing a data license?

Ask which records and years are covered, for what uses, whether it is exclusive, who the end user is, who de-identifies the data and when, when each payment is due, what can be held back, and what happens to the data and trained models at the end.

### How do I negotiate a data license?

Get a second offer if you can, because buyers price very differently. Then trade terms for money: exclusivity only for a higher fee or minimum payments, and a narrow field of use unless the buyer pays for a broad one.

### Should I accept an exclusive data license?

Only if the price reflects it and the exclusivity is limited in time, field or dataset. Exclusivity stops you licensing the same records to other buyers, so it is the most expensive term you can give.

### Who should de-identify the data, me or the buyer?

The seller should control it. Retain a de-identification provider that you direct, finish the work before any data leaves, and get the buyer's written promise not to re-identify anyone.

### Is a holdback normal in a data license?

A small one can be. It should be released on objective checks such as file counts, formats and a de-identification report, by a fixed date, and not depend on the buyer's opinion of the data.

### Can the buyer keep models trained on my data after the license ends?

Usually yes. Buyers almost always ask to keep trained models, so price the license as a permanent use and require deletion of the raw data and copies at the end.

## Sources

- [Quartz: AI training data is becoming a seller's market (May 2026)](https://qz.com/ai-training-data-pricing-licensing-deals-market-052126)
- [Axios: Google wins bankruptcy auction for Spirit Airlines emails, chats, documents (Aug 2026)](https://www.axios.com/2026/08/17/google-spirit-airlines-bankruptcy)
- [Tempus AI, Form 8-K, data license agreements (SEC, Oct 2026)](https://www.sec.gov/Archives/edgar/data/1717115/000119312526429533/d135568d8k.htm)
- [Tempus Reports Third Quarter 2026 Results (BioSpace, Nov 2026)](https://www.biospace.com/press-releases/tempus-reports-third-quarter-2026-results)
- [MIT Technology Review: AI labs are buying data from dead startups (Oct 2026)](https://www.technologyreview.com/2026/10/13/1145755/ai-labs-buying-data-from-dead-startups-to-train-models/)
- [Latsoudis + Associates: data licence agreements, the terms that decide the deal (Oct 2026)](https://www.latsoudislaw.com/data-ai-greece/data-licensing-contract-key-terms/)
- [Tonic.ai: core tenets for de-identification in AI data licensing (Sep 2026)](https://www.tonic.ai/blog/de-identification-ai-data-licensing)
- [Davis+Gilbert: AI grounding licensing versus training deals (Aug 2025)](https://www.dglaw.com/digiday-wtf-is-ai-grounding-licensing-and-why-do-publishers-say-it-matters-over-training-deals/)
- [California Civil Code 1798.140 (CCPA definitions, including deidentified)](https://leginfo.legislature.ca.gov/faces/codes_displaySection.xhtml?lawCode=CIV&sectionNum=1798.140)
- [IRS Internal Revenue Manual 4.48.5: Intangible Property Valuation Guidelines](https://www.irs.gov/irm/part4/irm_04-048-005)

## Related reading

- [Data licensing agreement: what to include](https://origindatapartners.com/guides/data-licensing-agreement)
- [Exclusive vs non-exclusive data licenses: what sellers should know](https://origindatapartners.com/guides/exclusive-vs-non-exclusive-data-license)
- [Data licensing pricing models: flat fee, recurring and revenue share](https://origindatapartners.com/guides/data-licensing-pricing-models)
- [Company data licensing due diligence checklist](https://origindatapartners.com/guides/data-licensing-due-diligence)
- [What is company data worth per employee? Benchmarks from 12 deals](https://origindatapartners.com/guides/company-data-value-per-employee)
- [How much do AI companies pay for company data?](https://origindatapartners.com/guides/company-data-value)
- [AI Data Brokers: What to Check Before You Accept an Offer](https://origindatapartners.com/guides/ai-data-broker-offer-what-to-check)