What a data licence turns on.

1. 1

   Licensed purpose and field of use

   What buyers ask

   A licence to train, fine-tune and evaluate machine learning models, sometimes also retrieval indexing; labs prefer the broadest purpose and no limit on which models.

   Where a holder should stand

   Name the permitted uses (pretraining, fine-tuning, evaluation, retrieval) and exclude what is not paid for, such as redistribution of the raw data, retrieval display of verbatim content, or use outside a field.

   What the record shows

   Templates bracket the purpose explicitly; NVIDIA's own published data agreement limits use to internal training.

   3 sources, with the words read

   • Terms.Law: AI training data licensing, a usable agreement(undated, read 2026-10-11)

     “the licensee may use the licensed data to train, fine-tune, and evaluate machine learning models”

   • ezel.ai AI Training Data License Agreement template, section 2(read 2026-10-11)

     “for [PRETRAINING / FINE-TUNING / EVALUATION / RETRIEVAL INDEXING / OTHER SPECIFIC USE]”

   • NVIDIA Data Agreement for Model Training (2025), via ScanCode LicenseDB(2025)

     “NVIDIA makes the Datasets available to Company solely for the purpose of internal training of Company AI Solutions”

2. 2

   Survival of trained models after the licence ends (perpetual model rights)

   What buyers ask

   A perpetual, worldwide right to keep and use models already trained, even after the data licence term ends or the copy is deleted, because models cannot practically be untrained.

   Where a holder should stand

   Price it: perpetual model rights are the default ask and should cost more than a term licence. Any survival must be written in; silence grants none. In health and personal-data contexts limit post-term rights to de-identified data.

   What the record shows

   Practitioner memo recommends a perpetual licence for training and use of trained models; the published template says continuing rights must be express; Foley warns vendors claim trained models are no longer tied to the data.

   4 sources, with the words read

   • Terms.Law: AI training data licensing(undated)

     “a perpetual, worldwide license for training and use of models trained”

   • ezel.ai template, section 8(read 2026-10-11)

     “Any continuing right must be stated expressly; silence grants none.”

   • Foley & Lardner: AI Contracts in Health Care(2025-06-25)

     “A common trap is allowing the vendor to claim that once data is used to train a model, the model is no longer tied to the underlying data.”

   • NVIDIA Data Agreement for Model Training (2025)(2025)

     “Company's AI Solutions, or outputs resulting from Company's AI Solutions, developed by Company using the Datasets ... are and will remain the sole and exclusive property of Company.”

3. 3

   Term

   What buyers ask

   Disclosed platform and publisher deals run two to three years, often with recurring payments; one-off archive sales (defunct companies) are effectively perpetual transfers.

   Where a holder should stand

   For a living company, a two to three year term with renewal and refresh rights keeps the holder's leverage; for a one-off archive, price the sale as perpetual.

   What the record shows

   Reddit's S-1 (February 2024, older) disclosed $203.0 million of licensing contracts with two to three year terms; HarperCollins's offer to authors was reported as a three-year licence; Troveo says most reported deals are multi-year.

   3 sources, with the words read

   • Reddit, Inc. Form S-1(2024-02-22 (older))

     “In January 2024, we entered into certain data licensing arrangements with an aggregate contract value of $203.0 million and terms ranging from two to three years.”

   • Troveo: AI data licensing(2026-07-23)

     “Most reported deals run multi-year terms: News Corp and OpenAI at five years, Reddit's disclosed contracts at two to three.”

   • eWeek: HarperCollins book catalog to train Microsoft AI models for the next 3 years(2024-12-07 (older, reported))

     “Microsoft signed a three-year deal with Harper Collins to train an as-yet-unnamed AI model on the major publisher’s catalog. ... In addition, the deal only applies to select nonfiction books that were previously published, not fiction books.”

4. 4

   Exclusivity and exclusivity windows

   What buyers ask

   Frontier labs ask for exclusivity to keep rivals from the same data; platforms selling public archives (Reddit, Stack Overflow) license non-exclusively.

   Where a holder should stand

   Offer exclusivity only for a defined field and a defined window, and charge a multiple for it; disclose any platform vendor that already trained on the content (see checklist item vendor-already-trained).

   What the record shows

   RL environment founders put the exclusivity premium at roughly 4 to 5 times; Troveo reports 95 percent of its licensors signed exclusively; the published template offers non-exclusive or exclusive within a field. Not in the public record: No public source states typical exclusivity window lengths in months for corporate data deals.

   3 sources, with the words read

   • Epoch AI Gradient Updates: The state of RL environments (Denain, Barber)(2026-01-12)

     “Two RL environment founders independently agreed that exclusive deals are roughly 4-5x more expensive than non-exclusive ones”

   • Troveo: AI data licensing(2026-07-23)

     “95 percent of them exclusively”

   • ezel.ai template, section 2(read 2026-10-11)

     “For the term and subject to payment, Licensor grants Licensee a [NONEXCLUSIVE / EXCLUSIVE WITHIN THE DEFINED FIELD] license.”

5. 5

   Price structure: fixed fee, per-unit, variable and revenue share

   What buyers ask

   Disclosed deals mix a fixed annual minimum with a variable component, per-item fees, or aggregate multi-year contract values.

   Where a holder should stand

   Anchor on a fixed minimum plus refresh or usage fees; per-unit pricing (per title, per hour, per record) makes consent and payouts traceable to individual contributors.

   What the record shows

   OpenAI pays Dotdash Meredith at least about $16 million a year as the fixed component, with a variable component on top (IAC disclosures, as reported); HarperCollins offered $5,000 per title split 50-50 with authors; Reddit reported $66.4 million minimum revenue for 2024 on $203.0 million of contracts.

   4 sources, with the words read

   • Adweek: OpenAI is paying Dotdash Meredith at least $16 million to license its content(2024-11 (older, from a search excerpt))

     “a minimum of around $16 million per year to license its content, according to public financial documents from IAC”

   • Authors Guild: HarperCollins AI licensing deal(2024-11-19 (older))

     “provides for a $5,000 fee per title, which will be split 50-50”

   • Reddit, Inc. Form S-1(2024-02-22 (older))

     “We expect a minimum of $66.4 million of revenue to be recognized during the year ending December 31, 2024 and the remaining thereafter.”

   • Troveo: AI data licensing(2026-07-23)

     “Revenue share Rights holders paid from model revenue”

6. 6

   Holder warranties: title, lawful collection, non-infringement, privacy

   What buyers ask

   That the licensor owns or has the right to license the data, collected it lawfully, and that training on it does not infringe third-party rights; labs add privacy and de-identification warranties.

   Where a holder should stand

   Holders commonly give title and authority without qualification and qualify the rest by knowledge, tied to a disclosed rights inventory (schedule), so the warranty is that the disclosed limits are accurate rather than that the data is perfect.

   What the record shows

   Practitioner memo lists the three core warranties; the published template uses a knowledge-qualified warranty tied to a rights schedule and disclaims model-performance warranties.

   2 sources, with the words read

   • Terms.Law: AI training data licensing(undated)

     “That the data was collected in compliance with applicable law”

   • ezel.ai template, section 6(read 2026-10-11)

     “represents that the rights and source limitations it has disclosed in Schedule A are accurate to its knowledge”

7. 7

   Representations on consent and opt-outs

   What buyers ask

   A warranty that no opt-out has been invoked and that later opt-outs will be honoured; for creator content, proof that each rights holder opted in.

   Where a holder should stand

   Warrant to knowledge, keep a consent ledger per person or per work, and agree a process (not a guarantee) for later opt-outs; exclude people who did not opt in rather than promising consent you cannot show.

   What the record shows

   Practitioner memo drafts a knowledge-qualified no-opt-out warranty plus commercially reasonable efforts on later opt-outs; HarperCollins sought express permission from each author.

   2 sources, with the words read

   • Terms.Law: AI training data licensing(undated)

     “the licensor warrants that, to its knowledge, no opt-out has been invoked”

   • Authors Guild: HarperCollins AI licensing deal(2024-11-19 (older))

     “HarperCollins is requesting express permission from each author individually”

8. 8

   Indemnities and liability caps

   What buyers ask

   Licensor indemnity for third-party claims that the data infringes or was unlawfully collected, often outside or above the general cap.

   Where a holder should stand

   Limit indemnity to precisely defined claim types, cap it (a multiple of fees), and carve out claims caused by the buyer's own use, combination with other data, or outputs. Note the reverse also exists: NVIDIA's own data agreement makes the recipient indemnify the provider.

   What the record shows

   Practitioner memo drafts a licensor indemnity for third-party claims about the data; the published template makes indemnity optional and leaves the cap to negotiation; NVIDIA's published terms run the indemnity from recipient to provider. Not in the public record: No public source read gives typical indemnity cap levels in AI data deals (a 2023 Cozen O'Connor presentation on super-caps was not re-verified).

   3 sources, with the words read

   • Terms.Law: AI training data licensing(undated)

     “the licensor indemnifies the licensee for any third-party claim alleging that the licensed data”

   • ezel.ai template, section 7(read 2026-10-11)

     “IF INDEMNITY IS AGREED: Each party will defend and indemnify the other against [PRECISE THIRD-PARTY CLAIM TYPES].”

   • NVIDIA Data Agreement for Model Training (2025)(2025)

     “Company will indemnify and hold harmless NVIDIA from and against any claim by any third party ... arising out of or related to your use or distribution of the Datasets.”

9. 9

   Return or deletion of the data copy, with certification

   What buyers ask

   Buyers accept deleting raw data at term end if trained models survive; providers' own forms require deletion within days.

   Where a holder should stand

   Require deletion of the dataset (not the models, unless agreed) within a fixed number of days, with a written certificate, and deletion on breach.

   What the record shows

   NVIDIA's form: stop using and delete within 14 days of termination; the published template: return or delete within a set number of days and certify on request; Foley recommends a data return or destruction clause.

   3 sources, with the words read

   • NVIDIA Data Agreement for Model Training (2025)(2025)

     “within fourteen (14) days, Company will stop using the Datasets and delete and destroy copies of Datasets”

   • ezel.ai template, section 8(read 2026-10-11)

     “return or delete copies within [NUMBER] days ... certify completion to Licensor on request”

   • Foley & Lardner: AI Contracts in Health Care(2025-06-25)

     “Otherwise, build in a data return or destruction clause that outlines how and when data must be returned or destroyed.”

10. 10

    Records, audit and certification

    What buyers ask

    Buyers resist audit of their training pipelines; holders get, at most, record-keeping and certification.

    Where a holder should stand

    Ask for records sufficient to show compliance with the licensed uses, an annual compliance certificate, and audit only on reasonable suspicion of breach. Holders should expect the mirror request: buyers want certification of de-identification before each transfer.

    What the record shows

    The published template has a record-keeping duty but no audit right; the Spirit Airlines ombudsman required a de-identification certification filed with the court before every transfer.

    2 sources, with the words read

    • ezel.ai template, section 4(read 2026-10-11)

      “keep records reasonably sufficient to show compliance with the licensed uses and restrictions”

    • The Next Web: Spirit Airlines data and the ombudsman(2026-10-09)

      “The agent must file a certification with the court before every transfer.”

11. 11

    No re-identification, no onward transfer, flow-down

    What buyers ask

    Labs accept no-re-identification and resale bans when the holder insists; in regulated settings they are required.

    Where a holder should stand

    In every licence we take part in: the buyer will not attempt re-identification, will not resell or sublicense the raw data, and will bind any permitted recipient to the same terms; for foreign buyers add the DOJ onward-brokerage ban and 14-day violation notice where bulk thresholds could be met.

    What the record shows

    CCPA deidentified status depends on contractual flow-down; the Spirit ombudsman required a public no-re-identification commitment and flow-down; 28 CFR 202.302 requires the onward-transfer clause for foreign buyers.

    2 sources, with the words read

    • California Civil Code 1798.140(m)(3)(posted January 2025)

      “Contractually obligates any recipients of the information to comply with all provisions of this subdivision.”

    • 28 CFR 202.302(a)(read 2026-10-11)

      “Contractually requires that the foreign person refrain from engaging in a subsequent covered data transaction involving data brokerage of the same data with a country of concern or covered person”

12. 12

    Output and memorisation limits

    What buyers ask

    Publishers have obtained caps on verbatim reproduction in model outputs; labs accept them for creator content.

    Where a holder should stand

    For text, code or media where verbatim regurgitation would leak confidential or personal content, require output filters and a cap on verbatim reproduction.

    What the record shows

    HarperCollins's terms limit outputs to 200 consecutive words or 5 percent of a book; Venable notes training data is typically memorised and can be reproduced.

    2 sources, with the words read

    • Authors Guild: HarperCollins AI licensing deal(2024-11-19 (older))

      “limiting outputs to no more than 200 consecutive words and/or 5 percent of a book's text”

    • Venable: Contracting for AI Model Training(2026-06-05)

      “Training data is typically "memorized" by the model, which can then reproduce similar or identical content when prompted.”

13. 13

    Delivery format, refresh and performance obligations

    What buyers ask

    For recurring licences, buyers set delivery and API performance requirements and renewal depends on them.

    Where a holder should stand

    Specify format, schedule and acceptance; avoid service levels the holder cannot meet; tie refresh fees to new data.

    What the record shows

    Reddit's S-1 risk factors note its data licences carry API performance requirements and may not be renewed.

    1 source, with the words read

    • Reddit, Inc. Form S-1, risk factors(2024-02-22 (older))

      “Our data license agreements are subject to terms and conditions including API performance requirements, that we may be unable to meet.”

14. 14

    Who controls de-identification, and use restrictions

    What buyers ask

    Some buyers offer to run de-identification themselves after receiving raw data.

    Where a holder should stand

    The holder selects or jointly selects the de-identification provider, approves the protocol and requires certification before anything leaves; the licence restricts buyer use and allocates privacy risk.

    What the record shows

    Frankfurt Kurnit's October 2026 sell-side checklist.

    2 sources, with the words read

    • Frankfurt Kurnit (Goldberg), Legal 500(2026-10-06)

      “Select (or jointly select) the de-identification provider, approve the protocol, and require certification to the company's satisfaction.”

    • Frankfurt Kurnit (Goldberg), Legal 500(2026-10-06)

      “The agreement should restrict how the buyer can use the data, allocate privacy risk, and address what happens if something goes wrong.”

15. 15

    Disclosed deal example: bankruptcy sale with exclusivity and privacy conditions (Spirit Airlines to Google)

    What buyers ask

    In the Spirit Airlines bankruptcy, Google won an August 2026 auction for de-identified company data (including corporate email) at a reported $10 million; transactional customer records were excluded.

    Where a holder should stand

    Shows the shape proposed to the court, which had not approved it when read: carve out consumer transaction data, independent scrubbing (Tonic.ai) with certification before each transfer, a public no-re-identification commitment and flow-down.

    What the record shows

    Ombudsman report filed 5 October 2026; the approval hearing set for 14 October 2026 was adjourned on 9 October to 6 November 2026 (Stretto docket summary, 14 October), so the sale was still not approved.

    1 source, with the words read

    • The Next Web: Spirit Airlines data and the ombudsman(2026-10-09)

      “The potential privacy losses or costs to consumers have been eliminated or mitigated.”

Research from public sources, read on the day shown. It is not legal advice: a sale needs a lawyer who has read the records and the contracts.

“Read first” marks a clause that costs a holder the most, such as a licence for ever with the right to pass the data on. “Not published” means we found nothing on it in the buyer's public pages: ask for it in writing.