TL;DR
- micro1's public Data License Master Terms (updated Oct 7 2026) take a "worldwide, royalty-free, fully paid-up, nonexclusive, irrevocable, perpetual license" that the buyer can sublicense through multiple tiers (micro1).
- Under those terms the seller can invoice only after the buyer accepts the data, and the buyer has 30 business days after delivery to reject it (micro1).
- Deletion falls due 120 days after final delivery, but nothing requires the buyer to delete or retrain models, weights or derivative works (micro1).
- Each side's liability is capped at three times the fees, except for fraud, gross negligence and wilful misconduct, and the buyer can claw back every payment if the seller's description of the data proves materially inaccurate (micro1).
- Per-unit pricing appears in public filings: Tempus licenses de-identified neurology records to Pathos at $100 per delivered record, invoiced monthly, plus $5.0 million for record updates. It is a related-party deal, since the two companies share executives and owners (Tempus 8-K, Oct 23 2026).
- Exclusivity can be a window rather than forever: Tempus agreed not to license the same neurology records to other third parties until 18 months after the later of delivery and Oct 22 2027 (Tempus 8-K).
- Terms are finite. Reddit's AI data licenses run two to three years (Reddit prospectus, SEC), and Tempus's oncology license with Bristol Myers Squibb ends Dec 31 2031, with 12 months' notice to terminate early (Tempus 8-K).
- Some clauses are required by law. In California, a contract to sell or license de-identified patient information must prohibit re-identification (Cal. Civ. Code § 1798.148(c)).
What is a data licensing agreement for AI training?
A data licensing agreement for AI training is the contract under which a company lets a model developer use a defined copy of its data, for defined purposes, for a defined time, in exchange for payment. Under a license, unlike a sale, the company keeps ownership. The agreement usually has two layers: master terms that cover every delivery, and an order form or statement of work for each dataset. Tempus's 2026 deals follow that pattern. The Bristol Myers Squibb oncology license is a statement of work under a master services agreement dated 2018, and the Pathos neurology license is an order form under a master agreement restated in 2024 (Tempus 8-K).
Most of these contracts are confidential, so this guide uses the two most detailed public sources: micro1's published Data License Master Terms, a buyer's standard form, and the Tempus 8-K of Oct 23 2026, which describes negotiated terms. A standard form shows where a negotiation starts, not where it has to end.
What should the data description schedule contain?
The schedule should describe the data precisely enough that both sides can tell whether a delivery matches it, because acceptance, pricing and every warranty point back to it. A good schedule lists:
- Sources and contents: each system or archive, and what it holds.
- Volume and span: hours, files, records or scans, and the dates covered.
- Formats and fields: resolution, frame rate, sensor channels, metadata.
- Exclusions: what is deliberately left out, such as HR files, named clients or audio.
- De-identification: what was removed or transformed, and how.
- Delivery method and schedule: in one batch, in tranches, or updated over time.
Datasheets for Datasets (Gebru et al.) gives a template: document each dataset's motivation, composition, collection process and recommended uses. That documentation also speeds up the buyer's own compliance work.
What does the grant clause decide?
The grant clause decides what the buyer may do with your data, and it is where most of the value changes hands. Read every adjective in it. micro1's grant is "worldwide, royalty-free, fully paid-up, nonexclusive, irrevocable, perpetual" with sublicensing "through multiple tiers" (micro1, s. 1.2). In plain terms:
| Word in the grant | What it means for the seller |
|---|---|
| Worldwide | No territorial limit on the buyer's use |
| Royalty-free, fully paid-up | One price, with no further payment however the data or the models are used |
| Nonexclusive | You can license the same data to others |
| Irrevocable | You cannot take the rights back, even after a dispute |
| Perpetual | The rights never expire |
| Sublicensable through multiple tiers | The buyer can pass rights to others, who can pass them on again |
A negotiated grant can be much narrower. Tempus lets Pathos use the neurology data "for its internal purposes" and to help one named partner, and a side letter confirms Tempus's consent to the sublicense to that partner (Tempus 8-K). Points to fix in the grant:
- Purpose. Training, fine-tuning, evaluation, or all three.
- Users. Internal use only, or named affiliates.
- Sublicensing. None, or only with your consent.
- Term. A fixed term, not a perpetual one.
- Exclusivity. Whether it applies, and its scope (see exclusive vs non-exclusive data license).
How are fees and payments usually structured?
AI data licenses are usually paid as a fixed fee, a per-unit fee, or a mix, and the public examples show all three:
- Fixed payments over a term. Bristol Myers Squibb pays Tempus $137.5 million in aggregate fixed payments to the end of 2031, invoiced quarterly, for a data license bundled with software access and research services (Tempus 8-K).
- Per unit delivered. Pathos pays $100 per delivered record, invoiced monthly, for up to 500,000 records over five years (Tempus 8-K).
- Fees for updates. The same order form adds $5.0 million for the availability of record updates, invoiced quarterly (Tempus 8-K).
- Multi-year contract value. Reddit disclosed data licensing arrangements with an aggregate contract value of $203.0 million and terms of two to three years (Reddit prospectus, SEC).
A per-unit price moves volume risk onto the seller; a fixed price moves it onto the buyer. Whichever you choose, the contract should say when each payment falls due and what triggers it.
How do delivery and acceptance clauses affect when you are paid?
Acceptance clauses decide when you are paid, because many buyers pay only for data they have accepted. Under micro1's terms the buyer can reject data within 30 business days of delivery, and the seller "may invoice Licensee only after Licensee has given written notice that the Data has been accepted, or such Data has been deemed accepted" (micro1, ss. 1.3 and 3).
Acceptance does not end the seller's exposure. The same terms give the buyer 60 days after acceptance to claim a latent defect or a material mismatch with the Data Manifest (micro1's name for the data schedule), and that time limit does not apply to claims about rights, privacy, consents or provenance (micro1, s. 1.3).
Protect yourself with objective acceptance criteria tied to the data schedule: volume, format, fields, and de-identification checks that can be measured. Add deemed acceptance if the buyer stays silent, and a short cure period for fixing a rejected delivery.
How do privacy and de-identification clauses work?
Privacy clauses say who is responsible for personal data in the delivery, how it was de-identified, and what the buyer must never do with it. The most important is a ban on re-identification. For de-identified patient information, California requires one: the contract must state that the data includes de-identified patient information, prohibit re-identification by the buyer, and bar onward disclosure except to parties bound by the same terms (Cal. Civ. Code § 1798.148(c)). California's general definition of "deidentified" also requires the holder to bind every recipient by contract (§ 1798.140(m)).
If any personal data leaves the EU or EEA, the agreement needs a transfer mechanism. The European Commission's standard contractual clauses, modernised on June 4 2021, are the usual one. The legal guide covers when data counts as de-identified in each region. fearofgod.ai de-identifies the data before it goes to buyers.
What warranties and indemnities will you give?
You will warrant that you had the right to collect and license the data, and you will indemnify the buyer if that proves false. micro1's terms ask the seller to warrant that it obtained "all consents, opt-ins or other valid legal bases" required by privacy law, obtained the releases needed for anyone whose likeness appears, and did not scrape or harvest the data from online sources without the owner's authorization (micro1, s. 5). The seller also warrants that it has not already granted anyone rights to the same data "on the same or substantially similar terms" (micro1, s. 5.1).
Two clauses give those promises their weight. micro1 caps each party's total liability at three times the fees under the order form, except for fraud, gross negligence and wilful misconduct (micro1, s. 7.2). And if the Data Manifest proves "inaccurate or misleading in any material respect", the buyer can terminate at once, nothing more is payable, and the seller refunds everything already paid, even after acceptance (micro1, s. 4.3). Ask for:
- Knowledge qualifiers ("to the seller's knowledge") on facts you cannot check, such as what every former employee consented to.
- A materiality bar and a cure period before a description error can unwind the whole deal.
- A clear split of responsibility: your duty covers the data as delivered, and the buyer's covers how it uses the data and what its models produce.
Buyer forms also contain protections for the seller, and they are worth keeping. micro1's terms bar the buyer from knowingly trying to re-identify anyone in de-identified data, from using the data to compete with the seller's business, and from reselling the same data as a standalone product (micro1, s. 1.4).
What happens to the data and the models when the license ends?
Usually the files are deleted and the models are kept, so read the deletion clause for what it leaves out. micro1 requires deletion 120 days after final delivery or on termination, whichever comes first, with written confirmation at the seller's request, and adds that this does "not require Licensee to delete or retrain any model, weights, or derivative works created from Data" (micro1, s. 4.4). A license for a fixed term does not stop a model trained during that term from being used after it.
Points to negotiate:
- Certification. A written confirmation once deletion is done, as of right rather than on request.
- Backups. Whether copies in backups are covered, and by when.
- Surviving rights. Whether models trained on your data may still be used, sold or sublicensed after the term, and whether that is reflected in the price.
Raw data is deleted according to a schedule agreed with the partner.
Which confidentiality, term and governing-law clauses matter?
These clauses decide who knows about the deal, how long it lasts, and where disputes go.
- Confidentiality of the seller. micro1's terms say the licensee "shall not disclose Licensor's identity or the source of the Data without Licensor's prior written consent" (micro1, s. 2.3). Ask for the same wherever your name could harm your relationships with staff or clients.
- Term and termination. Tempus's oncology license runs to December 31 2031, and the buyer can end it early on 12 months' notice and payment of a substantial termination fee (Tempus 8-K).
- Governing law and disputes. micro1's terms choose the laws of California and the United States, with arbitration in California (micro1, ss. 9.3 and 9.4). A seller elsewhere should know in advance where a dispute would be heard.
Which clauses should a seller negotiate first?
Negotiate the grant, acceptance, deletion and liability clauses first, because they decide how much of the data's value you keep. This checklist covers the full agreement:
| Clause | What to ask for |
|---|---|
| Data schedule | Precise sources, volume, span, formats, exclusions and de-identification |
| Grant | Named purposes, no open-ended sublicensing, a fixed term |
| Exclusivity | None, or limited by field, records and time, and priced as sales you give up |
| Fees | Payment triggers and due dates; per-unit or fixed, chosen on purpose |
| Acceptance | Objective criteria, deemed acceptance, a cure period |
| Privacy | A ban on re-identification and onward disclosure; transfer clauses where needed |
| Warranties | Knowledge qualifiers on facts you cannot verify |
| Liability | A cap that covers the indemnity, and limits on clawback for description errors |
| Deletion | A deadline, certification, backups, and an answer on models |
| Confidentiality | Your name kept private unless you agree |
| Term and law | A fixed term, a termination fee, a forum you can reach |
The commercial context is in how to sell data to AI companies. What a schedule should list for each kind of data is on the data-type pages, such as 3D scans and CAD and industrial sensor and telemetry data, all on the sell hub. If you hold data you want to license, talk to our team.
This guide is general information, not legal advice. Have counsel review any agreement before you sign.