Perpetual vs Term Licenses for AI Training Data

A license for training data grants rights to use that data for some scope of purpose (research, commercial training, redistribution) for some duration. Duration is the axis this guide is about. A perpetual license has no end date: once granted, the rights do not expire and do not need to be renewed. A term license grants the same category of rights — commercial use, for example — but only for a defined window, such as one year or the length of a specific contract, after which the license lapses unless it is renewed.

Duration is a separate question from scope. A license can be commercial and still be term-limited, and a license can be research-only and still be perpetual. The research-only versus commercial distinction (covered in /guides/research-only-vs-commercial-dataset-licenses) determines what the data can be used for; the perpetual-versus-term distinction determines how long that permission lasts. Both questions have to be answered independently before a training program is built on any dataset.

This page explains how license duration typically works for training data. It is general information, not legal advice — have counsel review the specific contract before relying on it.

This is the question that makes license duration consequential rather than academic. When a term license expires, what happens next depends entirely on what the specific contract says — and contracts vary. Some term licenses only restrict continued ingestion of the raw data after expiration, leaving any model already trained on it unaffected. Others extend the restriction to the trained model itself, meaning a model built during the license term can no longer be used, sold, or deployed once the underlying data license lapses, even though the model itself is a separate artifact from the data. A buyer who has not read this clause specifically does not actually know which situation they are in.

This is why the research-only-vs-commercial guide on this site treats a trained model as a potential "restricted use" in its own right (/guides/research-only-vs-commercial-dataset-licenses) — the same reasoning applies to term limits. A model is a derivative work of the data it was trained on, and a license that only covers the data for a fixed window may not say anything explicit about what happens to that derivative once the window closes. The only reliable way to know is to read the renewal, termination, and derivative-works clauses in the specific agreement, not to assume either outcome.

A model trained on data whose license has since expired, with no clear derivative-works carryover, is a legal risk that surfaces after deployment — when the model is already shipped. See /guides/how-to-license-training-data-for-commercial-models for the pre-training checklist this is meant to prevent.

Yes. A perpetual license and buyer-owned data solve overlapping but distinct problems. A perpetual license removes the expiration-date risk: the rights granted do not lapse. But a license — even a perpetual one — is still a grant of permission from whoever holds the underlying rights; the provider retains the ability to define the scope of that grant, and in principle a license can still be challenged, limited by indemnification caps, or otherwise constrained by the specific contract language, even if it never technically "expires" on the calendar.

Buyer-owned data goes one step further, as explained on /guides/buyer-owned-training-data: it means the buyer holds a perpetual, irrevocable commercial license to the delivered footage, backed by a documented consent chain and a data card, rather than a revocable subscription to a dataset the provider still controls. "Buyer-owned" does not mean copyright is transferred to the buyer — it means the rights granted cannot expire or be revoked once delivered. In practice, a buyer-owned, perpetual, irrevocable license is the strongest position a data buyer can be in on the duration axis: no renewal, no revocation, and no term to track.

A model’s training data underpins every later stage built on top of it — mid-training, post-training, fine-tuning, and eventually a shipped product. If the base layer carries a term limit or a revocable grant, that uncertainty propagates through every downstream investment made on top of it. A lab that has already spent months or years building mid-training and post-training stages on top of a pretraining corpus does not want to discover, close to shipping, that the foundation layer’s license has lapsed or could be revoked.

This is the same reasoning covered in /guides/licensing-pretraining-data-for-commercial-robot-models: pretraining specifically raises the derivative-work question early, because it is usually the first and largest training stage. A perpetual, buyer-owned commercial license removes that uncertainty at the foundation layer, which is why labs planning multi-year model programs weight license duration and ownership as heavily as they weight dataset scale or quality.

Exclusivity and duration are two separate contract terms that are often negotiated together but answer different questions. Duration (perpetual vs term) asks how long the buyer’s rights last. Exclusivity asks whether the same data can also be licensed to someone else — including a competitor. A dataset can be perpetual but non-exclusive (the default grant described on /provenance), meaning the buyer’s rights never expire, but the same footage could also be delivered to other buyers.

/guides/keeping-custom-training-data-exclusive covers how exclusivity is handled for custom-collected data specifically. The two terms are independent: a buyer can have a perpetual but non-exclusive grant, an exclusive but term-limited grant (unusual, but contractually possible), or — the combination that removes the most uncertainty — a perpetual, irrevocable, buyer-owned, exclusive grant. Which combination matters for a given program depends on whether a competitor training on the same footage is an acceptable risk.

Consent and license duration are linked in a specific way: a participant’s consent to be recorded is typically what the data license is actually grounded in. If a participant only consented to a time-limited use, a license covering that footage cannot legitimately be perpetual, no matter what the contract between the data provider and the buyer says — the consent is the ceiling. This is why a documented consent chain, not just a license agreement, is what makes a perpetual grant actually sound.

Provenance documentation is how a buyer verifies this rather than taking it on faith. A data card that records the consent basis and the jurisdictions of capture — the pattern described on /provenance and /guides/dataset-provenance-explained — is what lets a buyer confirm that the duration promised in the contract is actually backed by consent that supports it, rather than inheriting a liability that surfaces only after deployment.

• Is the license perpetual, or does it have a term, renewal date, or expiration clause?
• If it is term-limited, what happens to a model trained during the term once the term ends?
• Is the grant revocable, and under what conditions could the vendor revoke it?
• Is the data buyer-owned, or is the buyer licensing data the vendor still controls?
• Is the grant exclusive, or could the same footage be licensed to other buyers, including competitors?
• Does a documented, per-participant consent chain actually support the duration the contract promises?

Firsthand delivers custom-collected data under a perpetual, irrevocable, buyer-owned commercial license, backed by a documented, per-participant consent chain and a data card recording the jurisdictions of capture (see /provenance). The default grant is non-exclusive unless exclusivity is negotiated as part of the custom-collection spec (/guides/keeping-custom-training-data-exclusive). There is no renewal date to track and no term-expiration clause that could affect a model trained on the delivered batch, because the rights granted do not expire or get revoked once delivery is complete.

This describes the terms Firsthand states on /provenance and /terms as of this review. Always confirm the current contract language for a specific engagement, and have counsel review it before relying on it.