Per-model vs enterprise-wide training licenses

This page is general information, not legal advice. Confirm requirements with counsel for your jurisdiction and use case.

How the two license structures differ in practice

The two structures differ mainly in what triggers a new fee and who inside your company may touch the data. Licensing of training material is still a developing commercial practice, and the U.S. Copyright Office's Part 3 report, still a pre-publication version as of October 2026, describes voluntary licensing as one path the market is taking [1]. That means there is no standard "number of models" clause; you are drafting it.

A per-model grant usually names a model identifier, a parameter-count band or a product, and treats any other training use as out of scope. An enterprise-wide grant names a licensee group (parent plus defined affiliates), a field of use and a term, and leaves the count of models open. Some deals sit between the two: a capped number of "Licensed Models" per year, or a named model family plus a fixed number of additional slots.

Illustrative example: invented to show structure; it does not describe an available dataset.

Defining a "model" and a "training run" for counting

The counting definition is the clause that decides the economics, so write it before you negotiate price. A useful definition counts a "Licensed Model" as a distinct set of weights released to production or to a customer, not every intermediate checkpoint saved during training. Without that, a single pre-training run that writes thousands of checkpoints could be read as thousands of models.

Model documentation gives you a workable unit. Model cards propose recording intended use, training data and evaluation for each trained model [4], and your license can require a registry entry per Licensed Model that references the licensed dataset ID. That registry becomes your audit evidence rather than a new reporting burden.

Watch the edge cases that break naive definitions:

• Adapters. LoRA and similar low-rank adapters are trained separately from base weights and behave differently from full fine-tunes [5]. Decide whether each adapter is a new model or a variant of its base.
• Distillation. A student model trained on a teacher's outputs may never see the licensed records directly. Say whether it counts.
• Merges and continued pre-training. Merged checkpoints and continued pre-training on a new mix can be argued either way.
• Per-run pricing. "Per training run" licenses sound precise but collapse on restarts, resumed jobs and failed runs. Count completed, released models instead.

Successor versions of the same model are a separate question, covered in our guide to derivative and successor model rights; this page is about breadth across different models and teams.

Research experiments and ablations that create many small models

Ablations are where per-model licenses most often fail, because research creates far more models than products do. Generalization error scales predictably with training-set size [6], so teams often train many small proxy models to forecast large-run results before committing compute. Data-mix ablations, tokenizer tests and contamination checks can each produce dozens of throwaway models.

Small licensed datasets also carry weight in fine-tuning. LIMA fine-tuned a 65B-parameter model on 1,000 curated examples [7], so a modest SFT set can sit behind many experimental variants across product lines. A per-model license that counts each experiment makes research unaffordable; one that is silent invites a dispute.

Negotiate an explicit research carve-out: models trained solely for internal evaluation, never deployed or shared outside the licensee, with weights deleted or quarantined after a set period, do not count toward the cap. Pair it with audit and usage-reporting rights the supplier can live with, such as an annual list of deployed Licensed Models.

Cost predictability: per-model fees vs an enterprise flat fee

Per-model fees are cheaper when the roadmap is narrow and expensive when it widens, while an enterprise flat fee costs more on day one and stays fixed as usage grows. Per-unit pricing is common in adjacent markets; book publishing deals, for example, have been priced per title [3]. Applied to models, per-unit pricing makes every new product launch a renegotiation.

Model the break-even before you choose. If your plan includes a general model, two domain fine-tunes and a distilled on-device version, a per-model structure means four fees and four approval cycles. For a fuller comparison of flat-fee, per-record, per-token and subscription models, see AI data license pricing structures compared.

Illustrative example: invented to show structure; it does not describe an available dataset.

Break-even worksheet (fill with your own quotes)

1. Per-model fee quoted: P
2. Enterprise flat fee quoted for the same term: E
3. Models you expect to release on this data during the term: N (include distilled and on-device variants)
4. Probability-weighted extra models from new product lines: M
5. Choose enterprise if E < P × (N + M), plus the internal cost of renegotiating each added model.

Affiliates, contractors and acquisitions under enterprise-wide grants

An enterprise-wide grant is only as wide as its definition of the licensee group, so define affiliates by control and fix the test date. Sample data license clauses commonly restrict use to internal purposes while permitting disclosure to contractors and authorized affiliates [2]. Copy that pattern deliberately, not by default.

Points to settle in the definitions section:

• Affiliate test. Control by majority voting interest, measured continuously, so a divested subsidiary loses rights on the day it leaves.
• Acquisitions. Whether a company you acquire mid-term joins the grant automatically, joins for an uplift fee, or stays out. Suppliers may resist automatic inclusion of a large acquiree.
• Being acquired. Whether the grant survives a change of control and whether the acquirer's other models may then train on the data. Expect a termination or consent right.
• Contractors. Labeling vendors and cloud training partners may process data only for the licensee's benefit, under confidentiality terms no weaker than yours.
• Joint ventures. Minority-owned entities usually fall outside "Affiliate"; name them if they need access.

Teams that want strict containment can combine enterprise breadth with an internal-use-only restriction, and limit sectors through field-of-use restrictions.

Choosing the structure for your roadmap

Choose per-model when one product justifies the purchase and the roadmap beyond it is uncertain; choose enterprise-wide when three or more teams will plausibly train on the same data within the term. A capped multi-model license is the practical middle for most multi-product companies.

Illustrative example: invented to show structure; it does not describe an available dataset.

Scope request template to send with your inquiry

Licensee group: Parent Co. and affiliates under >50% voting control (continuous test)
Licensed Models: up to 3 production models per contract year; registry entry each
Counting unit: distinct weights released to production or customers
Not counted: internal research models, never deployed, deleted within 90 days
Adapters/distillation: [counts as variant of base | counts as new model]
Field of use: [pre-training | SFT | evaluation] for [product lines]
Acquisitions: acquiree joins on written notice + uplift fee
Reporting: annual list of Licensed Models and dataset IDs used

Before signing, check how the license treats deletion, exclusivity and term, which our owner guide on AI data license terms explains. The cluster hub on AI training data licensing covers the full set of rights and pricing questions, and the AI data hub maps every buyer guide.

How SourceX supports multi-model licensing requests

SourceX sources operational datasets from US companies on request and manages the commercial process, including licensing agreements and ongoing purchases. Pricing and allowed uses are agreed in a license defining records, uses, term and delivery, and nothing is contracted until the supplying company agrees. If your team needs data for several models or product lines, describe the data and the scope you need.

Scope your per-model or enterprise-wide license

Describe the data you need and the models or teams it should cover, and SourceX will look for US businesses that hold it. Every dataset is rights-reviewed and delivered under a license, and a request does not guarantee a match. Start a buyer request at SourceX.