Foundation model licensing — price the compute, not the access.

Foundation model contracts price compute, not access.

Enterprise contracts for foundation models — OpenAI's GPT and o-series, Anthropic's Claude family, Google's Gemini, Mistral, Cohere and the major open-weight commercial offerings — diverge sharply from traditional software licensing. There is no per-seat metric, no perpetual licence, and no support tier matrix. Pricing is structured around token consumption (input plus output), provisioned throughput units (reserved capacity), or fine-tuned model hosting. Each pricing model has a different cost profile under enterprise scale, and the wrong model can cost a buyer 3–5× nominal list price over a three-year horizon.

In our experience advising on AI procurement, the most common mistake is signing a token-priced contract when the workload pattern argues for provisioned throughput, or vice versa. The vendor sales motion routinely defaults to whichever model is most profitable at the customer's expected volume — which is rarely the model that minimises buyer cost. Getting the metric right at signature is the foundation of any serious software license optimization effort on AI spend.

Token, throughput, hosted — when each makes sense.

• Token-metered (pay-as-you-go): priced per million input and output tokens, billed in arrears. Best for variable, exploratory workloads with unpredictable volume. Worst for production workloads that run continuously — the per-token rate stacks faster than buyers model.
• Provisioned Throughput Units: reserved capacity priced per unit per hour, regardless of actual token consumption. Best for high-volume production workloads with predictable load. Worst for spiky workloads — idle capacity is unrecoverable cost.
• Hosted fine-tuned models: the buyer fine-tunes a base model on private data, and the vendor hosts the resulting model with dedicated infrastructure. Pricing includes both hosting and inference, typically at a premium per token. Best for regulated workloads or those requiring custom behaviour that prompting cannot achieve.

The breakeven math

The crossover from token-metered to provisioned throughput typically sits between 60% and 75% capacity utilisation, depending on the model and vendor. Below that, pay-as-you-go is cheaper. Above it, reserved capacity wins by 20–40%. Buyers should model the crossover for their specific workload before committing to either, and should include the right to switch — without renegotiation — at defined intervals.

What the vendor does with your inputs is contract-level, not policy-level.

Every enterprise foundation-model contract should specify, in the agreement itself, what the vendor may and may not do with customer inputs. The vendor's published privacy policy is not adequate — those policies are unilaterally amendable. The contract should explicitly state that customer inputs and outputs are not used to train or fine-tune the vendor's foundation models, are not retained beyond stated operational windows, and are not shared with third parties. Most enterprise contracts now offer this language, but it is rarely included in the base order form — it must be requested.

For regulated workloads, the data-residency commitment matters equally: which region the data is processed in, whether sub-processors have access, and what audit rights the buyer has over the vendor's data handling. The major hyperscaler-hosted foundation model offerings (Azure OpenAI, Bedrock, Vertex) inherit hyperscaler data residency commitments, which is a material reason regulated buyers default to them.

Who covers third-party IP claims on outputs.

All major enterprise foundation-model providers now offer some form of IP indemnification for outputs — Microsoft Copilot Copyright Commitment, Google Generative AI Indemnification, Anthropic and OpenAI enterprise variants. The protections differ materially. Some indemnify only when the customer uses content filters as published; some exclude outputs subsequently modified outside the platform; some cap liability at fees paid. In every case the indemnification must be referenced in the master agreement, with carve-outs enumerated and defence-and-hold-harmless commitments made explicit.

Common questions about foundation-model licensing.