Enterprise generative AI procurement in 2026 looks nothing like it did in 2023. The vendor landscape has consolidated to a small set of foundation-model providers, the pricing models have stabilised around tokens and consumption commits, and the contract templates have matured to the point where redlines are routine. What has not changed is that most enterprises are still buying generative AI like it is a software product - one-off purchases, single-vendor commitments, narrow use cases. The buyers getting the most value treat it as a platform decision, with a multi-vendor strategy, a portfolio of use cases, and a procurement framework that scales with adoption.
A generative AI procurement framework that holds up across a portfolio of use cases is built on five decisions, in this order: model strategy (which foundation models, in what configuration), deployment topology (shared inference, dedicated capacity, private deployment), commercial structure (per-token, consumption commits, ELA), governance (who buys, who approves, who tracks), and exit posture (portability, data deletion, model deprecation). The sequence matters because decisions made earlier in the sequence constrain decisions made later. Buyers who start with commercial structure - "let's negotiate a discount on our OpenAI commit" - usually end up with a deal that does not survive the model strategy they will adopt 18 months later.
The model-strategy question has three answers: single-vendor (one foundation-model provider for all use cases), best-of-breed (different models for different use cases, often via a model router), or open-source-first (Llama, Mistral or Qwen as the default, with proprietary models for specific use cases). Single-vendor is operationally simplest and commercially weakest - the vendor has full leverage at renewal. Best-of-breed is operationally complex but commercially strong - every renewal benchmarks against a real alternative. Open-source-first is operationally heavy (you run the inference) but commercially strongest - the cost floor is the cost of compute, not the price of an API.
In our experience across 340+ engagements, the buyers getting the best commercial outcomes are running a best-of-breed strategy with two or three vendors in active production and at least one open-source model deployed for low-sensitivity, high-volume use cases. The capital efficiency on the open-source workload subsidises the premium on the proprietary workload.
Shared inference is the default - the vendor's standard API. Dedicated capacity reserves GPU instances for your workload, with predictable throughput and (usually) more permissive data terms. Private deployment runs the model in your cloud account, often via Bedrock, Azure OpenAI dedicated capacity or Vertex. Each topology has a different cost profile - dedicated capacity is 2-4x shared at equivalent throughput; private deployment is 4-8x at equivalent throughput - and a different rights profile. The decision should be driven by the regulatory sensitivity of the workload, not by the procurement team's preference.
The cost and rights differences are larger than most buyers realise. We have benchmark data across 200+ deals.
Three commercial structures dominate in 2026. Per-token pay-as-you-go is the entry-level structure: no commitment, no discount. Consumption commits are the enterprise standard: an annual dollar commit, a discount of 15-35% off list, and overage charged at the discounted rate. Enterprise Licensing Agreements (ELAs) are the largest structure: a multi-year commit with a substantial discount (often 40-60%), platform features (admin, SSO, audit logs, dedicated support), and sometimes capacity guarantees. The right structure depends on workload predictability. For predictable workloads above $500k annual, an ELA is usually the right answer. For workloads with high variance, consumption commits with overage protection are more efficient. See our usage-based pricing article for the deeper analysis.
Governance is what scales the framework from one deal to fifty. The components: a central buying authority that signs all AI contracts (no shadow procurement); a use-case intake process that documents the business purpose and data sensitivity for every new deployment; a model registry that tracks which models are in production and what data they have access to; and a renewal calendar with sufficient lead time to benchmark and re-negotiate. The buyers who scale efficiently centralise the buying authority but distribute the use-case discovery. The buyers who lose control let use-case discovery happen in business units without central visibility.
Exit posture is the part of the framework that vendors most resist - and that pays back hardest at renewal. The components: data deletion on termination with a certificate; derivative artifact (fine-tunes, embeddings) deletion with a certificate; a transition period of 30-60 days where the customer can continue using the service while migrating; portability of any custom training data and evaluation sets; and a contractual right to a model that is being deprecated for at least 12 months past the deprecation announcement. The exit posture is set at contract signature - it cannot be added at renewal.
Twenty-three clauses to redline, with exact replacement language for each.
The 2026 enterprise generative AI vendor shortlist for most use cases includes: OpenAI (via direct or Azure), Anthropic (via direct, Bedrock or Vertex), Google (via Vertex), AWS Bedrock (as a multi-model platform), and Microsoft (via Azure OpenAI). For open-source deployments, the typical platforms are Bedrock, Vertex, and self-hosting on AWS or Azure. The differentiation among the proprietary vendors has narrowed considerably on raw capability - the leading models are within a few percentage points of each other on most benchmarks - and broadened on commercial terms, data residency options, and enterprise platform maturity.
The selection criteria that matter most in our experience: data residency match for regulated workloads; track record of model deprecation (some vendors deprecate aggressively, others maintain models for years); pricing predictability over multi-year horizons; and the depth of the enterprise commercial team. The selection criteria that matter less than buyers think: raw benchmark performance, marketing-led feature comparisons, and vendor PR cycles.
A few reference points from deals we have benchmarked in the past 12 months. These are not exhaustive but they are directionally correct for 2026:
We can tell you what comparable deals are signing at, in your industry and at your scale.
For background on the broader contract design, see our AI vendor contracts pillar and data rights clauses deep dives. Once the deal is live, the recurring win is license cost reduction — right-sizing AI seats and consumption against real usage so you are not paying for capacity the organisation never adopts.
For commercial leverage, no. For operational simplicity, maybe. The buyers getting the best commercial outcomes run multi-vendor strategies with at least two proprietary models and one open-source deployment in active production.
One year. Two-year and three-year commits earn larger discounts but the technology is moving fast enough that locking in for three years usually costs more in foregone optimisation than it saves in upfront discount.
It depends on your existing Microsoft footprint. Customers with a large Azure commit usually get better effective economics on Azure OpenAI through MACC consumption. Customers without an Azure footprint get cleaner terms and faster feature access on OpenAI direct.
Buy for the first 18 months while the market matures. Re-evaluate at month 24 - by then the choice between continuing to buy and bringing it in-house should be obvious from your operational data.
Our AI & SaaS team builds procurement frameworks tailored to your portfolio of use cases.
Weekly compliance intelligence for IT leaders.