LLM purchasing guide
Buying LLM capacity is different from buying software. Usage is variable, pricing changes, and the headline rate is rarely the final rate. This guide helps platform teams buy AI capacity without overcommitting.
Match the pricing model to the workload
- Pay-per-token — best for variable, hard-to-predict usage.
- Volume discounts — useful once usage is steady and large.
- Committed use discounts — cheaper, but only if you are confident in the forecast.
- Batch API pricing — steep discounts for non-urgent work.
- Reserved throughput — for latency-sensitive production features.
- Enterprise agreements — for large, multi-product commitments.
Do not commit too early
Committed discounts look attractive, but they lock you into a forecast. Start with pay-per-token, collect three months of data, then negotiate. Most teams overestimate how much of their usage is stable.
Require attribution and guardrails
No deal should close without a plan for tagging, dashboards, and spend limits. If you cannot attribute usage, you cannot manage it. If you cannot cap spend, you cannot protect the budget.
Keep fallback providers
Single-provider dependency is a pricing and availability risk. Your contract and architecture should support fallback to at least one alternative. This also gives leverage at renewal.
Platform team checklist
- Workload mapped to pricing model
- Three months of usage data before commitments
- Attribution tags defined and enforced
- Cost caps and alerts configured
- Fallback provider tested
- Data residency and security terms confirmed
- Overage and price-change clauses reviewed
For the finance perspective on the same decisions, see the LLM purchasing guide for CFOs.
Want this applied to your own LLM spend? FinOps LLM runs a free audit of your AI costs and shows where the savings are. Book free audit →
FAQ
What pricing models do LLM providers offer?
The main models are pay-per-token, volume discounts, committed use discounts, batch API pricing, reserved throughput, and enterprise agreements.
When should you commit to an LLM volume discount?
Commit only after you have three or more months of stable usage data. Most teams overestimate stable usage because growth and model switching are unpredictable.
What should be in an LLM procurement checklist?
The checklist should cover usage attribution, cost caps, fallback providers, data residency, rate limits, billing granularity, overage pricing, and contract flexibility.