This chapter covers the cost of AI features: how language models and the compute behind them are priced, how model providers and cloud platforms bill for them, and how that cost is allocated, budgeted and reduced per feature and per customer. It applies the groundwork of the other chapters to a kind of spend that is metered differently from servers and storage and often arrives on a different invoice.

What this chapter covers

The discipline is still FinOps; what changes are the units and the data sources. That is why the FinOps Foundation lists FinOps for AI as a technology category of its framework in its own right, beside public cloud, SaaS, data cloud platforms and the data centre. Its reasons are the ones this chapter is built on: AI cost follows its own units, swings with experiments and reaches the company through several channels at once.

The Units section answers what you are actually paying for. It explains what a token is and why a request pays for its whole context, how model price lists are built, what training, fine-tuning and inference each cost, and how GPU capacity is charged when a model runs on infrastructure you operate. There you no longer pay per token but per hour of compute, and the cost question becomes how much of that capacity is actually working.

The Billing section answers where that cost lands. Bought directly from a model provider, AI usage arrives on an invoice of its own, often in US dollars and with its own export interface; bought through a cloud, it appears on the cloud bill under that cloud's meters. The section describes both routes and the AI cost that sits outside the cloud bill altogether, such as AI tools and assistants paid by credit card or licence agreement, which therefore appear in no cloud billing export.

The Practice section answers how AI cost is managed. It carries allocation, unit cost and budgets over to AI features, shows how retrieval, vector databases and agents making tool calls can multiply the cost of a single request, and treats caching, batching and routing between models as the levers that act on the price per token and on the volume. At its centre is what a single answer or a completed task costs, and who owns that cost.

Key terms

  • Token: the unit in which language model APIs count and bill input and output.
  • Unit economics: cost per request, per outcome or per customer, which shows whether an AI feature pays for itself.
  • Cost allocation: splitting provider and cloud cost across features, teams and customers, here through projects, workspaces and API keys.
  • Budget: the plan with an owner and thresholds that provider usage and spend limits do not replace.
  • Anomaly: an unexpected rise in cost, with AI often caused by a growing context, an agent stuck in a loop or a new model.
  • FinOps: the discipline whose practice this chapter applies to AI spend.

Neighbouring chapters

The chapter follows "Governance and the operating model" and applies its roles, approvals and guardrails to AI features, where an experiment turns into running cost quickly. Allocation and budgets themselves are the subject of "Allocation and ownership" and "Budgets, forecasts and finance"; this chapter assumes them and shows what changes with tokens, provider invoices and model upgrades. It is followed by "Europe: currency, tax and sovereignty", which takes up the questions an invoice in US dollars from a provider outside the European Union raises. Every chapter is listed in the handbook.