AI Spend Broke the FinOps Playbook

The FinOps Foundation's State of FinOps 2026 surveyed 1,192 practitioners who between them manage more than $83 billion of annual cloud spend. One number in it reframes the rest.

98% of FinOps teams now manage AI spend. Two years ago it was 31%.

That is a whole new cost category arriving in twenty four months, and "FinOps for AI" comes out as the top forward looking priority in the same survey.

The trouble is that the techniques that made cloud spend manageable were built on an assumption that AI spend breaks.

Old cloud cost is a provisioning problem

Rightsizing, reserved capacity, savings plans, shutting down idle environments, storage tiers. All of it works because cost follows a provisioning decision.

Someone picks an instance type. That choice sticks around. It shows up in infrastructure code, you can review it, and you can change it once to get a saving that lasts.

Even the reactive parts assume things stay put. An idle resource sits there for weeks before anyone notices, which is annoying, but it does mean there is time to notice.

AI cost is a usage problem

Inference spend does not follow a provisioning decision. It follows behaviour. How many users. How many calls per interaction. How much context each call carries. How many retries. How many agent steps. How many tokens the model decides to produce.

That changes the cost in ways that matter:

It moves without a deploy. Edit a prompt template, put a longer document in the context window, or run a marketing campaign, and spend multiplies with no infrastructure change at all. Nothing in your change review is looking at that.

Last month does not predict next month. A retrieval step that grows the average context, or an agent loop that sometimes runs twenty times instead of three, breaks the link between history and forecast.

The unit is unfamiliar. Most teams have ten years of instinct about what an instance hour costs. Almost nobody has instinct about what a token costs at scale. Estimates stay guesses until the first real bill lands.

Failures cost full price. A request that fails after consuming a large context still bills. A runaway agent loop bills every step. In classic cloud, errors are mostly free. Here they are not.

What carries over, and what has to change

Work out unit economics before launch. Total AI spend tells you almost nothing. Cost per conversation, per document, per resolved ticket is the number that says whether the feature makes sense. Have it before you ship, not after the invoice surprises someone.

Set hard limits, not just alerts. A budget alert tells you the money is already gone. Usage driven spend needs enforced caps: per tenant and per feature limits, a maximum step count on any agent loop, token limits per request, and circuit breakers that degrade instead of billing forever. This is the biggest break from the old playbook, and the one most teams add only after an incident.

Treat cost as part of the change. Spend moves with prompt and orchestration changes, not infrastructure changes. So cost review belongs in code review for those paths. A pull request that doubles the average context length is a cost change and should be discussed as one.

Try the cheap options before the bigger model. Cache. Use smaller models for classification and routing. Cut context that adds nothing. Do not call the model at all when a simple rule will do. Most of the large savings we see come from removing unnecessary calls, not from negotiating a better rate.

The organisational half

The same report notes that 78% of FinOps teams now report to the CTO or CIO, and that teams with VP level engagement have several times more influence over which technology gets chosen. That matters more here than it did for classic cloud.

Rightsizing an instance is something a FinOps practitioner can drive alone. Capping an agent loop, picking a smaller model, or changing what goes into the context window is a product and engineering decision with quality trade offs attached. It only happens if cost is in the room where those trade offs get made.

That is the real shift behind the jump from 31% to 98%. AI spend did not just add a line to the bill. It moved the decisions that control the bill out of infrastructure and into the product.