FinOps for AI starts with a unit cost, not a bigger dashboard

What happened

InfoQ’s Cloud and DevOps Trends Report 2026, published in August, maps practices by how widely they have been adopted. Mature cloud FinOps and observability sit in the late majority: most organisations know how to manage deterministic cloud costs. At the other end, among the innovators, the report places FinOps for AI and tokenomics — organisations have started tracking token consumption, but connecting that spending to business outcomes remains largely unsolved.

Alongside it at the frontier are enterprise AI platforms and AI gateways, which centralise model governance, and sovereign cloud strategies. Platform engineering and the Model Context Protocol, by contrast, have moved into the early majority.

Why it matters

In the engagements we see, AI spend follows a familiar pattern.

It starts invisible. A handful of teams call model APIs on separate accounts or on expense cards. The amounts are small, so nobody builds reporting.

Then it compounds. A pilot becomes a production agent. Context windows grow, agents call tools and retry, and usage scales with customers rather than with infrastructure. Monthly cost can move by an order of magnitude with no change anyone would recognise as a provisioning decision.

Then finance asks the only question that matters — is it worth it? — and nobody can answer, because spend is reported per model or per account, not per outcome.

That last step is exactly the gap the InfoQ report describes. More dashboards do not close it. A unit cost does.

What we would do

1. Route every model call through one place. An AI gateway, or at minimum a shared client library, that tags each call with the team, product and use case. Without attribution, nothing else works.
2. Define a unit of value per use case. Cost per resolved ticket, per contract reviewed, per qualified lead. Agree it with the business owner, and record the value of that unit alongside.
3. Measure cost per completed task, not per request. A cheaper call that needs three retries is not cheaper. Include retries, tool calls and human review time.
4. Tune the design before the price. Right-size models to the task, trim context, cache stable prompts, and batch work that does not need an immediate answer. These typically cut more than any commitment discount.
5. Put guardrails in the pipeline. Budgets and anomaly alerts per use case, and a cost estimate in the review of any change that alters prompts, context or model choice.

Treat AI cost the way good FinOps teams already treat cloud: owned by a team, explained in business units, and designed in rather than cleaned up afterwards.