State of FinOps 2026: 98% now manage AI spend

Published 15 August 2026

The FinOps Foundation's State of FinOps 2026 survey covered 1,192 practitioners representing more than $83B in annual cloud spend. The headline number: 98% now manage AI spend, up from 63% in 2025 and 31% in 2024. In two years AI cost management went from a minority specialism to something practically every FinOps team owns.

The adoption curve

Roughly a doubling, then a saturation. There is nowhere left for that line to go, which makes 2026 the last year the interesting question is whether teams manage AI spend. From here the question is how well.

And "manage" is doing heavy lifting in that 98%. It counts everyone who has been handed the problem, including the large share who have a monthly bill, a rough split by provider, and no idea which feature caused which line. Adoption is universal. Capability is not. Do not read the number as maturity.

FinOps is no longer a public-cloud discipline

The survey's scope numbers make the point better than the AI ones do:

• 90% manage SaaS spend
• 64% manage licensing
• 57% manage private cloud
• 48% manage data center

FinOps in 2026 is a technology-spend discipline that happens to have started in public cloud. AI is the fastest-moving line inside it, and it crosses every one of those categories at once: an AI feature can bill as API usage, as a SaaS seat, as a licence, and as reserved GPU capacity in the same month. That is precisely why it resists the tooling built for any single category.

The most-requested capability does not exist yet

The finding worth the most attention is not the 98%. It is this: granular AI spend monitoring - tokens, LLM requests, GPU utilization - is the single most-requested capability in the entire survey. Not most-requested among AI capabilities. Most-requested, full stop.

A capability at the top of a request list of 1,192 practitioners is a capability the market has not delivered. Vendors have shipped AI cost dashboards; what they have not shipped is allocation at the granularity that makes a dashboard actionable. The reason is structural, not lazy:

• Provider invoices arrive pre-aggregated. A monthly total per key or per project cannot be decomposed into features after the fact. If the request did not carry a tag when it was made, the information is gone.
• The unit is not stable. Input, output, cached, and reasoning tokens price differently and the ratios shift every time you change a model. See token types explained.
• GPU utilization and token spend are different accounting systems that have to be reconciled by hand to produce one number for one workload.

A vendor can solve dashboards from billing exports alone. Nobody can solve allocation without instrumentation inside your request path.

What it means that nobody has shipped it

The practical consequence: the visibility layer is yours to build for at least another cycle. Teams waiting for a procurement decision to solve token allocation will spend 2026 without token allocation.

The good news is that the build is smaller than the wait. The instrumentation that produces granular AI cost is a gateway tagging contract plus a monthly reconciliation, and it is a matter of weeks:

1. Route every model call through one gateway. Everything below depends on this and nothing else does.
2. Require four tags on every request: feature, environment, team, customer or tenant. Reject untagged traffic in staging first, then production. See cost attribution.
3. Log token counts by type, not just totals. Input, output, cached, reasoning. See reasoning token attribution.
4. Reconcile to the provider invoice monthly. An unreconciled internal number is a guess with a decimal point. See invoice reconciliation.
5. Publish one unit-economic metric - cost per request, per ticket, per active user. See cost per request as a KPI.

Emit the tags using OpenTelemetry GenAI semantic conventions rather than a private schema. When tooling does catch up, the conventions are what it will read.

The skills signal: value, not cost

The top skillset FinOps teams say they are hiring for is AI value management. That is a different job from cost reduction, and the difference shows up in what gets asked in the meeting. Cost reduction answers "why is this so expensive." Value management answers "what did the spend produce, and would we buy it again at this price."

You cannot answer the second question with a cost dashboard. It needs a denominator - tickets resolved, PRs merged, deals sourced - attached to the same tag the cost was attributed to. Which is the same instrumentation as above, used for a better question. This is also the gap the Tokenomics Foundation was launched to standardize, and its first workstreams are aimed squarely at it.

What to take from this if you own AI spend

• Being in the 98% is table stakes and not a position. The peer group is everyone now. Differentiation is granularity.
• Do not buy on the promise of token-level allocation. Ask any vendor how they attribute a request that was made without tags. There is no good answer, which tells you the tagging is yours regardless of what you buy.
• Instrument before you optimize. Optimization on unattributed spend is guessing which knob moved the total. See the six levers once attribution is in place.
• Budget in consumption as well as dollars. Unit prices keep falling while bills rise; a dollar budget hides the volume growth that actually explains the invoice. See cost trends 2025-2026.

Related

• What is LLM FinOps? - the discipline this survey measures.
• What is the Tokenomics Foundation? - the standards body aimed at the value gap.
• What is LLM cost attribution? - the capability nobody shipped.
• Cost per request as a KPI - the first unit metric to publish.
• How to budget for AI spend - what to do with the numbers once you have them.

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 →

Back to research