The State of FinOps 2026 confirms it. AI ROI is the question no one can answer yet. Dashboards will not solve it. A different layer of infrastructure will.
Fifteen years ago, cloud spending was the line item no one could explain with confidence.
Engineering teams moved quickly. Billing lagged. Environments multiplied. Responsibility for results stayed vague.
I saw that up close. At RightScale and later at Flexera, we helped enterprises understand what they were spending on cloud and why. At McKinsey, I helped formalize that work into a practice called FinOps.
FinOps made accountability practical. It created a shared way to connect spend to decisions. Now the cycle is back, and it is moving faster.
The FinOps Foundation just published its State of FinOps 2026 report, and the numbers are telling.
It is the sixth annual survey, with input from 1,192 practitioners representing more than $83 billion in annual cloud spend. The headline is direct: FinOps for AI is the top forward-looking priority. 98% of respondents now manage AI spend, up from 31% two years ago.
One practitioner quote from the report really captured the problem underneath the stats:
"Is your AI providing value? No one can answer that question yet."
This is the pivot point.
The industry’s question is not “what did it cost?” It is “did it produce value?” And that is a very different measurement problem.
Cost and activity are easy to count. Value is harder to prove. It requires an AI economic control system that can tie AI activity to the business outcome it was meant to drive, and preserve that evidence when the outcome is disputed, audited, reversed, or priced.
The governance gap came back fast
If you’ve followed with my recent posts, you know I’ve been saying this for a while. Cloud FinOps followed a familiar sequence. Uncontrolled adoption turned into bill shock. Visibility came next, then allocation, then optimization, and finally governance. The arc took most of a decade.
AI compressed that arc into about two years.
This report shows that acceleration. In 2024, 31% of respondents managed AI spend. In 2025, it was 63%. Now it is 98%. AI moved from emerging concern to daily scope faster than any category the Foundation has tracked.
The word "manage" still carries a lot of weight. Most teams can show that AI spend exists. Some can locate the account. A few can identify the services.
That does not answer what the business will demand in the room. The business will demand attribution that ties AI spend to the specific features, customers, and workflows it served. The business will also demand visibility into which agent behaviors are quietly eroding margin.
Awareness is the starting point for governance.
What 1,192 practitioners just told us
The report surfaces three challenges that show up repeatedly when teams apply FinOps to AI.
Visibility into AI costs is hard because pricing differs across providers and services. One workflow might hit an inference API, a vector database, a tool API, and a GPU cluster for fine-tuning. Each one has its own pricing model, billing cycle, and unit of measure.
Allocating AI costs to business units is even harder than traditional infrastructure. Cloud resources map to accounts, projects, and tags. AI usage is embedded inside product features, internal workflows, and agent chains. It crosses teams and systems by design.
Determining AI value and ROI is still unclear. Many investments are exploratory. Returns are hard to define early. Outputs are not deterministic, so cost-to-value ratios shift run to run.
These are one problem viewed at three altitudes. Allocation depends on visibility. ROI depends on allocation. Skip a layer and everything above it becomes a story instead of a measurement.
The number-one tooling request is necessary, and incomplete
The report asked an open-ended question about missing tooling. The top answer was:
"Granular monitoring of AI spend (tokens, LLM requests and GPU utilization)."
That is the foundation layer. It is event-level telemetry for AI workloads. Telemetry is necessary, but it does not finish the job.
Granular monitoring gives technical truth. It tells you what ran, how often, and what it consumed. It helps with debugging and optimization.
But boards ask for economic truth. They want cost per customer. They want margin by feature. They want to know whether a workflow should scale or stop.
Telemetry without attribution is an expensive log. The missing join is between what happened in the AI stack and what it meant to the business. This is also where the market signal gets practical.
If the #1 missing feature is granular monitoring, then the near-term advantage goes to the teams who can deploy it without waiting for a multi-year roadmap from legacy FinOps tooling.
Revenium already provides this measurement layer: token, request, and GPU-adjacent telemetry captured as event-level records, with context hooks that let you attribute usage to the customer, feature, workflow, and agent that triggered it.
The integration requirement matters. The winning tool here is the one that fits into the execution path with minimal friction. If engineers have to redesign their stack to get visibility, adoption stalls. If you can add it in days, it becomes a platform primitive.
Cost per token is an infrastructure metric. Cost per outcome is an economic metric. Most organizations are still guessing in the space between those two numbers.
FinOps moved up. Tooling has to move closer to execution.
FinOps matured in a world where costs were mostly predictable. AI does not behave like that.
A single prompt can trigger a non-deterministic chain: retrieval, tool calls, retries, fallbacks, and re-synthesis. Forecasting breaks when autonomous systems can make spend decisions without checkpoints.
Then accountability shifts. When an AI outcome touches a customer, compliance posture, or margin, the business needs one case-level record that can reconstruct what happened end to end.
Dashboards summarize fleets. Governance requires unit-level proof.
Dashboards summarize. Ledgers prove.
The report included a second signal that matters just as much.
78% of FinOps teams now report into the CTO or CIO organization. Reporting to the CFO is down to 8%.
That shift changes where governance lives.
FinOps is not a finance reporting function anymore. It is a technology capability rooted in architecture, engineering, and platform decisions. The report links higher executive engagement with greater influence over technology selection.
If FinOps sits in engineering, then the economic control system for AI cost and value needs to sit in engineering too. It belongs in deploy pipelines, orchestration layers, API gateways, and agent runtimes.
This is where a lot of tooling quietly fails.
If “FinOps for AI” requires engineers to bolt on a heavyweight process, adoption collapses. The winning approach is developer-native: instrumentation that fits the way teams already ship software, and guardrails that run in the hot path.
That is the practical meaning of “shift left” for AI economics. You want cost and policy to show up at design time and deploy time, not as a retrospective after the invoice.
Revenium is built for this reality. Low-friction integration gives engineering teams granular measurement without re-architecting, and the control layer prevents the 3 a.m. runaway loop from becoming the seven-figure surprise.
Automation is not optional because teams are lean. Even organizations managing more than $100M in annual cloud spend often run FinOps with 8 to 10 practitioners. Headcount does not scale fast enough to keep up with machine-speed decisions.
This detail is easy to skim past, but it changes what “governance” can realistically mean.
If your enablement team is eight people and your agents can generate eight thousand economic events before breakfast, you cannot govern with human review. The only scalable path is automated guardrails.
That is the heart of an economic control system: budget enforcement and anomaly detection that run continuously, plus remediation mechanisms that can pause, cap, reroute, or block unprofitable behavior before it compounds.
Manual review arrives after the compounding has already happened.
Controls need to live in the hot path. Budget caps. Cost-velocity limits. Circuit breakers that stop spend before it accelerates.
What the next infrastructure layer looks like
The report describes the symptoms. The architecture comes next.
The problem breaks into four capabilities that build on each other.
Measure. Capture every AI transaction with full attribution to the customer, feature, workflow, and team that triggered it. Do not rely on sampling or aggregation. Treat it as an economic record that can support audit, governance, and billing.
Optimize. Compute true costs across models, providers, and infrastructure. Find waste in retry loops, model routing, and context growth. Enforce budgets and guardrails so efficiency is not left to best intentions.
Prove. Connect AI activity to outcomes. Calculate unit economics per workflow and per agent. Measure margin per customer, feature, and segment. Produce cost-per-outcome metrics that finance leaders can use.
Monetize. Turn attributable cost data into pricing and chargeback. Create rate cards and overages. Send billable events into billing systems.
This is the shift the report is pointing at.
Traditional cloud optimization still matters, but many practitioners are seeing diminishing returns on the old playbook. The “big rocks” have been moved. The next wave of influence is about technology selection, unit economics, and business value.
AI accelerates that shift because it is not just infrastructure spend. It is a product input.
When you can attribute usage at the transaction level, monetization stops being a debate and becomes an implementation detail. Cost-per-outcome becomes the unit you can price. Usage becomes a set of billable events you can meter. Rate cards, overages, and chargeback become outputs of the same AI economic control system, not a separate spreadsheet model.
Put bluntly: you can stop treating AI as a cost center and start treating it like a profit engine.
The question has an answer
The FinOps Foundation updated its mission this year. It moved from advancing people who manage the value of cloud to advancing people who manage the value of technology.
That wording matters because AI is not just another line item. It is a decision stream that needs ledger-grade records tied to outcomes.
The State of FinOps 2026 reinforces what practitioners already feel. Optimization is table stakes. Value is the goal. AI raised the stakes and shortened the timeline.
The organizations that win will not be the ones with the most agents. They will be the ones that can answer one question with data.
What did this accomplish?
Not how many tokens it consumed. Not how many model calls it triggered. Not what a dashboard reported.
What outcome did it produce, what did it cost, what margin did it leave, and should the business repeat it.
If you cannot trace it, you cannot tune it.
If you cannot price it, you cannot scale it.
If you cannot prove it, you cannot trust it.
The report's question has an answer, but it requires a different kind of infrastructure. Not more dashboards. A ledger. Not after-the-fact reporting. Economic control in the execution path.
It is time to put a meter on the machine.
See how Revenium connects AI activity to business outcomes →
