Flexera’s 2026 State of the Cloud Report, published March 18, 2026, reports wasted cloud spend at 29% — the first increase in five years. The same report finds 63% of organizations with an established FinOps team and 71% operating a Cloud Center of Excellence.

Those two results sitting in the same document is the finding. Cloud financial management has never been more widely staffed, and the waste number went the wrong way anyway. The interesting question is not whether FinOps works. It is what changed underneath it.

What the Report Actually Says

The methodology first, since it bounds everything else: this is the 15th annual edition, built on a global survey of more than 750 cloud decision-makers and users, published March 2026.

The headline numbers:

  • Wasted cloud spend: 29%, up after five consecutive years of decline
  • FinOps teams: 63% of organizations
  • CCOE or equivalent: 71%
  • Generative AI in use: 81% of respondents, with 45% using it extensively
  • GenAI is now the third most-used public cloud service at 58%, up from 50%
  • Value delivered to business units: 64%, up 12 percentage points year over year
  • Unit economics adoption: up from 40% to 49%
  • Cost efficiency: down 6 percentage points as a stated top priority

One methodological caveat worth carrying through the rest of this: the waste figure is respondents’ own estimate of their wasted spend. Self-estimated waste is a floor, not a measurement — it reflects what organizations believe they are wasting, which is bounded by what their tooling shows them. A team without granular allocation for a workload type cannot estimate waste in it accurately. Given that the report also shows GenAI usage climbing steeply, that limitation is not evenly distributed.

Why AI Workloads Break the Optimization Playbook

Flexera attributes the reversal to cloud-based AI workloads and the cost complexity of newer IaaS and PaaS services. That is worth unpacking, because the standard FinOps toolkit is genuinely poorly matched to this workload class.

The mature optimization plays all depend on predictability. Rightsizing assumes a workload has a stable resource profile you can measure and trim to. Committed-use discounts and reserved capacity assume you can forecast a baseline far enough ahead to commit to it. Autoscaling assumes demand signals that map cleanly to capacity. Scheduling assumes idle windows you can identify.

GenAI workloads violate most of that. Usage patterns are driven by experimentation rather than by production traffic curves. Accelerator capacity is expensive, frequently supply-constrained, and often reserved defensively — which means paying for idle capacity is sometimes the rational choice rather than a mistake. Inference demand can be bursty in ways that do not resemble web traffic. And a large share of 2025–2026 AI spend went to workloads that had no prior-year baseline to forecast from at all.

The result is a category of spend that is large, growing, hard to forecast, and hard to attribute — which is the exact profile that produces waste. Flexera’s own reporting notes organizations struggling specifically with forecasting unpredictable GenAI usage patterns.

FinOps Maturity Did Not Prevent This, and That Is Not a Failure

The uncomfortable reading — that 63% FinOps adoption coexisting with rising waste means FinOps is not delivering — is the wrong one, but it needs an actual argument rather than a dismissal.

The argument is that FinOps disciplines are workload-specific in practice even when they are general in principle. A FinOps team that spent five years building genuine competence in EC2 rightsizing, storage lifecycle policies, commitment portfolio management, and Kubernetes cost allocation has real capability that does not transfer automatically to accelerator fleets and token-metered API services. The practice matured against one cost structure; a new one arrived.

The supporting evidence is in the governance numbers. CCOE adoption at 71% and FinOps teams at 63% are not the profile of an organization ignoring cloud cost. They are the profile of an organization that built the function and then had the workload mix change under it.

That reframes the response. The action is not “do more FinOps.” It is “extend FinOps coverage to the workload class that is currently outside it” — which is a scoping and tooling problem with a clear owner, not a cultural one.

The Metric Shift Is the Real Story

The quieter finding may matter more over a multi-year horizon. Value delivered to business units rose 12 percentage points to 64%, unit economics adoption climbed from 40% to 49%, and cost efficiency dropped 6 points as a top priority.

Read together, that is FinOps moving from a savings function to a value function. The question shifts from “how much did we spend” to “what did each unit of spend produce” — cost per transaction, per customer, per inference, per business outcome.

This is the correct direction and it has a specific hazard attached. Unit-economics framing can absorb waste invisibly: if revenue per customer grows faster than infrastructure cost per customer, the unit metric improves while absolute waste grows. A team reporting only on unit economics can show a healthy trend line on top of a genuinely wasteful platform.

Both numbers belong on the dashboard. Unit economics answers whether the spend is productive. Absolute waste answers whether it is efficient. The 2026 report is a fairly direct demonstration of what happens when attention shifts from the second to the first.

What to Change

Get AI spend into the same allocation model as everything else. If accelerator fleets, model-inference APIs, and vector storage are not tagged and attributed with the same rigor as compute and storage, the waste in them is invisible rather than absent — and self-estimated waste figures will keep understating it.

Treat idle accelerator capacity as a policy decision, not an anomaly. Sometimes holding constrained capacity is correct. Make that an explicit, owned decision with a review date rather than something the optimization tooling flags and everyone ignores.

Separate experimentation spend from production spend. They have different economics and different acceptable waste levels. Merging them means the acceptable waste in one hides the unacceptable waste in the other.

Keep reporting absolute waste alongside unit economics. The metric shift is right; dropping the old measure is not.

The five-year decline in cloud waste was real, and so is this reversal. What it indicates is that a new workload class arrived faster than the discipline built to manage the old one could extend to cover it. That is a solvable problem, and the organizations solving it will be the ones that stopped treating AI infrastructure as a special case exempt from cost governance.