Cloud waste rose for the first time in five years

For five years running, the industry’s estimate of its own cloud waste went down. In Flexera’s 2026 State of the Cloud survey of 753 cloud decision-makers, it went up: respondents now estimate 29% of their IaaS and PaaS spend is wasted, versus roughly 27 to 28% in the years before. One point of drift would be noise. A reversal after five years of steady improvement is a signal, and the report’s own explanation names the suspect: “growing cost complexity from AI and new IaaS and PaaS services.”

Standard caveats first, because this number will get quoted carelessly all year. It’s 29% of spend estimated as waste, a share, and a self-reported one; it is emphatically not “waste grew 29%.” And Flexera sells cost tooling, so the survey deserves the same skepticism as any vendor research. That said, the trend line is the interesting part, and a vendor has no incentive to report its customers’ problem getting worse five years into buying the tooling.

The same survey has cost outranking security as the top cloud challenge for the fourth consecutive year, 85% to 82%. Read those two numbers together with a third one: 63% of these organizations have implemented FinOps practices. Most of the market has the tooling, the dashboards, and the practice, and the waste went up anyway.

Why the trend reversed

I spend my working life inside cloud bills, and the 2026 reversal matches what the bills look like from the inside. Cost programs built over the last five years were tuned to a particular kind of waste: the oversized instance, the unattached volume, the forgotten environment, the storage that should have aged into a cheaper tier years ago. That work paid off, which is exactly what the five declining years show.

AI spend doesn’t look like that. It scales with usage rather than with what you provisioned, it arrives on new and expensive line items, and the teams shipping it are usually moving too fast for the design review where cost questions get asked. A GPU fleet at low utilization, an inference endpoint nobody metered per-feature, an agent retrying its way through a bad prompt: none of these get caught by the idle-instance report. The waste-hunting muscle the industry built simply doesn’t reach the fastest-growing part of the bill. Meanwhile 17% of organizations report blowing through their cloud budgets, a rate Flexera describes as stable year over year, with AI positioned to push it higher.

The reversal, in other words, is a coverage gap. The old discipline still works on the old waste. Nobody extended it to the new spend in time.

What reverses the reversal

The answer is not another dashboard, and I say that as someone who reads the bill as a design document. Three moves, in order.

Extend attribution to the AI line items before optimizing anything. Shared inference endpoints and GPU pools have the same property as shared EKS clusters: tagged perfectly, they still attribute 100% of the spend to “the platform.” You need per-request usage data, which feature, how many tokens, what model tier, before any of the downstream decisions are possible.

Put the AI spend under the same review cadence as the rest of the bill. The workloads that reversed the trend are the ones that skipped the design review. A cost estimate at design time, and a monthly look at actuals against it, catches an expensive architecture while it’s still cheap to change. I’ve written about what that takes for AI workloads specifically.

Then do the unglamorous work. The 63%-have-FinOps number is the tell: visibility is mostly a solved problem, and the remaining gap is execution. Right-size the GPU fleet against measured utilization. Meter the endpoints. Set budgets that page engineering before finance notices. That’s not a tooling purchase. It’s a few weeks of engineering attention pointed at the newest part of the bill, which is precisely the part nobody owns yet.

Waste climbing after five years of decline isn’t a scandal. It’s what happens every time a new spending category outruns the discipline built for the old ones. The organizations that pulled waste down for five straight years did it by treating cost as an engineering property. The fix this time is the same, aimed at a new target.