Cloud Waste Was Falling for Five Years Straight. Then AI Showed Up.
For five consecutive years, the trend line on cloud waste pointed in exactly the direction every finance leader wants to see: down. FinOps teams got more disciplined, tagging and rightsizing got more automated, and the industry-wide percentage of cloud spend going to waste crept steadily lower, year after year. Then, in Flexera's 2026 State of the Cloud Report, that five-year streak broke. Wasted cloud spend rose to 29 percent, the first increase since the firm started tracking the reversal-worthy version of this number.
That's a strange thing to see turn around after half a decade of consistent progress, and it's worth sitting with why, because the answer isn't that FinOps stopped working. It's that something new got layered on top of it faster than governance could keep pace.
The Report Behind the Number
Flexera's 2026 report, its fifteenth annual edition, surveyed more than 750 cloud decision-makers and users globally. Alongside the waste figure, it found that generative AI adoption inside organizations climbed to 81 percent, up from 72 percent the year before and just 47 percent two years ago. Perhaps more tellingly, the share of organizations running AI extensively in actual production environments, not pilots, not proofs of concept, jumped to 45 percent, up from 36 percent the prior year.
Put those numbers side by side and the reversal in cloud waste stops looking like a mystery. AI workloads are computationally expensive, often provisioned quickly under real business pressure to ship something, and frequently owned by teams who are optimizing for getting a model into production, not for what that model costs to keep running. Eighty-five percent of organizations in the same survey named managing cloud spend their top challenge, a discomfort that's climbing at almost exactly the pace AI adoption is.
The Part That Doesn't Fit the Easy Narrative
It would be a tidy story to say cloud waste rose because AI teams are reckless with infrastructure spend, but Flexera's own framing pushes back on that read, and it's worth taking seriously. The report's authors describe cloud as maturing, not regressing, noting that organizations are moving past treating cloud purely as a cost-cutting exercise and toward measuring the actual value it generates. Sixty-three percent of organizations now report having an established FinOps team in place, and 71 percent operate a formal Cloud Center of Excellence. That's not an industry that abandoned financial discipline. That's an industry that built real discipline for one kind of workload, right as an entirely different kind of workload, with a different cost profile and a different pace of adoption, showed up and outran it.
This distinction matters, because the instinct after seeing a waste number tick upward is usually to tighten the existing FinOps process, run another rightsizing pass, chase down another round of orphaned storage volumes. Those are the right moves for the workloads FinOps was originally built to govern. They don't touch the newer problem, which is a category of spend that didn't exist in most organizations' cost models three years ago and is growing at a rate the governance layer hasn't caught up to yet.
Why AI Spend Breaks the Old Playbook
Conventional cloud waste has a familiar shape: an oversized instance nobody resized, a dev environment left running over a weekend, a storage tier that should have been archived months ago. It's wasteful, but it's also relatively legible. Someone can look at a utilization dashboard and see the gap between what's provisioned and what's actually used.
AI workloads don't fail in that same visible way. A model in production consuming GPU capacity around the clock doesn't necessarily look wasteful on a utilization dashboard, because it may genuinely be running, genuinely processing requests, and genuinely doing the job it was built for. The waste hides inside a different question entirely: whether the model needed to be that large, whether the inference workload needed to run continuously rather than on demand, whether a smaller or more efficient architecture would have delivered the same business outcome at a fraction of the compute cost. Answering that question requires a level of technical and financial fluency that most existing FinOps processes, built around infrastructure utilization rather than model architecture decisions, simply weren't designed to provide.
Governance Is Playing the Same Game of Catch-Up
The same Flexera research found that only 47 percent of large enterprises have established dedicated AI governance, even as production AI usage climbs toward half of all organizations. Fifty-three percent cite security and compliance as their top AI challenge, and 40 percent flag data quality concerns specifically tied to AI training. None of those numbers describe an industry that's ignoring the problem. They describe an industry where governance, in every dimension at once, cost included, is running behind adoption rather than ahead of it.
That pattern should sound familiar. It's the same shape that's shown up in nearly every recent wave of enterprise technology adoption: the capability arrives, gets adopted because the business case for using it is immediate and obvious, and the governance structure needed to use it responsibly gets built afterward, under pressure, once the gap has already become expensive. Cloud computing itself went through this cycle roughly a decade ago. Software-as-a-service sprawl went through a milder version of it more recently. AI infrastructure spend appears to be running the same loop, just considerably faster, because the underlying compute costs so much more per unit than the SaaS licenses that caused the previous round of sprawl.
What This Actually Costs, Beyond the Waste Number
Public cloud spending overall isn't a small line item to begin with. Gartner has projected worldwide public cloud end-user spending in the neighborhood of 723 billion dollars, and growing considerably from there as AI workloads scale. A jump in the waste percentage against a base that large translates into a genuinely significant amount of money moving from "productive infrastructure spend" to "spend nobody is getting value from," even before accounting for how much larger AI-specific compute costs run compared to the conventional workloads that dominated cloud bills a few years ago.
The organizations feeling this most acutely are the ones running significant cloud footprints already. Seventy-six percent of large enterprises in Flexera's survey report spending more than 5 million dollars a month on cloud infrastructure. At that scale, a few percentage points of additional waste isn't a rounding error. It's a budget line that starts drawing real scrutiny from finance, often before the engineering side has built the tools to explain where the money actually went.
What Separates the Organizations Managing This Well
The pattern separating organizations that are handling this well from those that aren't isn't whether they've adopted AI aggressively. Adoption is happening everywhere, at similar rates, regardless of how well-governed the spend behind it is. The difference is whether cost visibility got extended to cover AI workloads specifically, with the same rigor that traditional infrastructure spend already has, rather than assuming the existing FinOps process would simply absorb a new category of spend it was never built to measure.
That's a genuinely different kind of governance work than rightsizing a virtual machine. It requires someone asking, before a model goes into production, what it will actually cost to run at the volume the business expects, and whether that cost was ever weighed against the value the model is expected to generate. Most organizations, based on this year's numbers, haven't built that question into their process yet. The ones that have are the ones for whom rising AI adoption and falling waste aren't actually in tension with each other.
What This Actually Means
Five years of steady progress on cloud waste wasn't undone because anyone stopped caring about efficiency. It was undone because a new category of spend arrived faster than the discipline built to govern the old one could extend to cover it. That's not a reason to slow AI adoption, and nothing in this data suggests the organizations moving fastest are moving recklessly. It's a reason to treat AI infrastructure cost as its own governance problem, with its own visibility requirements, rather than assuming the FinOps maturity that took five years to build for conventional cloud spend will automatically stretch to cover a workload it was never designed to measure.