Somewhere in a data center right now, a server is running that nobody needs.
It hums, draws power, generates heat and lands on an invoice every month. The team that launched it moved on. Finance never saw it. Multiply that machine across every cloud account at every company in America and the number becomes genuinely hard to look at. How did a problem this expensive stay so invisible for so long?
How AI turned a manageable leak into a flood
Cloud waste is not new, but its engine just changed. For years the culprit was the over-provisioned virtual machine, a server sized for peak demand that rarely arrived, quietly burning budget at 5 to 8 percent of its available processing power. Finance teams learned to spot those.
Then AI arrived with a completely different kind of invisible cost. An idle inference endpoint, the cloud service that fields AI queries in production, can still appear active on a dashboard even while serving almost nothing. Unlike a forgotten virtual machine, it carries no obvious visual signal that it has gone cold.
Enterprises are spinning up GPU clusters and half-finished model experiments faster than finance teams can tag or shut them down. Fast launches and slow visibility together are the mechanism behind the number that just landed on annual cloud bills everywhere.
What a 29 percent waste rate looks like inside a real cloud account
Twenty-nine cents of every dollar spent on cloud IaaS and PaaS this year is estimated to be producing nothing useful. That is the finding from Flexera’s 2026 State of the Cloud Report, and it marks the first time the rate has risen in five years.
The types of waste are specific. Unattached storage volumes, unused network addresses, orphaned snapshots and idle load balancers are among the most common sources. Each one appears on an invoice every month, attached to nothing anyone is actively using.
Ghost instances compound the problem. The key misconception is that servers not doing work cost very little. In reality, idle servers in typical deployments still consume around 60 percent of their peak power draw, so even a machine doing nothing costs real money every hour it stays on.
Seventy-six percent of large enterprises now spend more than $5 million a month on public cloud. Small percentage swings in efficiency translate into enormous dollar amounts before anyone notices.
Where the $44.5 billion figure comes from and who verified it
The 29 percent waste rate is drawn from the Flexera 2026 State of the Cloud Report, based on a survey of 753 cloud decision-makers. The waste figure had been declining for five years, and this year’s uptick reflects the increased cost complexity brought on by AI and new IaaS and PaaS services.
A Harness “FinOps in Focus 2025” report estimates that an equivalent of roughly $44.5 billion in cloud infrastructure waste was projected for 2025, driven by a persistent disconnect between financial operations strategy and developer behavior. That figure represents Harness’s estimate of 21 percent of enterprise cloud infrastructure spend, a narrower definition than Flexera’s broader IaaS and PaaS measure. At roughly $675 billion in total global cloud spend, even a definition limited to genuinely idle resources produces a very large number.
Separately, 52 percent of engineering leaders say the gap between FinOps and development teams is the primary driver of wasted spend. Fewer than half of developers have access to real-time data on idle cloud resources, and only 33 percent can see whether workloads are over- or under-provisioned.
That visibility gap is where the dollars disappear. Much of the resulting waste, as GCN has covered in the context of record startup investment, is flowing into companies that promise to solve exactly this problem. Harness’s senior director of product management described the core issue as developers not viewing cost optimization as a priority, leaving overprovisioned resources and idle instances to quietly drain budgets.
Where the 29 percent number breaks down and who it does not hit equally
The 29 percent is a mean, and the spread beneath it is wide. Most organizations report 28 to 35 percent waste at baseline, while mature programs sustain roughly 15 to 20 percent. That gap shows the problem is not structural. It is a governance gap some teams have already closed and others have not started on.
Industry type matters significantly. Retail, media and manufacturing trend higher due to bursty traffic, content pipelines and edge data flows. A retailer running weekend flash-sale infrastructure that stays provisioned on Tuesday morning is paying for capacity it has zero use for.
AI inference is a sharp new blind spot. Analysts broadly agree that inference now accounts for a growing share of AI cloud spend, and that is exactly where waste is concentrated. An endpoint built for a pilot, then replaced by a better model, may run untouched for months. For companies already tracking the full financial exposure of their cloud footprint, idle AI endpoints represent a newer and less-mapped category of risk.
What is changing and what the waste bill still cannot tell you
The major cloud platforms are responding. The reversal in waste rates has triggered a scramble to ship AI-powered cost tools that can flag an idle endpoint automatically, recommend a shutdown and estimate the monthly saving.
The catch: those tools only work if the resource was tagged correctly at launch, and tagging discipline remains the weakest link in most engineering organizations. Three forces sustain today’s waste: complexity across multi-cloud, containers and AI clusters; velocity from self-service provisioning without guardrails; and fragmented cost visibility. Fixing any one without the other two produces only a partial result.
That is why the waste rate has remained stubbornly elevated across years of measurement. The Flexera figure stood at 28 percent in both 2022 and 2023, then eased to 27 percent in 2024 and 2025, matching the period when dedicated financial operations teams were both forming and growing fast. The question now open is whether those same teams can absorb AI infrastructure quickly enough to bend the curve back down, or whether 29 percent becomes the new floor before anyone figures out how to tag a GPU cluster nobody asked for.
