29% of IaaS and PaaS spend is wasted. The figure comes from Flexera, in a March 2026 survey of 753 organizations, and it carries an uncomfortable novelty: for the first time in 5 years, waste grew.
The curve had been descending. A 32% peak in 2021, a drop to 27% in 2023, and now the reversal. In dollars, Red9 estimates roughly 122 billion wasted in IaaS and PaaS in 2025, in a public cloud market that passes 1 trillion dollars in 2026, according to IDC.
Why waste is rising again
Flexera points to three causes: AI cost complexity, new pricing models, and underutilized commitment discounts.
All three share the same root. AI brought a spending category that grows faster than the governance around it: tokens, inference, dedicated GPUs, experiment environments nobody shuts down. FinOps processes designed for VMs and storage do not capture that spend in the same cycle. The bill grows before the control exists.
The idle GPU is the symbol of the decade
The harshest number in the report is not the 29% average. It is average GPU utilization in enterprise environments: 5%, according to Cast AI, measured across 23,000 Kubernetes clusters.
The math is simple and brutal: at 5% utilization, effective cost is 20 times the nominal price. An idle H100 costs 8,850 dollars a month. The GPU is not expensive: you are paying for 20 GPUs to use 1. We have already detailed the 8% utilization crisis in Kubernetes here on the blog, and the pattern repeats in the AI layer with worse numbers.
Commitment without criteria became a liability
The other side of the coin: underutilized commitment discounts. According to VendorBenchmark, 31% of 3-year AWS Reserved Instances generate zero net benefit, with utilization below break-even.
A Reserved Instance is a bet that the workload will stay there, in the same shape, for 3 years. Bought by inertia or in the shock of an on-demand invoice, it becomes the very waste it was supposed to fight. The market is bringing workloads back to owned infrastructure precisely when the commitment math stops closing.
63% have FinOps teams. And waste grows
The central paradox of the report: 63% of organizations have a FinOps team, yet practitioners report diminishing returns. The big rocks have been captured. What remains is harder.
The discipline’s answer was to expand: the FinOps Foundation changed its mission from managing the value of cloud to managing the value of technology, including AI, SaaS, licensing, and private cloud. The number that measures the urgency: 98% of FinOps teams now manage AI spend, across 1,192 organizations that together account for 83 billion dollars in cloud. In 2024, it was 31%.
And a new mechanism emerged, self-funding: organizations being required to finance AI investment with optimization savings. Cutting cloud waste to pay for GPUs. With 17% of organizations already over budget and 27% expecting spend to increase, optimization has stopped being a savings program and became a prerequisite for investment.
Where the fix starts
Tech86 implements FinOps for AI infrastructure in LATAM, and the entry diagnosis is almost always the same: the waste is not hidden in exotic metrics, it lives in architecture. Resources with no owner, 24/7 test environments, fear-driven overprovisioning, unreviewed commitments.
The observability tool is the last step. Before it comes the sequence that resolves most of the 29%: map resources with owners, kill the obvious, rightsize with real usage data, audit commitments against current usage, and build governance into the architecture (mandatory tags, budget alerts, automatic shutdown).
Conclusion
The 29% waste is not a character flaw of the cloud market: it is the default price of self-managed infrastructure without governance. The good news is that it can be attacked in order of impact, and the first 10 percentage points are almost mechanical.
For anyone about to sign the next 3-year commitment or buy the next GPU, the right question comes before the invoice: who owns it, what is the measured real usage, and what is the review plan. If the answer does not exist, the waste is already under contract.