Kubernetes Clusters Run at 8% CPU, 5% GPU: The FinOps Bill Comes Due

The average production Kubernetes cluster used 8% of its CPU in 2025, down from 10% a year earlier. Cast AI's 2026 Kubernetes optimization report puts memory utilization at 20%, down from 23%, and GPU utilization at 5%. For teams paying for AI inference nodes, those three numbers define the cost problem of late 2026.

The report states it plainly: "Average CPU utilization across analyzed production clusters fell to 8% in 2025, down from 10% the prior year." A cluster can look healthy on a dashboard while most paid capacity sits idle. Requests and limits, not real demand, drive what a node costs.

What do the numbers leave out?

Two caveats belong next to these figures. Cast AI sells automation, so its sample may not resemble any given fleet. No official Kubernetes, AWS, Microsoft or Google page confirms the 8% figure. The cloud providers publish no fleet-wide utilization data of their own. Treat 8% as a strong signal, not a universal constant.

The direction is still clear. Utilization fell two points in a year when autoscalers and cost dashboards were common tools, which suggests the bottleneck is organizational. Developers set generous requests to avoid pager alerts. Platform teams leave them alone. Finance sees one line item per cluster.

Why does 5% GPU utilization matter more?

A CPU node idling at 8% wastes a few dollars an hour. A GPU node idling at 5% wastes a rate finance teams now track hourly. The report states: "GPU utilization across the analyzed Kubernetes clusters was 5%."

The waste is structural. Inference services reserve whole accelerators for bursty models. Training jobs hold devices between epochs. A GPU is not shared across pods by default, so a half-used device often counts as fully used. The FinOps Foundation's 2025 State of FinOps report found 63% of organizations tracking AI spend, up from 31% the previous year. Tracking is not controlling. It marks where AI spend became a monthly finance review.

What are the providers charging for idle capacity?

All three managed services bill for worker nodes regardless of pod usage. Published figures:

• Amazon EKS standard control plane: $0.10 per cluster-hour; extended support $0.60, so old versions cost six times more.
• Provisioned Control Plane tiers: $3.40 (2XL) and $13.90 (8XL) per cluster-hour.
• Azure AKS agent nodes: billed as ordinary Azure VMs; reservations apply automatically.
• AKS on Azure Local: no additional AKS charge, though host costs apply.

AWS gives each version 14 months of standard support plus 12 months extended. For a company with 200 clusters on old versions, the $0.50 surcharge is pure waste. Microsoft documents a node_cpu_usage_percentage metric, so AKS teams can measure utilization without third-party tools. Microsoft also cites savings of up to 65% versus pay-as-you-go for Savings Plans under program conditions — a vendor-stated figure with a wide range.

What changed in Kubernetes 1.37?

Kubernetes 1.37.1, released September 15, 2026, is the newest release. The project supports 1.37, 1.36 and 1.35; version 1.34 reached end of life on October 27. AWS's EKS support notice of October 2 highlights three utilization-relevant changes: the Metrics API metrics.k8s.io/v1 is generally available; Dynamic Resource Allocation (a scheduler mechanism for sharing devices) gained taints and tolerations; and Horizontal Pod Autoscaler scale-to-zero is in beta, enabled by default. DRA observability, which lets operators see which device a pod claimed, is generally available and on by default. Version 1.38 entered beta November 4, with secondary coverage listing December 16 availability — a date the project has not confirmed.

On November 9, AWS announced EKS Auto Mode and Karpenter support for DRA on NVIDIA GPUs and EFA devices, enabling GPU sharing across pods in most Regions excluding GovCloud and China. AWS published no utilization target for the feature. It gives teams a lever; whether they pull it is another matter. AWS also cites a Deloitte finding of 51% to 70% TCO reduction for Auto Mode — a vendor-selected study with a wide range.

Google moved in parallel. GKE dynamic slicing reached general availability October 21. On November 9, Google announced Cloud Pathways GA, TPU 8tx and TPU 8i Max. Each provider is adding ways to split accelerators, an implicit admission that whole-device allocation wastes money.

What is the FinOps Foundation doing?

The Foundation published its 2026 FinOps Framework on March 19, adding Executive Strategy Alignment as a capability. FOCUS 1.4 (a billing-data standard), ratified June 4, adds 47 columns and a Price Sheet dataset. FOCUS 1.5 is in development, scoped to surface AI model identity and token consumption, input and output. The Foundation also published "The Experimental AI Spend Endgame" on October 30. AI Conformance moved into SIG Architecture on October 29, with a Workload Conformance specification targeted within two to three release cycles.

The market logic is simple. Right-sizing is a revenue risk for hyperscalers and a revenue opportunity for vendors like CloudBolt, which announced bill-accurate GPU allocation at KubeCon. KubeVirt graduated November 10. Finance leaders gain the most: with 63% of organizations tracking AI spend, 2027 budgets will start with utilization questions platform teams have rarely had to answer.