{"slug": "kubernetes-clusters-use-8-cpu-5-gpu-finops-hit-2026", "title": "Kubernetes Clusters Use 8% CPU, 5% GPU: FinOps Hit [2026]", "summary": "Cast AI's 2026 Kubernetes optimization report found average production cluster CPU utilization fell to 8% in 2025 from 10% the prior year, memory utilization dropped from 23% to 20%, and GPU utilization stood at 5%. The FinOps Foundation's 2025 State of FinOps report separately found 63% of organizations tracked AI spend, up from 31% a year earlier. Cast AI sells automation software and the 8% figure is not confirmed by Kubernetes, AWS, Microsoft or Google documentation, so the report is an industry signal rather than a universal constant.", "body_md": "The average production Kubernetes cluster used just 8% of its CPU in 2025, down from 10% the year before, according to Cast AI’s [2026 Kubernetes optimization report](https://cast.ai/reports/kubernetes-optimization-report/). Memory slid from 23% to 20%, and GPUs sat at 5%. For any team now paying for AI inference nodes, those three numbers define the Kubernetes cost optimization problem of late 2026.\n\nThe timing matters. Kubernetes 1.37 shipped with new device-allocation features, Amazon and Google pushed fresh GPU controls within the past three weeks, and the FinOps Foundation is rewriting its billing standard to count tokens. Teams that already ran a tool like [Kubecost](https://shattered.io/kubecost-setup-kubernetes-cost-optimization-2026/) have the visibility. Most still lack the habit of acting on it.\n\nDon't miss new tech stories on Google\n\nAdd Shattered once in the Google app and our stories show up in your news suggestions.\n\n[Add now](https://www.google.com/preferences/source?hl=en&q=https%3A%2F%2Fshattered.io%2F)\n\n## What the Cast AI report says about Kubernetes CPU utilization\n\nCast AI analyzed production clusters for its yearly optimization study. Its headline finding, in the report’s own words: “Average CPU utilization across analyzed production Kubernetes clusters fell to 8% in 2025, down from 10% the prior year.” The company published it in the 2026 edition of the study.\n\nMemory followed the same direction. Cast AI reported that “Average memory utilization across analyzed production Kubernetes clusters fell from 23% to 20%.” A cluster can look healthy on a dashboard because pods are scheduled and nothing is throttled, yet most of the paid capacity sits idle. Requests and limits, not real demand, drive what a node costs.\n\nTwo caveats belong next to these figures. First, the numbers come from a vendor that sells automation, and its sample may not resemble your fleet. Second, none of the official Kubernetes, AWS, Microsoft or Google pages reviewed for this story confirm the 8% number, and the cloud providers do not publish a fleet-wide utilization figure of their own. Treat 8% as a strong industry signal, not a universal constant.\n\nEven with those caveats, the direction is clear. Utilization fell by two points in a year when autoscalers and cost dashboards were already common tools. That suggests the bottleneck is organizational. Developers set generous requests to avoid pager alerts, platform teams leave them alone, and finance sees one line item per cluster rather than per service.\n\n## GPU utilization at 5% changes the AI FinOps math\n\nThe third number in the report deserves the most attention. Cast AI wrote that “GPU utilization across the analyzed Kubernetes clusters was 5%.” A CPU node that idles at 8% wastes a few dollars an hour. A GPU node that idles at 5% wastes a rate that finance teams now track by the hour.\n\nGPU waste has a structural cause. Inference services reserve whole accelerators for models that may serve bursts of traffic. Training jobs hold devices between epochs, and notebooks hold them overnight. Unlike CPU, a GPU is not shared across pods by default without extra scheduling work, so a half-used device often counts as fully used by the scheduler.\n\nThe FinOps Foundation’s data shows how quickly the spend category grew into a governance problem. Its 2025 State of FinOps report found that 63% of organizations were tracking AI spend, up from 31% the previous year, per the Foundation’s [published report data](https://data.finops.org/2025-report/). Tracking is not the same as controlling, but it marks the point where AI moved from an experiment budget to a line that finance reviews every month.\n\nFor teams that call hosted models instead of running their own, a proxy layer can capture some of the savings. A [Cloudflare AI Gateway setup](https://shattered.io/cloudflare-ai-gateway-setup-2026/) adds caching and per-key cost tracking in front of LLM calls, which attacks the token bill rather than the node bill. Self-hosters have the opposite problem. They own the idle silicon.\n\n## How EKS, AKS and GKE bill for idle capacity\n\nAll three managed Kubernetes services charge for the worker nodes whether or not pods use them. The control plane fee, where one exists, is small next to a fleet of 8%-utilized virtual machines. The table below lists the charges that AWS and Microsoft publish with numbers. The Google pricing pages reviewed for this story did not yield dated figures, so GKE appears without a price.\n\n| Service | Item | Published charge | Idle-cost implication | \n|---|---|---|---|\n| Amazon EKS | Standard support control plane | $0.10 per cluster-hour | Fixed fee, plus worker nodes billed separately | \n| Amazon EKS | Extended support control plane | $0.60 per cluster-hour | Standard fee plus $0.50, so old versions cost six times more | \n| Amazon EKS | Provisioned Control Plane, 2XL tier | $3.40 per cluster-hour | Pays for control-plane headroom, not workload demand | \n| Amazon EKS | Provisioned Control Plane, 8XL tier | $13.90 per cluster-hour | Only worth it for very large clusters | \n| Azure AKS | Free, Standard and Premium tiers | Regional prices not confirmed | Premium adds 24-month long-term support | \n| Azure AKS | Agent nodes | Billed as ordinary Azure VMs | Reservations apply automatically to AKS nodes | \n| Azure AKS on Azure Local | AKS component | No additional AKS charge | Azure Local host costs still apply | \n\n### Amazon EKS: the version-support tax\n\nAWS gives each Kubernetes version 14 months of standard support, then 12 months of extended support, for up to 26 months in total. The cluster-hour fee jumps from $0.10 to $0.60 once a version enters extended support. For one cluster, that is a modest difference. For a company with 200 clusters stuck on old versions, it adds up fast, and it is pure waste because it buys no extra capacity. Teams comparing node purchasing options can start with a look at [EC2 on-demand, Spot and reserved pricing](https://shattered.io/aws-ec2-pricing-on-demand-vs-spot-vs-reserved-2026/), since the node layer is where most of the money goes.\n\n### Azure AKS: VMs first, then tiers\n\nMicrosoft documents that AKS agent nodes bill as regular Azure virtual machines and that VM reservations apply to them. The company also points customers to Savings Plans, Reserved Instances and Azure Hybrid Benefit, and its documentation reportedly cites savings of up to 65% versus pay-as-you-go for Savings Plans under the program’s conditions. Its monitoring reference documents a `node_cpu_usage_percentage` metric that aggregates by node and node pool, so an AKS team can measure its own utilization today without a third-party tool.\n\nCPU time is a line item everywhere in the stack, not only in container requests. Even authentication code carries a price, as the [HMAC versus SHA-256 comparison](https://shattered.io/hmac-vs-sha256-2026/) showed with its 2x CPU cost. Multiply small costs by an 8% utilization rate and the waste compounds.\n\n### Google GKE: capacity controls and AI runtimes\n\nGoogle updated its GKE documentation on granular resource limits on November 6, 2026. Its Agent Sandbox page, dated August 9, reportedly described that feature as available at no extra charge, with the underlying resources still billable. GKE also gets the most aggressive AI roadmap of the three, covered below.\n\n## Kubernetes 1.37 gives platform teams new utilization tools\n\nKubernetes 1.37.1, released September 15, 2026, is the newest release on the project’s official list. The project currently supports three minor branches: 1.37, 1.36 and 1.35. Version 1.34 reached its listed end of life on October 27, 2026, so any cluster still on it receives no upstream security fixes. Teams that want a safe place to test the new release can follow the [kind 1.37 local cluster tutorial](https://shattered.io/kubernetes-tutorial-kind-1-37-local-cluster-12-steps-2026/) before touching production.\n\nAWS’s notice of Kubernetes 1.37 support on EKS, dated October 2, 2026, highlights three changes tied to utilization. The Metrics API `metrics.k8s.io/v1` is now generally available. Dynamic Resource Allocation gained device taints and tolerations. And Horizontal Pod Autoscaler scale-to-zero is in beta and enabled by default. Scale-to-zero matters for cost because an idle inference service can drop to no pods at all.\n\nKubernetes documentation dated October 14, 2026 also describes Dynamic Resource Allocation observability as generally available in 1.37 and on by default. That is the piece that lets operators see which device a pod claimed, a prerequisite for billing GPU time to the team that used it. Pod-level resource managers, by contrast, remain beta and disabled by default.\n\nAnyone planning an autoscaling rollout can use the [Kubernetes HPA autoscaling walkthrough](https://shattered.io/set-up-kubernetes-hpa-autoscaling-2026/) as a starting point. The next minor release, Kubernetes 1.38, entered beta on November 4 with v1.38.0-beta.0, and secondary coverage lists general availability for December 16, 2026. That date is not firm until the project confirms it.\n\n## EKS Auto Mode and Karpenter add GPU-aware scheduling\n\nOn November 9, 2026, AWS announced that EKS Auto Mode and Karpenter now support Dynamic Resource Allocation for NVIDIA GPUs and Elastic Fabric Adapter devices. AWS says the capability allows GPU sharing across pods, topology-aligned allocation and dynamic EC2 scaling, and it is available in most AWS Regions excluding GovCloud and China.\n\nThis answers the 5% GPU number directly. If two small models can share one accelerator, the scheduler no longer needs to burn a whole device for each. AWS did not publish a utilization target for the feature, and the announcement does not claim it will fix fleet-wide waste. It gives teams a lever. Whether they pull it is a separate question.\n\nAWS has also pushed Auto Mode as a cost story. Its product page cites a Deloitte finding that EKS Auto Mode reduced total cost of ownership by 51% to 70% against self-managed and standard EKS deployments. That is a vendor-selected study and the range is wide, so buyers should test it against their own workloads. Related changes from October 28 include Gateway API support in Auto Mode and an [EKS Dashboard update for fleet-wide certificate rotation visibility](https://aws.amazon.com/about-aws/whats-new/2026/10/amazon-eks-dashboard-fleet-wide-visibility-certificate-authority-ca-rotation/), which comes at no additional charge.\n\nFor the alerting side of the bill, a budget guardrail still matters more than any scheduler feature. The [AWS Cost Explorer alerts guide](https://shattered.io/aws-cost-explorer-budgets-anomaly-detection-alerts-12-steps-2026/) covers anomaly detection that catches a runaway GPU job before the invoice does.\n\n## GKE doubles down on TPUs, slicing and distributed inference\n\nGoogle’s announcements cluster around accelerator efficiency. Reports dated November 9, 2026 say Cloud Pathways became generally available on GKE, described by Google as a distributed machine-learning runtime that scales workloads across TPUs from a single controller. GKE dynamic slicing was announced as generally available on October 21. Google also announced TPU 8tx and TPU 8i Max on November 9.\n\nAt KubeCon + CloudNativeCon North America on November 10, Google presented an expanded AI-native Kubernetes stack that reportedly includes OpenRL, a Kubernetes-native post-training service, and llm-d 1.0, a distributed-inference control plane. Both are described as CNCF Sandbox projects. These items rest on a single source each, so readers should confirm details on Google’s own pages before building plans around them.\n\nThe pattern is consistent across the providers. Each is adding ways to split, slice or share accelerators, which is an implicit admission that whole-device allocation wastes money. Dynamic slicing and DRA-based sharing both target the same inefficiency from different directions. Google’s approach leans on its own TPU hardware, while AWS leans on NVIDIA GPUs and Elastic Fabric Adapter networking.\n\n## The FinOps Foundation moves from cloud bills to token bills\n\nThe FinOps Foundation published its 2026 FinOps Framework on March 19, 2026. Its biggest addition is Executive Strategy Alignment, a new capability in the Manage the FinOps Practice domain, alongside deeper guidance on FinOps Scopes and more technology categories. The group also widened its mission in February 2026 from managing cloud costs to managing the value of technology, according to its announcement.\n\nThe billing standard is changing too. FOCUS 1.4 was ratified June 4 and announced June 10, 2026, adding two datasets, 47 columns, six attributes, 17 glossary entries and two supported features. A Price Sheet dataset standardizes provider list-price data for price analysis and commitment optimization. Work on FOCUS 1.5 has begun, and its stated scope includes surfacing AI model identity and token consumption, both input and output, in the cost and usage dataset.\n\n| Date | Event | Why it matters for utilization and cost | \n|---|---|---|\n| March 19, 2026 | FinOps Framework 2026 published | Adds Executive Strategy Alignment and sharper Scopes guidance | \n| June 4, 2026 | FOCUS 1.4 ratified (announced June 10) | Price Sheet dataset, 47 new columns, reconciliation support | \n| Oct. 2, 2026 | EKS adds Kubernetes 1.37 | HPA scale-to-zero beta, GA Metrics API, DRA device taints | \n| Oct. 27, 2026 | Kubernetes 1.34 reaches end of life | Clusters must move to 1.35, 1.36 or 1.37 to stay in upstream support | \n| Oct. 29, 2026 | AI Conformance moves into SIG Architecture | Standardizes how AI workloads behave across providers | \n| Nov. 9, 2026 | EKS Auto Mode and Karpenter add DRA for NVIDIA GPUs | Enables GPU sharing and topology-aware placement | \n| Planned | FOCUS 1.5 in development | Token and model data in billing exports | \n\nThe Foundation also published “The Experimental AI Spend Endgame: Allocation, Governance, and Lifecycle Management” on October 30, 2026. The title tells the story. Teams that ran AI pilots on loose budgets now have to decide which experiments become products and which get shut off. Google Cloud’s own cost push, covered in this look at [Google Cloud FinOps and GPU commitments](https://shattered.io/google-cloud-finops-flexible-gpu-commitments-focus-1-4-2026/), points in the same direction.\n\n## Azure’s long-term support and Kubernetes standards work\n\nMicrosoft’s AKS pricing structure now has three cluster-management tiers: Free, Standard and Premium. Premium adds Microsoft maintenance beyond community support and 24-month long-term support for Kubernetes versions. AKS Automatic clusters are preconfigured to use the Standard tier at creation. Microsoft’s documentation, updated November 7, 2026, shows no regional prices in the pages reviewed here, so budget owners should check the Azure pricing calculator directly.\n\nAKS also reaches beyond the public cloud. Its hybrid releases on Azure Local carry Kubernetes versions through 1.34.4 in the 2610 release, and Microsoft states that AKS is included with Azure Local at no additional AKS charge. That helps edge deployments, though the underlying Azure Local host costs still apply. Microsoft’s [edge compute option guide](https://learn.microsoft.com/en-us/azure/architecture/guide/technology-choices/choose-kubernetes-edge-compute-option) lays out the choices.\n\nOn the standards side, the Kubernetes project posted on October 29, 2026 that its AI work moved from a working group into SIG Architecture: AI Conformance. The post identifies Janet Kuo, a Senior Staff Software Engineer at Google, as a Kubernetes maintainer. A separate Workload Conformance Working Group announced on October 27 targets an initial stable specification within two to three release cycles. Conformance matters for cost because it lets buyers move AI workloads between clouds without a rewrite, which strengthens their bargaining position.\n\n## The tooling market reacts to idle clusters\n\nVendors spotted the same gap Cast AI measured. CloudBolt announced GPU optimization and bill-accurate GPU allocation for Kubernetes at KubeCon North America, according to its [launch announcement](https://www.cloudbolt.io/company/news/cloudbolt-launches-gpu-optimization/). The phrase “bill-accurate” signals that chargeback, not just visibility, is the sales pitch. If finance cannot attribute a GPU hour to a team, no one owns the waste.\n\nSiliconANGLE’s KubeCon coverage on November 10 carried a headline about AI operations expanding to Kubernetes upgrades as teams shrink. The theme is that smaller platform teams are leaning on automation to handle upgrades and right-sizing. The CNCF also announced KubeVirt’s graduation on November 10, a governance milestone that matters to teams consolidating virtual machines and containers onto one cluster to raise density.\n\nInfrastructure-as-code teams have another checkpoint. A pull-request gate that estimates cost before merge, like the one in the [Infracost Terraform cost gate tutorial](https://shattered.io/infracost-terraform-cost-gate-tutorial-2026/), stops oversized node groups before they exist. Prevention beats cleanup, because cleanup needs someone to risk an outage by shrinking a request.\n\n## Market impact: who gains when clusters get right-sized\n\nRight-sizing is a revenue risk for cloud providers and a revenue opportunity for the vendors that sell it. If a customer lifts CPU utilization from 8% to 30%, it needs fewer nodes for the same work. Hyperscalers sell more AI capacity than ever, so they can afford the trade, and each is building the tools itself. Auto Mode, AKS Automatic and GKE’s slicing features all package efficiency as a managed feature.\n\nIndependent FinOps vendors face a harder question. When the provider’s own dashboard shows utilization for free, as the EKS Dashboard does at no charge, third-party tools must offer cross-cloud views, chargeback or automation to justify their fee. The FOCUS standard helps them by normalizing billing data, and FOCUS 1.5 token fields would extend that to LLM spend.\n\nFinance leaders gain the most from serious Kubernetes cost optimization. Cloud spend used to grow with traffic, which made it hard to question. AI spend grows with experiments, which makes it easy to question. With 63% of surveyed organizations now tracking AI spend as a distinct category, 2027 budget conversations will start with utilization questions that platform teams have rarely had to answer.\n\n## Competitive comparison: which cloud offers the most cost levers\n\nThe three clouds offer different levers rather than a clear winner. The comparison below uses only features that the providers’ own pages or announcements describe. Where a cell says “not confirmed,” the reviewed sources did not state it.\n\n| Lever | Amazon EKS | Azure AKS | Google GKE | \n|---|---|---|---|\n| Managed auto-provisioning | EKS Auto Mode with Karpenter | AKS Automatic | Not confirmed in reviewed sources | \n| GPU sharing support | DRA for NVIDIA GPUs (Nov. 9, 2026) | Not confirmed | Dynamic slicing GA (Oct. 21, 2026) | \n| Long-term version support | 26 months total, extended costs $0.60 per hour | 24-month LTS on Premium | Not confirmed | \n| Utilization metrics | Metrics API GA in 1.37, EKS Dashboard | node_cpu_usage_percentage metric | Granular resource limits docs (Nov. 6, 2026) | \n| Hybrid or edge option | EKS Anywhere, Hybrid Nodes | AKS on Azure Local, Edge Essentials | Not confirmed | \n| Accelerator runtime news | EFA and placement groups for Auto Mode | Not confirmed | Cloud Pathways GA, TPU 8tx and 8i Max | \n\nAWS leads on managed GPU scheduling for NVIDIA hardware. Google leads on TPU-first tooling. Microsoft has the most flexible support policy for slow-moving enterprises and the clearest hybrid story. Teams that also weigh self-managed options can compare distributions in the [K3s vs K8s distribution breakdown](https://shattered.io/k3s-vs-k8s-kubernetes-distributions-rke2-openshift-talos-2026/), and the wider [cloud computing coverage](https://shattered.io/category/cloud-computing/) tracks the provider announcements as they land.\n\n## Historical context: from the density promise to idle fleets\n\nKubernetes won enterprise adoption by promising density. Pack many containers on fewer machines, and the infrastructure bill shrinks. For years that argument held for stateless web services. It weakened as teams added staging clusters, per-team clusters, and safety buffers on every request.\n\nFinOps emerged as the discipline that argued with those buffers. Its early work centered on tagging, showback and commitment purchases. The 2026 Framework update shows how far it has moved: Executive Strategy Alignment is a capability about boardroom priorities, not tag hygiene. The mission change in February 2026 to managing the value of technology completes the shift.\n\nAI changed the stakes in two steps. First, GPU-backed services pushed per-node costs well above CPU-only services, so a 5% utilization rate hurts more. Second, token-based pricing from model APIs created a spend category that no node-level tool sees. The Foundation reportedly announced at FinOps X in June 2026 an intention to form a Tokenomics Foundation, and FOCUS 1.5 aims to bring tokens into the billing export.\n\nSupport policy also tightened. AWS’s 14-plus-12-month schedule and the $0.50 extended-support surcharge turned version drift into a billing event. Kubernetes 1.33 reached end of life on June 28, 2026, and 1.34 followed on October 27. Each cutoff pushes laggard clusters toward either an upgrade project or a bigger invoice.\n\n## Five predictions for Kubernetes and AI FinOps in 2027\n\nThese are analyst judgments, not reported facts, and each rests on the trends above.\n\n1. **GPU sharing becomes the default for inference.** With DRA support landing on EKS and slicing generally available on GKE, platform teams will stop reserving whole accelerators for small models.\n2. **FOCUS 1.5 ships with token data, and major providers adopt it.** Vendors such as Oracle already support earlier FOCUS versions, and customers will ask for comparable AI line items.\n3. **Utilization reporting moves into the cloud consoles.** Providers already publish node-level metrics, and a fleet-level utilization score is a small step that would pressure customers and competitors alike.\n4. **Extended-support fees drive upgrade projects.** With versions retiring every few months, finance will start asking why clusters pay $0.60 per hour instead of $0.10.\n5. **The next Cast AI-style report shows a smaller gap.** If scale-to-zero and DRA features see real adoption, the 8% CPU figure should stop falling, though a rebound to healthy levels in one year is unlikely.\n\n## A Kubernetes cost optimization plan for this quarter\n\nStart with measurement, then policy. Pull your own numbers before accepting an industry average, since the 8% figure may be better or worse than your fleet. The table lists the checks that map to the features above.\n\n| Check | What to measure | Action if it fails | \n|---|---|---|\n| CPU request versus usage | Per-namespace ratio over 30 days | Lower requests and use vertical or horizontal autoscaling | \n| Memory request versus usage | Peak working set against limits | Right-size with headroom, avoid hard OOM risk | \n| GPU occupancy | Busy time per device, not allocated time | Share devices with DRA or slicing where supported | \n| Version age | Clusters outside the 1.35 to 1.37 range | Schedule upgrades before extended-support billing hits | \n| Idle services | Pods with near-zero traffic for 14 days | Enable scale-to-zero or retire the service | \n| AI spend attribution | Cost per team, model and environment | Add allocation tags and review monthly | \n\nSequence matters. Lowering requests before you have good metrics can trigger outages and make developers distrust the exercise. Lead with the safest wins: delete idle clusters, move old versions off extended support, and turn on scale-to-zero for non-critical services. Then tackle GPU sharing, which carries the largest dollar upside and the highest risk.\n\n## Frequently asked questions\n\n### What is the average Kubernetes CPU utilization in 2026?\n\nCast AI’s 2026 report found that average CPU utilization across analyzed production clusters fell to 8% in 2025, down from 10% the prior year. Memory utilization fell from 23% to 20%. The data comes from a vendor study, not an official cloud-provider statistic.\n\n### How low is GPU utilization on Kubernetes?\n\nThe same Cast AI report put GPU utilization across the analyzed clusters at 5%. That means most paid GPU capacity sat idle, which is why scheduling features like Dynamic Resource Allocation and GPU slicing are getting attention.\n\n### How much does Amazon EKS charge for the control plane?\n\nAWS lists $0.10 per cluster-hour for a Kubernetes version in standard support and $0.60 per cluster-hour in extended support. Worker nodes and other AWS resources cost extra. Provisioned Control Plane tiers cost more, with the 2XL tier at $3.40 and the 8XL tier at $13.90 per cluster-hour.\n\n### What is the newest Kubernetes version?\n\nKubernetes 1.37.1, released September 15, 2026, is the newest release on the project’s official list. The project supports 1.37, 1.36 and 1.35. Version 1.38 entered beta on November 4, with general availability reported for December 16, 2026.\n\n### What does FOCUS 1.5 change for AI costs?\n\nThe FinOps Foundation says FOCUS 1.5 is scoped to surface AI model identity and token consumption, input and output, in the cost and usage dataset. It is in development and not yet released, so teams cannot rely on it today.\n\n### Does AKS charge for the control plane?\n\nAKS offers Free, Standard and Premium cluster-management tiers. Microsoft’s reviewed documentation confirms the tier structure but not current regional prices. Agent nodes bill as regular Azure VMs, and AKS on Azure Local carries no additional AKS charge.\n\n### Where should a team start with Kubernetes cost optimization?\n\nMeasure request versus real usage per namespace for 30 days, retire clusters on versions past end of life, and turn on scale-to-zero for services that sit idle. GPU sharing comes after those basic steps because it carries higher risk.", "url": "https://wpnews.pro/news/kubernetes-clusters-use-8-cpu-5-gpu-finops-hit-2026", "canonical_source": "https://shattered.io/kubernetes-cost-optimization-8-percent-cpu-5-percent-gpu-finops-2026/", "published_at": "2026-11-11 08:15:50+00:00", "updated_at": "2026-11-11 08:22:08.423430+00:00", "lang": "en", "topics": ["ai-infrastructure", "mlops", "ai-chips"], "entities": ["Cast AI", "Kubernetes", "FinOps Foundation", "Amazon", "Google", "Microsoft", "Kubecost", "Cloudflare"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/kubernetes-clusters-use-8-cpu-5-gpu-finops-hit-2026", "markdown": "https://wpnews.pro/news/kubernetes-clusters-use-8-cpu-5-gpu-finops-hit-2026.md", "text": "https://wpnews.pro/news/kubernetes-clusters-use-8-cpu-5-gpu-finops-hit-2026.txt", "jsonld": "https://wpnews.pro/news/kubernetes-clusters-use-8-cpu-5-gpu-finops-hit-2026.jsonld"}}