The platform team cost scorecard.

Kubernetes hides cost in two ways. Shared clusters mix many teams on the same nodes, so no invoice line maps to an owner. And the request model lets teams reserve far more CPU and memory than their pods use, so the cluster scales to capacity that sits idle. The scorecard attacks both: allocation metrics make spend attributable, efficiency metrics make waste visible. We unpack the underlying mechanics in the Kubernetes cost guide.

Why does request to usage ratio sit at the top?

Because it is the single clearest signal of waste in a cluster. The cluster autoscaler and Karpenter add nodes to satisfy pod requests, not pod usage. If a team requests two cores and uses half a core, the cluster still provisions for two, and you pay for the gap. Track actual usage as a percentage of requests at the cluster and namespace level, and aim to keep usage around 60 to 80 percent of requests. Below that band you are funding reserved capacity nothing touches; above it you risk eviction and throttling. Right setting requests, ideally with the vertical pod autoscaler in recommendation mode first, is usually the largest single win on a Kubernetes estate.

What does bin packing efficiency tell you?

Bin packing efficiency is the share of provisioned node capacity that is actually requested by scheduled pods. A cluster can have well sized requests and still waste money if pods pack poorly onto nodes, leaving stranded fragments of CPU and memory on each node that are too small to schedule anything else. Track the ratio of requested to allocatable capacity across the fleet. Low efficiency points to node shapes that do not match workload shapes, or a scheduler configuration that spreads pods when it should consolidate. The fix is node rightsizing and consolidation, not more nodes.

How do idle hours and commitment coverage fit?

Idle node hours capture capacity running with little or nothing on it, the classic weekend and overnight waste in non production clusters. Commitment coverage measures how much of the steady state node fleet is covered by the right instrument: AWS Savings Plans or Reserved Instances on EKS, Azure Reservations or the Azure Savings Plan on AKS, GCP Committed Use Discounts on GKE, and OCI Universal Credits on OKE. The base layer of nodes that always runs should be covered for the deeper discount; the variable layer belongs on Spot or on demand. Coverage and efficiency work together: there is no point committing to nodes you only need because requests are inflated, so fix the ratio first, then size the commitment.

What makes spend attributable on a shared cluster?

The last two metrics. Allocated versus unallocated spend tracks how much of the cluster cost maps to a namespace, team, or product, and how much falls into a shared bucket nobody owns. Shared costs such as the control plane, system add ons, and idle headroom will never reach zero, but unallocated spend above roughly 15 percent means your namespace and labelling discipline is slipping. The unit cost metric, cost per service or per thousand requests, ties the whole thing to value, so leadership sees efficiency improving even as absolute spend grows with the business. Showback turns these numbers into accountability without billing teams internally, which we cover in showback for platform teams.

The scorecard at a glance

Run the scorecard alongside a regular review so the numbers drive action rather than sit on a dashboard. The cadence and the fixes that follow each metric are laid out in the Kubernetes cost review playbook.

Stand up the scorecard with us

We build the platform team cost scorecard on your real cluster data, drive the early wins on requests and bin packing, and leave your team owning the numbers. We take zero provider commissions. Our guarantee: we reduce your cloud spend or we reimburse our service fee.