Kubecost on EKS

The AWS bill says what a cluster’s EC2 instances cost. It does not say which team, namespace or deployment is responsible. Kubecost fills that gap: it combines resource usage from the cluster with pricing to show cost per Kubernetes object, and how much of what you pay for is actually used.

It is built on OpenCost, the CNCF open-source cost model, and adds a UI, savings recommendations and longer retention. AWS distributes an EKS-optimized bundle as an add-on at no extra licence charge.

How it calculates cost

metrics (kube-state-metrics, cAdvisor, node labels)  ─┐
                                                      ├─► cost model ─► cost per container, per hour
prices (public list prices, or your actual bill)     ─┘

For each container, per hour:

cost = max(requested, used) CPU × CPU price
     + max(requested, used) memory × memory price
     + storage, load balancers and network attributed to it

Using the larger of request and usage matters: a pod that requests four cores and uses one still reserves four, and is charged for four. The difference between requested and used is reported as idle cost.

Three kinds of cost need a decision about who pays:

Cost	What it is	Options
Idle	Node capacity nobody requested	Show separately, or spread across tenants
Shared	kube-system, ingress controllers, monitoring	Split evenly, or in proportion to each tenant’s usage
Out-of-cluster	RDS, S3 and other resources a team uses	Attribute by AWS tag

Decide these with finance before publishing numbers; they change each team’s figure noticeably.

Install

aws eks create-addon --cluster-name my-cluster --addon-name kubecost_kubecost

Or with Helm, from the EKS-optimized chart in Amazon ECR Public:

helm upgrade --install kubecost \
  oci://public.ecr.aws/kubecost/cost-analyzer \
  --namespace kubecost --create-namespace \
  -f https://raw.githubusercontent.com/kubecost/cost-analyzer-helm-chart/develop/cost-analyzer/values-eks-cost-monitoring.yaml

It installs the cost model, a bundled Prometheus and the UI, and needs a persistent volume (EBS CSI). Reach the UI with a port-forward to begin with:

kubectl port-forward -n kubecost deploy/kubecost-cost-analyzer 9090

Do not expose it publicly: it reveals the whole cluster’s structure. Put it behind an internal load balancer with authentication.

To avoid running a second Prometheus, point Kubecost at an existing one or at Amazon Managed Prometheus.

Making the numbers accurate

Out of the box, Kubecost uses public On-Demand list prices. Real bills differ because of Spot, Savings Plans, Reserved Instances and negotiated discounts.

• Cloud billing integration. Give Kubecost read access to the Cost and Usage Report through Athena, and it reconciles its estimates against what you were actually charged, typically a day or two behind.
• Spot data feed for accurate Spot prices.
• Node labels for instance type, zone and capacity type must be present — they are by default on EKS.

Without billing integration, treat the absolute numbers as indicative and the relative numbers — which namespace costs most — as reliable.

What to look at

View	Question it answers
Allocations	What does each namespace, deployment, label or team cost over time?
Assets	What do the nodes, disks and load balancers cost?
Efficiency	How much of what each workload requests does it use?
Savings	Where are over-sized requests, underused nodes, orphaned volumes, idle workloads?
Budgets and alerts	Has a team exceeded its budget, or did spend jump unexpectedly?

The allocation view grouped by a team label is the report most organisations actually need. It only works if workloads carry that label consistently — enforce it with admission policy.

Acting on recommendations

Kubecost suggests request sizes from observed usage. Treat them as input, not instructions:

• Look at a window long enough to include peaks — at least a week, ideally a full business cycle.
• Leave headroom on memory; an under-sized memory request ends in an OOM kill.
• Apply changes through the normal deployment path, one workload at a time, and watch latency and restarts.
• Recommendations for nodes assume pods can move; check PodDisruptionBudgets first.

The full optimisation sequence is in EKS cost optimization.

Useful API calls

Everything in the UI is available over HTTP, for dashboards and chargeback reports:

# Cost by namespace for the last 7 days
curl -G http://localhost:9090/model/allocation \
  -d window=7d -d aggregate=namespace -d accumulate=true
 
# Cost by team label, with idle shared proportionally
curl -G http://localhost:9090/model/allocation \
  -d window=30d -d aggregate=label:team -d shareIdle=weighted

Limits

Limit	Detail
Free tier scope	Single-cluster views and limited retention; a unified multi-cluster view is a paid feature
Network cost attribution	Needs an optional network-cost DaemonSet, and is approximate
Fargate	Supported, with less granular data
Accuracy without billing integration	List prices only
Its own footprint	The bundled Prometheus needs memory proportional to cluster size

Alternatives

Option	Notes
OpenCost	The open-source core: the same allocation model and API, a simpler UI, no recommendations
Split cost allocation data for EKS	AWS-native: pod-level cost in the Cost and Usage Report, no agent to run, a day behind
CloudWatch Container Insights	Usage, not cost — Container Insights
Commercial FinOps platforms	Multi-cloud and automation features

A reasonable minimum is split cost allocation data for monthly chargeback, plus OpenCost or Kubecost when engineers need a live view to act on.

Related

• EKS observability overview
• AWS cost management
• Kubernetes cost optimization
• Kubecost
• EKS Workshop: Kubecost

Across the wiki

• AKS FinOps, Cost Allocation, and Cloud Spend Optimization — Kubernetes cost (Azure)
• GKE Cost Optimization, FinOps, and GKE Cost Allocation Architecture — Kubernetes cost (GCP)
• GCP Cost Optimization, Committed Use Discounts (CUDs), and FinOps — Kubernetes cost (GCP)