Cutting Kubernetes Costs by 40%

Cloud bills are a debt that grows quietly. Before long, your Kubernetes cluster is paying for twice the resources you actually use. Last quarter we cut the bill by 40% — here's how.

1. End over-provisioning

Most teams inflate pod requests "just in case." The result: nodes look full, but CPU sits idle.

Measure actual usage

Set requests from data, not guesses:

kubectl top pods --all-namespaces --sort-by=cpu
# If p95 usage is 30% of the request, cut it in half

Aligning requests to just above p95 usage alone saved a quarter of the bill.

2. Reclaim idle resources

Running dev environments 24/7 is pure waste.

• Scale development clusters to zero outside working hours.
• Move batch jobs to spot/preemptible nodes.
• Delete orphaned LoadBalancer and disk resources.

An idle resource is a subscription you pay for but never use.

3. Pick the right machine type

General-purpose nodes are rarely the cheapest. Profile your workloads:

Workload	Previous type	Next type	Savings
API	general	compute-optimized	22%
Database	general	memory-optimized	18%
Batch	general	spot	60%

The takeaway

Cost optimization isn't a one-time project — it's an ongoing discipline. Once we automated these three interventions with panel-connected alerts, the savings became permanent.