Zhivko Todorov
ALL CASE STUDIES

CASE 17 · COBALT · 2025

EKSKARPENTERSPOTBIN-PACKING

Karpenter, Spot, and the EKS bill that fell sixty percent.

An AdTech company ran 14 EKS clusters with Cluster Autoscaler, a fixed set of node groups, and an on-demand-only policy. The compute bill was $180k/month and the autoscaler routinely ran clusters at 38% utilisation. We replaced it with Karpenter, opened the cluster to Spot, and pulled utilisation past 70%.

INDUSTRY

AdTech

DOMAIN

COST

DELIVERED

2025

STACK

EKS·KARPENTER·EC2 SPOT·GRAVITON·PROMETHEUS·GOLDILOCKS

RESULTS

What changed, by the numbers.

COMPUTE BILL

−60%

$180K → $72K / MONTH

CLUSTER UTILISATION

72%

WAS 38%

SPOT INTERRUPTIONS

< 0.1%

OF POD-MINUTES

PROVISIONING TIME

−83%

KARPENTER VS CA

HOW IT WENT

Cluster Autoscaler was making over-conservative choices because the node groups were over-specified. Every workload had a `nodeSelector`. Every node group was sized for the heaviest workload that might land on it. Nobody had touched the limit ranges in a year.

Karpenter replaced both the autoscaler and the per-node-group thinking. We let it choose instance types from a pool of seventeen families (mix of x86 and Graviton), with constraints on architecture only where the workload required it. Spot was enabled with capacity-optimized allocation strategy and pod disruption budgets sized per workload.

Interruption rate landed at 0.07% of pod-minutes — well within the noise floor for the workloads. Goldilocks recommendations tightened pod resource requests in the second month, freeing more headroom. The 60% bill reduction is now the team’s most-cited slide in their internal architecture reviews.

READY WHEN YOU ARE

Let's get your AWS bill (and architecture) in order.

The discovery call is free. You walk away with at least one concrete idea — even if we never work together.

Or email directly →