Loading page content
Enter the size of your cluster and how much CPU and memory is requested and actually used. The monthly bill is split into capacity doing work, capacity reserved by requests but unused, and capacity nothing has asked for.
SponsorThis spot is open to one DevOps or AI tooling company at a time. Ask about sponsoring
116 of 118 run entirely in your browser (Weather Widget uses Open-Meteo; What's My IP asks this site's server).
Share on XShare on LinkedInKubernetes clusters waste money in two different places, and they need different fixes. The first is request padding: a pod asks for 2 CPUs and uses 0.3, and the scheduler holds the other 1.7 for it whether or not it is ever used. Nobody is being careless; requests are set once, at launch, with a safety margin, and never revisited. The fix is right-sizing requests from observed usage, by hand or with the Vertical Pod Autoscaler in recommendation mode.
The second is capacity that no pod has requested at all: nodes kept around after a scale-down, or a node pool sized for a peak that comes once a month. The fix there is the cluster autoscaler or Karpenter, and node sizes that pack well. The calculator separates the two so you know which to work on, and estimates the node count you would need if requests tracked usage plus a headroom margin. Treat that as an upper bound on the saving: real clusters lose some capacity to system reservations and to pods that do not pack perfectly.
Enter your cluster size and how much CPU and memory is requested and actually used. You get the monthly cost split into what does work, what is padding in resource requests, and what sits empty.
0.192 is an m5.xlarge on demand in us-east-1.
Requested: sum of pod requests (kubectl describe nodes, Allocated resources). Used: a typical peak from kubectl top nodes or your metrics.
Room for spikes, rollouts, and losing a node.
$10,092 a year. Memory is the resource that sets the node count, so that is where right-sizing requests pays off first.
Assumptions: 730 hours per month, identical nodes, and the node price split evenly between CPU and memory. Control plane, storage, load balancer, and data transfer charges are not included. The node-count estimate ignores pod packing and system reservations, so treat it as an upper bound on the saving.