Who Ordered the GPU?
Kubernetes spend comes in three buckets that are easy to confuse: idle, shared and unallocated. On a GPU cluster, Kubecost often puts the most expensive pods, the training jobs run by operators such as Kubeflow and KubeRay, in __unallocated__. Why that happens, the two ways a cost ends up owned by nobody, and how to find and fix them with Kubecost for Grafana.


