Kubernetes cloud bills often balloon due to over-provisioned pod requests, underutilized nodes, and misconfigured autoscaling rather than any single large mistake. Practical cost reduction starts with Kubernetes-aware cost allocation and labeling to make spend visible by team and workload. Right-sizing CPU and memory requests using 2–4 weeks of 95th-percentile utilization data from Prometheus or Grafana can cut requests by 30–50% on the first pass. Node right-sizing follows pod right-sizing, matching instance families to actual CPU/memory/storage blend. Autoscaling must be configured to scale down as well as up, with HPA tied to meaningful metrics and Cluster Autoscaler unblocked by properly configured PodDisruptionBudgets. Spot/preemptible instances offer 60–90% savings for fault-tolerant workloads. Long-term control requires namespace ResourceQuotas, LimitRanges, and embedding cost reviews into existing SRE and team rituals to prevent drift back to over-provisioning.
