FinOps Right-Sizing Without Vibes: The 5-Number Recipe That Cuts $200k of Idle EC2

Short summary

EC2 fleets waste 77-88% of compute capacity, costing $150k-$250k annually. Use five CloudWatch metrics (p50/p99 CPU, p50/p99 memory, burst duration) to deterministically select instance types from a lookup table, replacing subjective sizing. The method composes with autoscaling (HPA, KEDA) and applies to Kubernetes node pools.

• •Typical EC2 fleet runs at 12-23% utilization; 77-88% is idle and continuously billed
• •Use 5 CloudWatch metrics over 30 days to eliminate guesswork from instance selection
• •T-family vs M-family decision hinges on burst duration; CPU credit math is critical

Generated with AI, which can make mistakes.