Cloud bills grow like gardens: quietly, and mostly with things nobody planted. When our monthly spend crossed a line nobody could justify, we spent one month on a diet instead of a migration.
What actually moved the needle
- Tag enforcement via SCP + a daily orphaned-resource report (unattached EBS, stale snapshots, idle ELBs)
- Rightsizing from real utilization data — 60% of instances were under 15% CPU
- Spot instances for stateless workers behind an ASG mixed-instances policy
- S3 lifecycle policies: infrequent access at 30 days, Glacier at 90
The unglamorous winner
Deleting things. A single forgotten dev RKS cluster running since 2023 cost more per year than the entire observability stack. Cost optimization is 20% clever architecture and 80% honest bookkeeping.