Start with visibility, not cuts
Most cloud waste is invisible rather than intentional: resources nobody remembers launching, environments that run all weekend, and storage that only ever grows. Before changing anything, make spend understandable.
- Tag resources by product, team and environment, and enforce tagging with policy
- Enable cost reports grouped by those tags
- Set budgets and anomaly alerts so surprises are caught within days
Remove what you do not need
Quick wins usually come from cleaning up:
- Unattached volumes, old snapshots and unused load balancers
- Idle development and test environments; schedule them off out of hours
- Over-provisioned logging and metrics retention
- Data transfer patterns that cross regions or zones unnecessarily
Right-size before you commit
Look at actual utilization of compute and databases over a representative period and move to smaller or newer instance types where headroom is consistently unused. Do this before buying commitments, so you do not lock in oversized capacity.
Use commitments and spot capacity deliberately
Once your baseline is stable, Savings Plans or Reserved Instances can substantially reduce the cost of steady workloads. Spot capacity suits fault-tolerant work such as batch jobs, CI runners and some Kubernetes node pools. Mix them according to how predictable each workload is.
Kubernetes and containers
Clusters hide waste behind abstraction. Check that:
- Pod resource requests reflect real usage
- Cluster autoscaling removes empty nodes
- Costs are allocated per namespace or team
- GPU nodes for AI workloads scale to zero when idle
Make it a habit
Cost optimization is not a one-off project. A short monthly review of the biggest changes in spend, owned jointly by engineering and finance, keeps costs aligned with growth, and builds the shared language that makes trade-offs easier.
Related services