Cloud Cost Optimization: Where the Money Actually Goes
Key takeaway
Most cloud bills are 20–40% waste, concentrated in idle non-production environments, over-provisioned instances, unattached storage, and egress. Tagging, rightsizing, and scheduled shutdowns typically recover the bulk of it within a month — before any architectural change.
Cloud bills grow by accumulation, not by decision. Nobody approves a 40% overspend; it arrives one forgotten test environment at a time. The good news is that the waste is concentrated in a few predictable places.
Find out what you're paying for first
You cannot optimize an unattributed bill. Enforce tagging by environment, team, and service, then produce a cost-per-service view. In most organizations this exercise alone surfaces resources nobody can identify an owner for — which are usually safe to turn off.
The quick wins
- Shut down non-production environments outside working hours — a dev environment running 168 hours a week for 45 hours of use is a ~70% saving on that line.
- Rightsize compute against actual utilization, not the size someone picked during setup.
- Delete unattached volumes, old snapshots, and idle load balancers; they bill forever and serve nobody.
- Set lifecycle policies to move cold object storage to infrequent-access or archive tiers.
- Check egress — cross-region and internet data transfer is often a top-three line item and is frequently accidental.
Commitment discounts, carefully
Reserved instances and savings plans cut 30–70% off steady-state compute, but they're a one-to-three-year bet. Rightsize first, then commit to the stable baseline only, and keep variable load on on-demand or spot. Committing to an over-provisioned fleet locks your waste in for three years.
Structural changes worth planning
- Autoscaling that actually scales to zero for bursty workloads.
- Serverless for genuinely spiky, event-driven work — and not for steady high throughput, where it usually costs more.
- Caching in front of expensive queries and repeated egress; a CDN often pays for itself on bandwidth alone.
- Data retention policies — storing logs forever is a decision, whether or not anyone made it.
Make it stick
A one-off cleanup decays within two quarters. Put the cost-per-service dashboard in front of the engineering team monthly, set budget alerts with real thresholds, and include a cost line in architecture reviews. Engineers optimize what they can see.