1. Start with visibility, not cuts
Before changing anything, make the bill readable. Enable the Cost and Usage Report, turn on cost allocation tags and group spend by account, service and tag. If you cannot say which client or project drives each line, you will struggle to prioritize.
In Cost Explorer, sort by service and by usage type. In most agency accounts, EC2, RDS, data transfer and EBS account for 70–80% of the bill. That is where to focus first.
- Enable the Cost and Usage Report and cost allocation tags
- Agree a minimum tag set: project, environment, owner
- Identify the five usage types that make up most of the bill
2. Remove what nobody uses
Idle resources are the safest saving because removing them changes nothing for users. Look for unattached EBS volumes, old snapshots and AMIs, idle load balancers, unassociated Elastic IPs and stopped instances whose volumes still bill.
Always confirm ownership before deleting. A simple rule we use is: tag, notify, wait seven days, snapshot if in doubt, then delete.
3. Schedule non-production environments
Staging, QA and demo environments rarely need to run at night or at weekends. Scheduling them for office hours cuts their running time by roughly 65%. Instance Scheduler, Lambda functions or simple CI jobs can start and stop them automatically.
4. Right-size EC2 and RDS against real utilization
Use at least two weeks of CloudWatch data, ideally with the CloudWatch agent for memory metrics. Compare P95 CPU and memory against instance capacity. Instances that rarely exceed 30–40% are candidates for the next size down.
Move to current generations at the same time. Newer instance families are usually faster and cheaper per vCPU, and Graviton instances are typically 20–40% better on price-performance for compatible workloads.
5. Fix storage: gp3, lifecycle rules and tiering
Migrating EBS volumes from gp2 to gp3 typically saves around 20% with no downtime, and lets you provision IOPS separately from size. For S3, apply lifecycle rules to logs and build artifacts, and use Intelligent-Tiering for data with unpredictable access.
6. Cut data transfer and NAT gateway charges
NAT gateways charge per hour and per gigabyte processed. Traffic from private subnets to S3 or DynamoDB should use gateway VPC endpoints, which have no hourly charge. Review cross-AZ traffic between chatty services, and cache static and API responses with CloudFront where appropriate.
7. Commit to the clean baseline
Only after right-sizing should you buy commitments. Compute Savings Plans apply across EC2, Fargate and Lambda and across instance families, making them the safest choice for most agencies. Use RDS and ElastiCache reserved instances for databases. Commit to the stable floor of usage, not the average, and consider no-upfront terms to protect cash flow.
8. Use Spot for interruptible work
CI runners, batch processing, rendering and stateless workers can run on Spot capacity at steep discounts. Use mixed-instance Auto Scaling groups and handle interruption notices gracefully.
9. Set CloudWatch log retention
New log groups keep data forever by default. Set retention per log group, filter noisy logs in the application and export long-term archives to S3 if you need them for compliance.
10–12. Keep it down: budgets, anomaly detection and reviews
Set AWS Budgets per account or project with alerts at 80% and 100%. Turn on Cost Anomaly Detection. Hold a 30-minute monthly review of the top movers. These three habits prevent most savings from eroding.
Where AWS savings typically come from
| Lever | Typical saving on that line | Risk |
|---|---|---|
| Idle resource clean-up | 100% of the idle spend | Low |
| Non-production scheduling | 60–70% | Low |
| EC2 right-sizing and Graviton | 20–45% | Medium |
| gp2 to gp3 | ~20% | Low |
| VPC endpoints and NAT consolidation | 30–60% of NAT cost | Low |
| Savings Plans and reservations | Up to ~40% vs on-demand (varies by term) | Medium |
| Spot for interruptible work | Up to ~70–90% vs on-demand | Medium |
Checklist
- Cost and Usage Report and tags enabled
- Idle volumes, snapshots, IPs and load balancers removed
- Non-production on schedules
- EC2 and RDS right-sized and on current generations
- EBS on gp3; S3 lifecycle rules in place
- VPC endpoints for S3 and DynamoDB
- Savings Plans sized to the clean baseline
- Budgets, anomaly detection and a monthly review
