AWS Cost Optimization: The Discipline That Actually Sticks
The initial cleanup is the easy part. The savings that last come from tagging, per-team budgets, and a monthly review, not a one-time audit.
Key takeaways
- Reduce an AWS bill in a fixed order.
- Delete orphaned resources, right-size EC2 and RDS, cap log retention, then commit to your baseline with Savings Plans, fix NAT and egress paths, tier S3, and stop non-prod overnight.
- Keep the savings with mandatory tags, per-team budgets, a monthly review, and cost estimates in pull requests.
On this page
To reduce an AWS bill, delete what does nothing, shrink what is oversized, and only then commit to Savings Plans for the baseline that remains. That first pass is the easy part. The savings hold only if tagging, per-team budgets, a monthly review, and pull-request cost estimates stop the waste from creeping back.
Verified againstChecked against official docs
Checked against the AWS Savings Plans User Guide, the AWS Compute Optimizer User Guide, the Amazon VPC gateway endpoints docs, and the CloudWatch Logs User Guide. Checked October 2026.
AWS savings that hold are the ones backed by process, not a single sweep.
The first pass, in order of return on effort
Sort the work by payoff per hour, not by how the console groups services. The top items are free and reversible; the bottom ones need a diagram or a meeting.
- Delete orphaned resources. Nothing on AWS cleans itself up. Unattached EBS volumes, old snapshots and the AMIs on top of them, unassociated Elastic IPs, and load balancers with no healthy targets all keep billing. Snapshot anything you're nervous about, then delete.
- Right-size EC2 and RDS. Opt in to Compute Optimizer, which analyzes the last 14 days of CloudWatch metrics by default and covers EC2, Auto Scaling groups, EBS, Lambda, and RDS and Aurora, among others. Databases are where fear-driven over-provisioning hides longest, so check CPU, freeable memory, and connection counts before dropping an instance class.
- Cap logging ingestion and retention. CloudWatch Logs keeps log data indefinitely by default, so every log group without a retention setting grows forever. Set retention per group, and sample high-volume, low-value logs before they ship to a third-party tool.
- Commit to your baseline with Savings Plans. Only after steps 1 to 3, or you commit to paying for waste. AWS cites savings of up to 72% on compute. Compute Savings Plans apply across instance family, size, OS, tenancy, and Region, and also cover Fargate and Lambda.
- Fix NAT and egress paths. Private subnets reaching S3 or DynamoDB through a NAT gateway pay processing charges for traffic that never needed it. Gateway endpoints for those two services carry no additional charge. Chatty services split across availability zones also pay for transfer, so keep tightly coupled services zone-aware.
- Tier S3 and add lifecycle rules. Unpredictable access goes to Intelligent-Tiering; known-cold data gets a lifecycle rule to Glacier; incomplete multipart uploads get an abort rule. Our S3 cost traps post covers the request-charge side.
- Stop non-prod on nights and weekends. Dev, staging, and QA rarely need to run around the clock. Tag them, and have an EventBridge schedule stop the tagged set in the evening and start it in the morning.
The commands for the first three steps:
# 1. Unattached EBS volumes and unassociated Elastic IPs
aws ec2 describe-volumes --filters Name=status,Values=available \
--query 'Volumes[*].[VolumeId,Size,CreateTime]' --output table
aws ec2 describe-addresses \
--query 'Addresses[?AssociationId==`null`].[PublicIp,AllocationId]' --output table
# 2. Compute Optimizer's right-sizing verdicts (after opting in)
aws compute-optimizer get-ec2-instance-recommendations \
--query 'instanceRecommendations[*].[instanceArn,currentInstanceType,recommendationOptions[0].instanceType]' \
--output table
# 3. Log groups with no retention, then set one
aws logs describe-log-groups \
--query 'logGroups[?retentionInDays==`null`].[logGroupName,storedBytes]' --output table
aws logs put-retention-policy --log-group-name /aws/lambda/my-func --retention-in-days 90
Why the bill grows back even when usage does not
Without discipline, AWS costs creep in predictable ways. New services get over-provisioned "to be safe." Reserved capacity gets renewed at a higher level because consumption crept up. Snapshots, orphaned volumes, and idle load balancers accumulate. New AWS services get adopted without a cost pass. Engineers default to a bigger instance type because tuning is work nobody has time for. Each is invisible day to day, so a one-time cleanup only resets the clock.
Tagging is the foundation, not a nice-to-have
Every resource carries tags: team, env, service, cost-center. Without them, attribution is impossible and accountability evaporates into "the AWS bill went up" with nobody able to say why.
We enforce this three ways: organization-level tag policies that block untagged resources where the provider supports it, cost allocation tags activated in the billing console, and a monthly script that flags anything untagged for its owner. New resources get tagged at creation; the existing fleet got tagged in one batch cleanup. The payoff is that the cost dashboard, filtered by team, becomes that team's bill.
Budgets per team, and a monthly review
Set a budget per team each quarter from last quarter's actuals plus planned changes, not an arbitrary cut target, with an alert well before the limit. When a team overruns, the conversation is not punitive: what changed, and is it justified?
A short meeting every month, same agenda every time: total cost against last month and the same month last year, the top movers, anything spending more than expected, and last month's action items. This catches silent drift early. Once a quarter, one team does a deep-dive on its own services: right-sizing, spot, storage tiering, reservation coverage, and zombie resources that never produce a spike but cost money every month.
Catch it before it ships
New infrastructure gets a cost estimate in the pull request: expected monthly cost at baseline and at peak, the reserved-versus-on-demand call, and a comparison against similar services. We generate these from Terraform plans with Infracost, so the estimate lands as a comment on the diff. Reviewers can ask why something costs twice what its neighbors do, and often the resource gets right-sized before it ever runs.
Cost Anomaly Detection wired to each team's Slack catches the spikes the monthly review would find weeks later: a retry loop after a deploy, unbatched small S3 writes, a Lambda invoking itself. Once native tooling stops being enough, usually with multi-account or multi-cloud spend, work through a FinOps tools comparison before building your own.
Questions people ask
What is the fastest way to reduce an AWS bill?
Delete orphaned resources first: unattached EBS volumes, old snapshots, unassociated Elastic IPs, and load balancers with no targets. It is free, reversible if you snapshot first, and needs only the CLI. Then set retention on CloudWatch log groups, which keep data indefinitely by default, and right-size instances using Compute Optimizer's 14-day analysis.
Should I buy Savings Plans or Reserved Instances?
For compute, start with Compute Savings Plans, which apply across EC2 instance families, sizes, Regions, Fargate, and Lambda, so a later re-architecture does not strand the commitment. Size the commitment to the bottom of your usage graph, not the peak, and buy it only after cleanup and right-sizing. Prefer a one-year term for workloads you might redesign.
Do VPC gateway endpoints cost money?
No. AWS states there is no additional charge for gateway endpoints, which exist for Amazon S3 and DynamoDB. Routing private-subnet traffic for those services through a gateway endpoint instead of a NAT gateway removes the NAT data processing charge on that traffic, which makes it one of the cheapest network fixes available.
The decision, concretely
- Starting from zero? Run the first pass in order, then tag everything. None of the ongoing discipline works without attribution.
- Tempted to buy a Savings Plan first? Don't. Commit only to what is left after deleting and right-sizing.
- Deciding what to automate? Automate detection and alerting, not deletion. An orphaned volume sometimes holds data someone needs; alert the owner and escalate if nobody answers.
The call we'd make
Run the first pass this week in the order above, and leave NAT and egress for when you have a traffic diagram in front of you. Then treat cost as ongoing discipline, not a project with an end date: tags from day one, a monthly review, a quarterly deep-dive, and cost estimates at the pull request. Teams that skip the discipline repeat the same cleanup every year. Teams that keep it only do it once.
Sources
Cite this article
Kiril Urbonas (2024). AWS Cost Optimization: The Discipline That Actually Sticks. DevOpsNess. Updated 2026-11-20. https://www.devopsness.com/blog/aws-cost-optimization-strategies-reduce-cloud-bill
[AWS Cost Optimization: The Discipline That Actually Sticks](https://www.devopsness.com/blog/aws-cost-optimization-strategies-reduce-cloud-bill) (Kiril Urbonas, DevOpsNess, updated 2026-11-20)
More from Cloud
Explore more articles in this category
FinOps and Cloud Cost Management for Engineering Teams
Embed cost ownership in engineering: tags, budgets, and showback.
.NET 8 Support Ends November 10: The Lambda and Azure Functions Checklist
.NET 8, .NET 9, the Lambda dotnet8 runtime and the Azure Functions in-process model all lose support on November 10, 2026. Inventory commands, the isolated-worker migration and a CI gate.
PlanetScale Best Practices: Schema Changes, Limits and Connections
The PlanetScale defaults that matter in production: safe migrations on main, deploy requests with a real revert plan, queries shaped for Vitess limits, and scoped credentials.
You might have missed
Evergreen posts worth revisiting.