The pattern is familiar. The AWS account was set up quickly to get a product out the door, a few people have had admin access over the years, and nobody’s job is to read the invoice. Each month the number is a little higher, and each month there’s a reasonable-sounding explanation.
Most of that growth doesn’t come from the product getting more popular. It comes from things that were started and never stopped, sized for a guess and never revisited, or left on default settings that charge forever. None of it needs a rewrite to fix.
This article is for managers who sign off on the bill and want to know where to look first. We run our own products on AWS and pay for them out of our own revenue, so we read our invoices closely. You won’t find a promised savings percentage here; every account is different, and the only honest number is the one your own bill shows after the cleanup.
1. Read the bill before you cut anything
Open Cost Explorer and group the last few months by service. Usually a handful of services make up most of the total, and that’s where your attention belongs. Then break the largest ones down by region and usage type. A line labelled NAT gateway bytes or data transfer tells you far more than “EC2-Other.”
Next, look at tags. Cost allocation tags such as environment, product or team let you see what production costs compared with staging, or one customer’s setup compared with another. They have to be activated in the billing console, and they only apply from activation onward, so switch them on now even if your tagging is patchy. A large untagged share is a finding in itself: it means nobody can say what that money pays for.
For managers: Ask for a breakdown by service, region and environment before you approve any cost project.
2. Idle machines and forgotten environments
Servers are often sized on launch day for traffic that never arrived, then left alone because nobody wants to touch production. CloudWatch shows how busy they really are. AWS Compute Optimizer suggests smaller sizes based on actual use, and the same check applies to container task sizes on ECS Fargate.
Forgotten environments are often the bigger find. A staging stack that runs around the clock but gets used a few hours a week. A demo built for one sales meeting. An experiment someone started in another region and never shut down. Check every region the account has used, because the console only shows one at a time. Non-production environments that only matter during working hours can be scheduled to stop overnight and on weekends.
For managers: Ask for a list of every running environment, who owns it and when it was last used.
3. Networking charges hide in plain sight
NAT gateways charge by the hour and for every gigabyte they process. If your private servers pull large files from S3 through a NAT gateway, you pay for that traffic on top of everything else. A gateway VPC endpoint for S3 or DynamoDB routes that traffic privately without the NAT processing charge, and it’s a configuration change that leaves your code alone.
Data transfer is the other quiet line. Traffic out to the internet is charged, and so is traffic between availability zones. Chatty services spread across zones, or a database in one zone serving app servers in another, add up. Since 2024, AWS also charges for public IPv4 addresses, so servers that don’t need a public address shouldn’t have one.
For managers: If NAT gateway or data transfer is near the top of your bill, ask why before you look anywhere else.
4. Storage that outlived its purpose
When a server is terminated, its disk can stay behind, depending on how it was set up. Unattached EBS volumes keep billing. So do old snapshots and machine images that a backup script creates and nothing ever deletes. In S3, buckets without lifecycle rules keep every object at the same storage class forever, versioned buckets keep every old version, and incomplete multipart uploads sit around unseen.
Logs deserve their own look. New CloudWatch log groups keep data indefinitely unless someone sets a retention period. A debug log level left on in production, plus no retention, means you pay every month to store chatter nobody will read. Pick a retention per log group that matches what you actually need for support and audit, and archive the rest to cheaper storage if it must be kept.
For managers: Set log retention and S3 lifecycle rules once, and ask for a monthly list of unattached volumes and old snapshots.
5. Databases sized for a launch that never came
Managed databases are often the largest single line after compute. A production instance picked for peak load that never materialized, a Multi-AZ standby on a development database, provisioned IOPS nobody needed, or a read replica that no application queries: each one keeps billing whether it’s used or not.
Be careful with storage. RDS storage can grow automatically, but you can’t shrink allocated storage in place, so oversizing it is expensive to undo. Resizing the instance class, on the other hand, is a routine change you can plan for a maintenance window. Look at CPU, memory and connection metrics over a few weeks, including your busiest days, before choosing the new size.
For managers: Review database size against several weeks of real metrics, and keep high-availability features for the environments that need them.
6. Commit only after the cleanup
Savings Plans and Reserved Instances give you lower rates in exchange for a one- or three-year commitment. Compute Savings Plans apply across EC2, Fargate and Lambda, which makes them flexible if your architecture is still moving. Reserved Instances cover services like RDS. They’re worth having once usage is steady.
The order matters. Commit before the cleanup and you lock in today’s waste at a discount for years. Do the cleanup first, let usage settle for a month or two, then commit to the baseline you’re confident about, and leave peaks on regular pricing.
For managers: Approve commitments only after a cleanup, and only for the steady baseline you’re sure you’ll still be running.
7. Keep it down without a rewrite
Almost everything above is configuration: sizes, schedules, endpoints, retention, lifecycle rules, commitments. A rewrite or a new architecture is sometimes the right call, for example moving a spiky workload to Lambda, but it should follow from what the bill shows rather than from a vendor’s preference.
To stop the number from creeping back, set AWS Budgets with alerts that reach a person, turn on Cost Anomaly Detection, and give one named person a monthly review. Keep infrastructure in Terraform so environments can be torn down and rebuilt on demand, and require tags in code so new resources are attributable from day one. When we review an account for a client, we read the bill, the architecture and the pipeline, then put what can go, what it touches and in what order on paper before anything changes.
For managers: Name an owner, set budgets and alerts, and review the bill monthly. Most of the savings come from that habit.
The short version
A climbing AWS bill usually comes from idle or oversized resources, forgotten environments, networking charges, leftover storage, logs without retention and databases sized for a guess. Read the bill by service and tag, clean up, let usage settle, then commit to the baseline. Keep it in check with budgets, alerts and a monthly review. None of that requires rewriting your software.