Amazon EC2 Cost Optimization Guide for 2026
An eight-step EC2 cost optimization guide: tagging, waste removal, rightsizing, pricing models, scaling and storage. Do the steps in order.
EC2 is often the largest line on an AWS bill, and most savings come from a short list of steps done in the right order. This guide gives that order, so you cut cost without surprising your engineers or your uptime.
1. See what you spend and who owns it
You cannot optimise what you cannot see. Tag instances by team, application and environment, and use AWS Cost Explorer and Budgets to watch trends. Untagged spend is the first thing to fix, because nobody feels responsible for it.
2. Remove waste
Look for the easy items first:
- Stopped instances that still have attached EBS volumes
- Unattached volumes and old snapshots
- Idle load balancers and unused public IPv4 addresses, which AWS charges for
- Non-production environments that run all night and weekend
An approved operating schedule can avoid unnecessary compute runtime; compare it with other options using the workload’s actual dependencies and costs.
3. Rightsize
Move over-provisioned instances to a smaller or better-suited type, using Compute Optimizer and your own metrics. Look at memory and peaks, not only average CPU. Our rightsizing playbook shows how to do this safely.
4. Choose the right instance family
Compare newer generations with the current workload using measured performance and current prices. AWS Graviton (Arm) instances are a candidate for compatible workloads; test the artifact and compare the complete cost before changing architecture. Match the family to the job: compute, memory or general purpose.
5. Use the right pricing model
| Model | Best for | Trade-off |
|---|---|---|
| On-Demand | Unpredictable or short workloads | No term commitment; compare effective cost |
| Savings Plans | Steady usage you can commit to | One or three year commitment |
| Reserved Instances | Steady, specific instance usage | Less flexible |
| Spot | Fault-tolerant, interruptible work | AWS can reclaim the instance |
AWS publishes the current discount ranges for each model, so check them before you plan. Buy commitments only after rightsizing, and cover the steady base, not your peak.
6. Scale with demand
Auto Scaling groups add and remove instances as load changes, so you stop paying for peak capacity all day. Check that scaling rules react fast enough and do not scale in too aggressively.
7. Watch storage and data transfer
EBS volumes, snapshots and network traffic cost money even when instances are cheap. Review volume types and sizes, and check cross-zone and cross-region traffic. See data transfer costs in common AWS architectures and cross-zone data transfer cost.
8. Review every month
Costs drift as teams launch new systems. Set a monthly review, track cost per customer or per transaction, and use the cloud cost drift tool to spot changes early.
Start here
If you only have a week, do steps 1 to 3. Choose reversible changes with adequate evidence; deletion and production resizing can be high risk and need their own review. For a structured review of your account, see our cloud cost optimization service or talk to our team.
Worked example: separate waste from required headroom
Consider a fictional application with a customer-facing API, a nightly worker and a development environment. All three appear under EC2-related spending, but they have different availability requirements. The API must remain available, the worker has a completion deadline and the development system may have an approved operating schedule. The first review therefore creates three workload boundaries rather than applying one shutdown or sizing rule to the entire account.
Use a common period and accounting basis. In an illustrative month, compute is $900, attached storage is $180, snapshot storage is $40 and network-related charges are $130, for a total of $1,250. These are invented amounts for arithmetic, not AWS rates or an observed bill. If a proposed schedule avoids $100 of compute but leaves those other lines unchanged, the modeled workload reduction is 8%, not a saving on every EC2-related category.
Before scheduling the development system, ask about overnight tests, remote users, patching and backup windows. A stopped EC2 instance can leave storage and other resources chargeable. The EC2 On-Demand pricing page is a primary reference for applicable charge families and billing conditions. Check the actual configuration and Region rather than applying a remembered price to a whole environment.
Next, separate idle resources from recovery assets. An unattached volume may be abandoned, or it may be retained for a documented recovery purpose. Require an owner, a retention decision and a safe recovery path before deletion. The absence of recent application traffic is useful evidence for investigation, but it is not permission to delete the only recoverable copy of a dataset.
Use a staged change register
| Candidate | Evidence before action | Verification after action |
|---|---|---|
| Development schedule | Approved hours and dependency review | Jobs complete and resources restart when needed |
| Instance resize | Representative demand and compatible artifact | Objectives hold under the same workload |
| Storage cleanup | Owner and retention decision | Required data remains recoverable |
| Commitment purchase | Stable eligible hourly demand after changes | Coverage and utilization match the chosen scope |
| Scaling-policy change | Demand profile and startup delay | Launches complete before queue or latency limits |
Treat each candidate as a hypothesis. A schedule might be more useful than a commitment for a short-lived environment, but only if shutting it down is operationally acceptable. A larger newer instance might complete a batch faster and reduce total runtime, while a smaller one could miss the deadline. Compare cost per completed job and required capacity, not just the displayed hourly compute price.
Compute Optimizer's resource requirements describe the metrics needed for different recommendation types. Use those recommendations alongside application evidence, not as a substitute for it. A quiet recent period can miss a launch or a seasonal enrollment window. Document that gap and choose an observation or test that represents the actual risk before applying the suggested configuration.
Model commitments after the operating shape is known
A Savings Plan exchanges a spending commitment for discounted eligible usage. It does not turn an unnecessary resource into necessary demand. If scheduling and resizing materially reduce the steady base, rerun the commitment analysis using the changed footprint. Include existing commitments, expiry dates and expected migrations when deciding how much additional spend to cover.
Use an hourly demand distribution rather than a monthly average alone. A service that consumes heavily on weekdays and almost nothing at weekends can have the same monthly total as an evenly used service, while presenting a different commitment risk. An illustrative $2 per hour commitment creates an obligation across the term's hours; consuming above that on busy days does not automatically recover unused eligible coverage from a quiet hour.
Keep pricing and capacity separate. EC2 Capacity Reservations address matching capacity in a specific Availability Zone under their documented conditions. A discounted spending arrangement does not by itself ensure that the needed instance can launch. If a major event needs assured capacity, review that requirement explicitly rather than maximizing a discount percentage and assuming resilience follows.
Detect failure modes in the optimization itself
Compare service quality before and after a change. A lower bill accompanied by a larger retry queue, more failed jobs or an expensive manual recovery may be a poor result. Monitor a representative workload period and record external causes such as a demand drop or a feature removal. Avoid attributing every month-over-month billing reduction to a resize when the business processed fewer transactions.
Cost allocation can change without the infrastructure changing. A tag correction can move spend between teams, and amortization can move the apparent timing of a commitment. Preserve the chosen cost metric and scope in the report. If finance uses cash invoice totals while engineering uses amortized workload cost, reconcile those views instead of treating one as automatically wrong.
Set explicit recovery triggers for each technical change. For a schedule, the trigger may be a required overnight job that cannot complete. For a resize, it may be sustained latency or job duration beyond the agreed boundary. For a storage deletion, recovery must have been established before action. Do not describe destructive cleanup as the universally lowest-risk first step.
Next action: one workload, one verified improvement
This guide is a sequence for review, not a complete account assessment or a promise of a particular saving. Specialized hardware, licensed software, regulated retention and hard recovery requirements can change the order. The workload owner should record those constraints before selecting a candidate. Confirm current pricing and service behavior before any purchase or production operation.
Start with one workload and produce a change register containing the proposed action, supporting evidence, expected financial effect, service objective, owner and rollback condition. Select a reversible candidate with representative evidence, verify the result and then review the next one. That produces a repeatable operating process without confusing activity, projected savings and actual financial outcomes.
Related services
AWS Cost Optimization & FinOps Consulting
Make your cloud bill explainable, forecastable, and defensible to finance. We audit your spend, eliminate waste, and build cost governance that holds.
Cloud Reliability Assessment & Production Architecture Review
Cloud reliability assessment covering production architecture, disaster recovery and incident readiness. Identify failure modes and prioritize resilience work.