Skip to content
Resources · Blog

Cloud Cost Optimization Checklist: AWS and Azure 

A cloud cost optimization checklist for AWS and Azure: visibility, right-sizing before commitments, idle environments, storage, logs and network.

November 14, 2026 5 min read
CloudCloud MigrationCost Estimation

Most of the waste in an AWS or Azure bill sits in five places: compute sized for a peak that rarely comes, non-production environments running around the clock, steady workloads paid at on-demand rates, storage and logs that are never deleted, and network charges nobody modeled. Work through them in a fixed order. Get visibility first, so every dollar has an owner. Right-size and schedule second. Buy commitments such as Savings Plans or reservations only after that, because a three-year commitment bought on an oversized fleet locks the waste in for three years. Storage, logs and network come next, and they often hold surprises larger than compute. This is the checklist we work through in a cloud cost review, a two-week engagement that ends with a ranked list of changes, each with its estimated saving and its risk. Most items on it are configuration changes, not re-architecture, so a team with access to its own billing data can run the same checklist without outside help.

Start With Visibility

You can't cut what nobody owns. Before changing a single instance, make every resource answer three questions: which service it belongs to, which environment it runs in, and which team pays for it. A tagging policy with those keys, enforced at creation time through AWS Organizations tag policies or Azure Policy, does most of the work; untagged resources should fail to deploy, not appear in a report a month later. Activate the tags for cost allocation so billing data can be grouped by them. Then set budgets per account or subscription with alerts going to the owning team, not only to finance, and turn on anomaly detection (AWS Cost Anomaly Detection, or cost alerts in Azure Cost Management) so a misconfigured job shows up within a day instead of on next month's invoice. The last step is a unit cost: cost per customer, per order or per thousand requests. A total bill that grows with revenue is healthy. A unit cost that grows is the signal to act.

Compute: Right-Size, Schedule, Then Commit

Right-size from data, not from instance names. Pull two to four weeks of CPU and memory utilization; on AWS, memory metrics need the CloudWatch agent, and without them you are guessing at half the picture. AWS Compute Optimizer and Azure Advisor give a starting list, and a person who knows the workload should approve each change. Next, schedule non-production. A development or staging environment running all week is billed for 168 hours; used during working hours, it needs about 50. Stopping it at night and over weekends removes roughly 70 percent of those instance hours, and it is often the fastest saving in the whole review. Try ARM-based instances, which both clouds offer and which are often cheaper per unit of work when your runtime and dependencies support them. Use spot capacity for interruptible work such as CI runners and batch jobs. On Kubernetes, the bill follows the resource requests in your pod specs, not actual usage, so compare requests with measured consumption and let a cluster autoscaler remove idle nodes. Only then commit: AWS Savings Plans, or Azure reservations and savings plans, sized to the steady baseline that remains, with Azure Hybrid Benefit applied if you already own Windows Server or SQL Server licenses.

Storage, Databases and Logs

Storage costs grow without anyone deciding they should. Delete unattached block volumes and managed disks left behind by terminated machines, and set a retention rule for snapshots instead of keeping every one forever. On AWS, move gp2 volumes to gp3, which is cheaper per gigabyte and lets you provision performance separately from size. Put lifecycle rules on object storage so data moves to infrequent-access or archive tiers on S3, or to the cool, cold and archive tiers on Azure Blob Storage, once it stops being read. Databases deserve the same right-sizing as compute, plus a few specific checks: development databases stopped outside working hours, read replicas that nothing reads, and provisioned capacity sized for a launch that ended long ago. Logs are the most common surprise. CloudWatch log groups keep data forever by default, Azure Log Analytics bills by the gigabyte ingested, and a service left at debug level in production can cost more in logging than in compute. Set retention on every log group, sample high-volume logs, and keep debug logging for the environments where someone reads it.

Network: The Charges Nobody Modeled

Network charges rarely appear in an architecture diagram and often appear on the invoice. On AWS, a NAT gateway charges for every gigabyte it processes, so private workloads pulling large volumes from S3 through a NAT gateway pay for traffic that a free gateway endpoint for S3 would carry. Chatty services spread across availability zones pay for the traffic between zones; where your availability requirements allow, keep tightly coupled services in the same zone or cache their calls. Data leaving the cloud for the internet is billed on both platforms, so serve static assets and downloads through a CDN with sensible cache headers. Cross-region replication, multi-region backups and analytics exports to another region all carry transfer costs on top of storage costs. For GCC workloads, data residency rules may require an in-country region such as AWS me-central-1 or Azure UAE North in any case, which is one more reason to choose regions deliberately at the start instead of discovering transfer charges after a migration. Our post on legacy modernization cost covers how to plan for this during a move.

The Checklist

  • Tag every resource with service, environment and owner, and block untagged deployments
  • Set budgets and anomaly alerts that reach the owning team within a day
  • Track one unit cost metric alongside the total bill
  • Right-size compute and databases from two to four weeks of CPU and memory data
  • Stop non-production environments outside working hours
  • Move interruptible work such as CI and batch jobs to spot capacity
  • Compare Kubernetes resource requests with measured usage, and autoscale nodes
  • Buy Savings Plans or reservations only for the baseline left after right-sizing
  • Delete unattached volumes and disks, and set snapshot retention
  • Move gp2 volumes to gp3, and add lifecycle tiers to object storage
  • Set log retention on every log group, and turn off debug logging in production
  • Route S3 traffic through gateway endpoints instead of NAT gateways
  • Review cross-zone, cross-region and internet egress charges by service

Make It a Monthly Habit

A one-off cost review produces a good quarter, then the bill drifts back, because new services ship with the same defaults the old ones had. Savings stay in place when someone owns them. Name an owner for cloud spend, review the bill against usage every month with the engineering leads in the room, and put cost checks into the delivery process, for example infrastructure defined in Terraform, where a reviewer sees an instance size change in a pull request before it ships. Commitments need their own calendar, since Savings Plans and reservations expire, and renewing them without a fresh look repeats last year's sizing. If your team doesn't have time to run this, a managed operations retainer can include the monthly review; that is part of what our cloud team does for clients after a migration. If a migration is still ahead of you, build the tagging, budgets and schedules into the landing zone on day one, which costs far less than retrofitting them across a running estate.

Have a project that needs this kind of thinking applied to it?