Cloud Cost Optimization: Find Waste in AWS, Azure and Google Cloud

Cloud Cost Optimization: Find Waste in AWS, Azure and Google Cloud

Cloud cost optimization for AWS, Azure and Google Cloud. Find unused resources, check continuing charges and calculate savings with a worked billing example.

Cloud cost optimization starts with the bill: identify what you pay for, establish who needs it, and measure the effect of one approved change at a time. Remove genuinely unnecessary resources, match retained capacity to demand, then evaluate commitments against the usage that remains.

Reviewed 16 October 2026: Rechecked provider billing, storage, recommendation, commitment and budget documentation, including Google Cloud spend-cap limits. The worked savings example uses explicit assumed rates and continuing costs.

TL;DR:

  • A stopped server can leave storage, networking and contractual charges behind.
  • An unattached disk is a review candidate, not automatic permission to delete it.
  • Budget alerts provide visibility. An alert alone does not cap spending.
  • Calculate savings after replacement costs and existing commitments.

If the cloud terminology is unfamiliar, start with IaaS, PaaS and SaaS explained. For choosing a platform, use our AWS, Azure and Google Cloud comparison.

Start With an Explainable Bill

Export a complete billing period and group charges by account or subscription, service, region and environment. Record whether you are looking at usage charges, amortized commitment costs or the cash invoice. Keep taxes, credits, currency conversions and support charges identifiable.

Give each significant charge an owner and a business purpose. Useful labels include the application, environment and person responsible for deciding whether it should continue. A name such as test-old is a clue to investigate, not evidence that the resource is disposable.

Pair resource ownership with an AWS IAM permissions review so the people and workloads that can change the infrastructure have the access they actually need.

Compare spending with the work delivered: orders processed, active customers or another workload measure relevant to your business. A growing total bill with a stable cost per order may have a different explanation from rising cost with unchanged demand.

Keep a short review record:

FieldExample of what to record
Resource and scopeVolume ID, account, region and environment
Current chargeInvoice amount and billed unit for the chosen period
Business ownerPerson who can confirm its purpose
EvidenceAttachment history, usage metrics, deployment configuration or backup policy
Proposed changeRemove, schedule, resize or retain
Approval and recoveryWho agrees, what must be preserved and how to recover
VerificationExpected saving, actual later charge and application checks

This record makes the next decision easier: keep resources whose purpose is established, investigate those whose purpose is unclear, and change only the items with an agreed plan.

Why Charges Continue After a Server Stops

Cloud resources have separate lifecycles. Stopping a VM, deleting a disk and ending a spending commitment are different actions.

ProviderBilling behavior to checkPrimary source
AWSEBS volumes preserved after an EC2 instance terminates continue to incur charges. Deleting a volume does not remove its snapshots.Preserved volumes, EBS volume deletion
AzureA stopped but allocated VM is billed for instance usage. A deallocated VM is not billed for that usage, but disks and some networking resources can still incur charges.VM states and billing
Google CloudStopping a Compute Engine instance retains resources such as attached disks, which can continue to incur charges.Stopping a Compute Engine instance

Use the detailed billing report to identify the continuing meter. Repeatedly stopping the VM will not remove a disk charge. Likewise, reducing resource usage may not lower the current cash bill if a contractual commitment still applies.

Find Unused Resources Without Losing Needed Data

Start with unattached volumes, retired development environments, old snapshots and resources left by previous deployments. Check every relevant account and region. A view of one region cannot establish that an entire estate has been reviewed.

For each candidate:

  1. Establish ownership. Find the team, deployment or former project that created it.
  2. Check dependencies. Review infrastructure configuration, scheduled jobs, recovery procedures and any services that still reference it.
  3. Check retention. Confirm whether a backup, contractual requirement or recovery objective requires the data.
  4. Choose a recovery method. Where a snapshot or export is needed, verify that it is usable and record its retention and access requirements.
  5. Approve the change. Record the resource, scope, responsible person and planned action.
  6. Verify the result. Check application behavior and the later billing period, allowing for billing delays and any retained resources.

For EBS, AWS documents deletion from the available state and explains that a snapshot can be used to recreate a deleted volume. It also notes that a matching Recycle Bin rule can retain a deleted volume. Include that retention behavior in the cost review. EBS deletion procedure.

Snapshots have their own lifecycle and storage charges. Removing one incremental EBS snapshot does not necessarily remove every block it referenced, because another snapshot may still need those blocks. How EBS snapshots work.

For recovery planning, see why the 3-2-1 backup rule still works. Cost cleanup should preserve the recovery arrangements your business actually needs.

Calculate the Saving Before Making the Change

Use consistent units, currency, billing periods and workload assumptions:

Net recurring saving = comparable baseline cost
  - expected cost after the change

Simple payback periods = one-off implementation cost
  / positive net recurring saving per period

Include replacement storage, monitoring, backup, licensing and operating costs. A discount against a public on-demand price is not necessarily a saving against your existing bill.

Worked Example With Assumed Rates

Consider an illustrative 30-day period with 720 hours. All amounts below are USD, excluding tax, credits and currency conversion. These are assumed rates for demonstrating the calculation; use your own invoice rates for a real decision.

Assume the owner confirms that a retired 500 GB development disk can be replaced by a retained snapshot with 250 GB of billed snapshot storage. An independent development VM can run for 220 hours instead of 720. Productive resources remain unchanged. There are no minimum storage charges, Recycle Bin charges or existing commitments affecting this example.

Line itemBaselineAfter the approved changes
Other retained resources$200.00$200.00
Development VM at $0.10/hour720 × $0.10 = $72.00220 × $0.10 = $22.00
Retired disk at $0.08/GB-month500 × $0.08 = $40.00$0.00
Retained snapshot at $0.05/GB-month$0.00250 × $0.05 = $12.50
Total per comparable period$312.00$234.50
Compute saving = (720 - 220) × $0.10 = $50.00
Storage saving after replacement = $40.00 - $12.50 = $27.50
Net recurring saving = $312.00 - $234.50 = $77.50

With $300 one-off implementation cost:
Simple payback = $300 / $77.50 = about 3.87 comparable periods
Net saving over 12 identical periods = 12 × $77.50 - $300 = $630.00

The snapshot’s billed storage is an explicit input, not an assumption that every snapshot is half its source volume. Scheduling also removes only the VM usage charges shown here. Add any other continuing costs to your calculation.

After the change, check the next complete billing period against the same workload and prices. Separate the effect of the change from traffic growth, expiring credits and unrelated price changes.

Right-Size and Schedule Against Real Demand

A low average CPU percentage alone does not prove that a smaller instance will work. Review memory, disk throughput and latency, network limits, application response times and important peak periods. Include reporting jobs, backups and recovery activity in the observation window.

AWS Compute Optimizer has resource-specific data requirements. EC2 memory information needs an appropriate metrics source, such as the CloudWatch agent, rather than an assumption that all guest metrics are already available. Recommendation requirements, EC2 metrics analyzed.

For a proposed resize, record the performance your business needs, test a representative workload and define a way to return to the previous configuration. A recommendation is a starting point for that decision.

For development scheduling, agree when engineers, CI jobs and integrations need the environment. Calculate the actual hours removed, identify persistent charges and test startup dependencies. The worked example saves 500 compute hours; it does not establish a universal percentage saving for non-production systems.

Review Storage and Network Costs Together

A cheaper storage class can cost more for your access pattern once retrieval, requests, transitions and minimum storage duration are included. Check restore time as well as price. Amazon S3 pricing, Azure Blob access tiers, Google Cloud Storage classes.

For networking, identify where bytes move: between components, zones, regions and the internet. Include gateways, load balancers, logging and CDN charges in the proposed design. A CDN can change both origin traffic and the services you pay for, so compare the complete arrangement.

Preserve availability and recovery requirements when evaluating fewer regions or replicas. A smaller bill produced by removing required resilience is a different service, and should be presented as that decision rather than equivalent optimization.

Buy Commitments Against the Usage You Will Retain

Evaluate commitments after cleanup, scheduling and right-sizing. Use a conservative retained baseline, check how often eligible usage falls below the commitment, and include plans to migrate or close applications.

AWS Savings Plans have different scopes. Compute plans cover eligible EC2, Fargate and Lambda usage, while EC2 Instance plans commit to a family and region. AWS also documents Database and SageMaker AI plans. Do not assume one plan covers every cloud service. Current AWS plan types.

Apply the same discipline to Azure savings plans and Google Cloud committed use discounts: check eligible resources, term, existing discounts and the effect of unused commitment before purchase.

Keep proposed savings separate so you do not count the same usage twice. If an instance will be removed, do not also claim a right-sizing saving and a new commitment saving for its original full-month usage.

Use Budgets as Alerts and Assign a Response

Set actual-cost and forecast alerts at thresholds appropriate to your business. Give each alert a recipient and an action: inspect an unexpected increase, contact the owner, or invoke a separately tested control.

AWS warns that billing and notification delays can allow costs beyond the notification threshold. Azure budgets notify without stopping resource consumption. AWS budget behavior, Azure budget behavior.

Google Cloud offers both alerts-only budgets and separately configured spend cap budgets for eligible services. Its documented caps have a single-project, single-service scope, use gross estimated costs and are not enforced instantly. They do not pause ongoing fixed usage of persistent resources such as compute and storage. Check eligibility and operational effects before treating a cap as protection for an entire cloud bill. Google Cloud budget types, spend cap scope and limits.

Start with the provider’s billing reports and recommendations, then check the price of any additional monitoring or automation. For example, AWS charges for Cost Explorer API requests and optional hourly-granularity records. Include those charges when comparing reporting options. Cost Explorer pricing.

Make the Next Review Repeatable

After each complete billing period, reconcile the largest changes, review open cleanup decisions and check whether the expected savings appeared. Revisit commitments when demand or architecture changes, and investigate anomalies when alerts arrive rather than waiting for the monthly review.

A useful handover contains the baseline, owner-approved action list, calculation, recovery arrangements, and later billing and performance results. That gives your team a process it can repeat instead of a percentage promise it cannot inspect.

If you need help explaining an AWS bill or planning the changes, our AWS infrastructure service covers architecture and cost optimization. Contact us with the workload, relevant billing period and constraints so we can discuss the review’s scope.

Frequently Asked Questions

Why am I still charged after stopping a cloud server?
Compute, storage, networking and commitments have different billing rules. Azure can still bill a stopped but allocated VM. Persistent disks, snapshots and other retained resources can continue to cost money after compute stops. Check the resource state and the actual invoice line items.
Can I delete every unattached disk?
No. Unattached means it is not connected to a VM at that moment. Confirm its owner, recovery purpose, retention requirements and restore arrangements before approving deletion. A replacement snapshot can also create ongoing charges.
Do cloud budget alerts stop spending automatically?
An alert alone does not stop resources. Billing and notification delays can allow spending beyond a threshold. Any automated action needs its own scope, permissions, testing and recovery plan.
How much can my business save on cloud costs?
Estimate it from specific invoice line items and approved changes. Subtract replacement costs and implementation effort, account for existing commitments, and compare equivalent usage periods. The worked example shows the arithmetic without assuming a universal savings percentage.
Should I buy a Savings Plan before cleaning up resources?
First establish the usage you expect to retain after cleanup and right-sizing. A commitment can discount eligible usage while leaving you paying for capacity or spending you no longer need. Check the plan’s service, region and term restrictions.
Does low CPU usage prove that a server is oversized?
No. Review memory, storage performance, network traffic, application latency and peak demand too. Test the proposed size against a representative workload and define how to reverse the change.