Skip to content
Cloud cost optimization services · AWS · Azure · Google Cloud

Reduce AWS and Azure costs in the right order

A working guide to cutting a cloud bill without hurting production: what to switch off first, which tool to trust for rightsizing, how to buy Savings Plans and reservations under the 2027 rules, and a fixed-scope review if you would rather we did it.

Laptop screen showing bar charts and trend lines on a dark dashboard
At a glance
  • Levers rankedBy saving, production risk and engineer effort
  • Week-one quick winsChanges that need no redesign and no commitment
  • Commitments under 2027 rulesSavings Plans, reservations and the new exchange limits
  • When not to optimiseWhere the saving is smaller than the risk

How do you reduce AWS and Azure costs?

Work in this order. Delete idle resources and put non-production on a schedule. Upgrade anything paying extended-support fees. Rightsize with AWS Compute Optimizer or Azure Advisor using at least 30 days of data. Move disks, objects and logs to cheaper tiers, and cut NAT and data-transfer charges. Only then buy Savings Plans or reservations, sized to the steady baseline. Keep the savings with tags, budgets and anomaly alerts that reach the team that owns each resource.

Cloud cost optimization services

Most of a cloud bill is old decisions nobody went back to

Ask why an AWS or Azure bill has grown and the answer is rarely one dramatic mistake. It is a database left on a version that now attracts extended-support fees, a NAT gateway carrying traffic that should have gone through a free endpoint, debug logs ingested at full price, a test environment that never sleeps, and a support plan percentage nobody has looked at since it was set up. Each line is small. Together they are often the difference between a bill the board accepts and one it questions.

This page is the method we use, written so you can run it yourself. It ranks the levers by what they save, what they risk and how long they take; gives separate checklists for AWS and Azure, with a short note on Google Cloud; explains the commitment rules that change in January and February 2027; and ends with the cases where optimising is the wrong use of an engineer’s week. For the wider FinOps practice behind it (allocation, unit costs, the monthly rhythm) read our FinOps playbook; for cluster-level detail, Kubernetes cost control.

If you would rather hand it over, our cloud cost optimization services start with a fixed-scope review of one cloud, then a savings sprint in which our engineers make the changes through your own code review, and optionally a monthly retainer. The work sits inside our cloud infrastructure practice and pairs with managed services if you want someone to keep watching the bill afterwards. We sell no cloud capacity, take no reseller margin and hold no partner status with AWS, Microsoft or Google, so nothing we recommend earns us a rebate.

Before you cut: find out where the money actually goes

Open Cost Explorer on AWS or Cost analysis in Azure Cost Management, pick the last three full months, and group by service and then by usage type (AWS) or meter (Azure). One month hides trends; three show which lines are growing. Most bills are dominated by compute and databases, but the lines that most often surprise owners are smaller ones further down:

  • Extended-support charges on database engines and Kubernetes versions that have aged past standard support. They appear automatically and keep growing.
  • NAT gateway processing and data transfer, usually traffic to S3 or between Availability Zones that could take a free or cheaper path.
  • Log ingestion, where verbose application logs are billed per GB on the way in.
  • Public IPv4 addresses on AWS, charged hourly whether attached or idle.
  • Licence-included Windows and SQL Server, often on machines where the licence could be cheaper, smaller or unnecessary.
  • The support plan, which on AWS is a percentage of the rest of the bill.

Write down the run-rate of each line before you change anything. Savings are measured against that number, not against a forecast.

Cloud cost levers ranked by saving, risk and effort

This is the order we work in on a typical AWS or Azure estate. Vendor figures are the vendors’ own published maximums, checked on 7 November 2026; your result depends on your usage, Region and contract. Where a row shows arithmetic rather than a vendor figure, it says so.

#LeverWhat it is worthRisk to productionTypical effort
1Remove what nobody uses: unattached disks, stale snapshots, idle load balancers, spare public IPsEach public IPv4 address on AWS costs $0.005 an hour, about $3.65 a month, idle or not (Amazon VPC pricing)Low, once the tagged owner confirmsHours to two days
2Schedule non-productionArithmetic: running 12 hours on weekdays is 60 of 168 hours, so that environment’s compute falls by about 64%None for productionOne to two days
3Leave extended support on RDS, Aurora and EKSAWS’s own RDS example: $0.100 per vCPU-hour in the first two years past standard support, $0.200 from the third (US East (Ohio), RDS for MySQL pricing). EKS: $0.60 per cluster-hour instead of $0.10Medium: a major-version upgrade needs testingDays to a few weeks
4Rightsize instances, databases and containersDepends on how oversized you are; tools estimate it per resourceMedium: load-test the smaller size firstDays
5Move to Graviton (AWS, Arm-based)AWS says Graviton instances cost up to 20% less than comparable x86 instances (AWS Graviton)Medium: arm64 images and native dependenciesDays per service
6gp2 to gp3 EBS volumesAWS’s pricing example: $0.08 against $0.10 per GB-month, with 3,000 IOPS and 125 MB/s included on gp3 (EBS pricing)Low: the change is made onlineHours
7Object storage tiersS3 Intelligent-Tiering has no retrieval charges, only a per-object monitoring fee; objects under 128 KB never move tierLowHours
8Cheaper log ingestion and retentionCloudWatch Logs Infrequent Access is priced at half the Standard ingestion rate, with fewer featuresLow: keep alarm-driving logs on StandardHours to days
9NAT and data-transfer pathsNAT gateway: $0.045 an hour plus $0.045 per GB processed in AWS’s US East (Ohio) example. Gateway endpoints for S3 and DynamoDB carry no hourly or per-GB chargeLow to mediumDays
10Commitments for the steady baselineAWS: Compute Savings Plans up to 66%, EC2 Instance up to 72%, Database up to 35%. Azure: savings plan for compute up to 65%, VM reservations up to 72%, savings plan for databases up to 35%Financial, not technicalHours to buy; one to three years to live with
11Spot capacity for interruptible workUp to 90% off on AWS and on Azure, up to 91% on Google CloudInterruptions: two minutes’ notice on AWS, 30 seconds on AzureDays to weeks
12Licences: Windows, SQL Server, RHELRemoves or shrinks the licence part of the hourly price; see the licensing sectionCompliance risk if misappliedDays
13Support plan (AWS)Business Support+ is the greater of $29 a month per account or a sliding share of usage; see belowNone, if you keep cover where it mattersHours

Rows 1 to 9 come before row 10 for a reason: every one of them shrinks the usage you would otherwise commit to for years.

How to reduce AWS costs: a service-by-service checklist

Start with AWS’s own tools

  • Cost Optimization Hub gathers more than 18 kinds of recommendation (rightsizing, Graviton, idle resources, RDS and Aurora, Savings Plans and reservations) across the accounts in your organisation, and nets the estimates against your existing discounts. Use its numbers rather than list-price estimates.
  • Compute Optimizer has no charge of its own. By default it looks at 14 days of CloudWatch metrics, optionally 32; the paid enhanced infrastructure metrics extend that to about three months for roughly $0.25 per resource per month. For anything with a month-end peak, 14 days is too short.
  • Cost Anomaly Detection costs nothing extra, and accounts that opened Cost Explorer after 27 March 2023 already have a services monitor with a daily email. Add monitors per linked account or cost category and send alerts to the owning team’s channel through SNS, not to a finance inbox.

Compute

  • List instances from older generations; moving to the current generation of the same family is usually cheaper per unit of work and needs no code change.
  • Try Graviton first on stateless services written in interpreted or JVM languages and on managed services that offer Arm instances, such as RDS and ElastiCache. Native extensions and third-party agents are where arm64 moves stall.
  • Put Auto Scaling groups for development and staging on scheduled actions, or use the Instance Scheduler solution AWS publishes.

Storage

  • Convert gp2 volumes to gp3 in place. Check large volumes first: gp2 gives 3 IOPS per GB, so a gp2 volume above 1 TB has more baseline IOPS than gp3 includes, and you may need to provision some.
  • Delete unattached EBS volumes and set snapshot retention with Data Lifecycle Manager or AWS Backup, rather than by hand.
  • Use S3 Intelligent-Tiering for buckets whose access pattern you cannot predict, add a lifecycle rule to abort incomplete multipart uploads, and expire old object versions in versioned buckets.

Databases

  • Check engine versions now. RDS for MySQL 8.0 left standard support on 31 July 2026 and RDS for PostgreSQL 13 on 28 February 2026; instances still on them are being billed for Extended Support per vCPU-hour, at a rate that rises in the third year.
  • Database Savings Plans are a one-year, no-upfront commitment that applies only to the latest provisioned instance generations. Older generations need an upgrade (or a reserved instance) to get a discount.
  • Stop or scale down non-production databases outside working hours, and check whether serverless options fit databases that sit idle most of the day.

Network

  • Add gateway VPC endpoints for S3 and DynamoDB so that traffic stops paying the NAT gateway’s per-GB processing charge.
  • Look for chatty services in different Availability Zones; traffic between zones is billed per GB in each direction.
  • Release Elastic IPs nobody uses and count public addresses on load balancers and instances.

Observability

  • Set a retention period on every CloudWatch log group; the default is to keep logs indefinitely.
  • Send verbose debug and access logs to the Infrequent Access log class and keep logs that drive metric filters and alarms on Standard.

How to reduce Azure costs: a checklist

  • Azure Advisor, with a longer lookback. Advisor’s VM resize and shut-down recommendations use seven days of data unless you change it; set the lookback to 30 days or more under Advisor configuration. Microsoft notes that Advisor’s savings figures use retail rates and ignore your reservations and savings plans, so they overstate what you will see, and a cross-series resize can even raise costs if a reservation covers the old series.
  • Budgets and anomaly alerts in Cost Management. Cost Management has no charge for Azure subscriptions. Budgets can sit on management groups, subscriptions or resource groups and alert on forecast as well as actual spend. Anomaly alerts are created per subscription only, at most five rules each, compare each day with a model trained on the last 60 days, and run about 36 hours after the day ends, so a budget forecast alert is your faster signal.
  • Azure Hybrid Benefit. If you own Windows Server or SQL Server core licences with Software Assurance or qualifying subscriptions, apply them to VMs, Azure SQL and AKS Windows nodes to remove the licence part of the price. During a migration you may use the same licences on-premises and in Azure for up to 180 days. Red Hat and SUSE subscriptions have their own version of the benefit.
  • Dev/Test pricing. Subscriptions created under the Dev/Test offer for active Visual Studio subscribers bill Windows and SQL Server VMs at Linux rates and discount several other services. They are for development and testing only, with no uptime SLA.
  • VM size and type. Follow Advisor’s burstable B-series suggestions for VMs with low average CPU, switch on auto-shutdown for development VMs, and run batch work on Spot VMs, which Azure evicts with 30 seconds’ notice through Scheduled Events.
  • Disks and snapshots. A managed disk keeps billing after its VM is deleted. Search for unattached disks and old snapshots every month.
  • Databases. Use serverless with auto-pause for Azure SQL databases that sit idle, and elastic pools for many small databases with uneven load.
  • Log Analytics. Put verbose tables on the cheaper Basic or Auxiliary plans, set table-level retention, and consider a commitment tier once daily ingestion is steady.

Google Cloud, briefly

The same order applies on Google Cloud, with three differences worth knowing. Sustained use discounts arrive automatically, but only on older machine series (N1, N2, N2D, C2, M1, M2 and shared-core), so moving to a newer series such as N4 or E2 can lose that discount even as the list price falls; compare totals, not rates. Committed use discounts come in resource-based and spend-based flexible forms. And Spot VMs, at up to 91% off, give no preemption warning by default: you must opt in to the 120-second notice. The Recommender in the console covers idle VMs, rightsizing and commitments.

Kubernetes: where cluster costs hide

In EKS, AKS or GKE, the bill follows what pods request, not what they use, so inflated requests turn directly into extra nodes. Three checks catch most of it: compare requests with real usage per namespace, let the node autoscaler (Karpenter, Cluster Autoscaler or the provider’s managed mode) consolidate half-empty nodes, and keep control planes on supported versions, because an EKS cluster in extended support costs $0.50 an hour more, about $365 a month per cluster. Our Kubernetes cost guide has the worked numbers, and our Kubernetes consulting team does the tuning.

Commitment strategy under the 2026–27 rules

Commitments are the largest single discount and the only lever that can lose you money, so the order of operations matters.

  1. Finish the clean-up and upgrades first. A commitment priced on an oversized or extended-support estate locks the waste in.
  2. On AWS, start with Compute Savings Plans. They follow you across instance families, Regions, Fargate and Lambda. Choose EC2 Instance Savings Plans only for a family you are sure to keep in one Region for the whole term.
  3. Treat Database Savings Plans as a one-year decision. They are one-year and no-upfront only, and cover current instance generations across engines; plan engine upgrades before you buy.
  4. Use the AWS return window. A Savings Plan with an hourly commitment of $100 or less can be returned within seven days of purchase, provided the return happens in the same calendar month (UTC). Buy early in a month and check the coverage report within the week.
  5. On Azure, decide before 1 February 2027. Reservations bought from that date for services that savings plans also cover (VMs, App Service, Azure SQL Database and similar) cannot be exchanged. Reservations bought earlier keep one final exchange. Refunds remain capped at USD 50,000 of cancelled commitment per rolling 12 months, and Microsoft says it may add a 12% early termination fee in future. In practice: reservations only for workloads whose VM series and region will not change for the term, the savings plan for everything else.
  6. Share commitments across accounts and subscriptions deliberately. On AWS, Savings Plans apply across the consolidated billing family unless discount sharing is turned off; on Azure, set the reservation or savings plan scope to shared, a management group or a single subscription on purpose, and make sure showback reports allocate the discount fairly.

Licences and support: the lines nobody owns

Windows Server and SQL Server

  • Check which SQL Server instances use Enterprise features. Many do not, and Standard edition costs a fraction as much per core.
  • Non-production SQL Server can use the free Developer editions. Amazon RDS for SQL Server has supported Developer Edition since December 2025, so test databases no longer need licence-included Standard or Enterprise.
  • Bringing your own Microsoft licences to AWS or Google Cloud is restricted: Microsoft treats them as “Listed Providers”, Windows Server licences cannot use License Mobility, and SQL Server needs Software Assurance or a subscription to move. Confirm the position with your licensing reseller before changing anything.
  • AWS runs a complimentary Optimization and Licensing Assessment (OLA) that models licence-included against bring-your-own options for Windows and SQL Server workloads. It is worth requesting before a large Windows estate moves or renews.

AWS Support plans change on 1 January 2027

AWS will discontinue Developer Support and Business Support on 1 January 2027, outside the GovCloud (US) Region. The paid plans become Business Support+, Enterprise Support and Unified Operations. Enterprise On-Ramp customers are being moved to Enterprise Support at renewal during 2026, and the Enterprise minimum has dropped to $5,000 a month from $15,000.

Business Support+ is priced at the greater of $29 a month per account or 9% of monthly AWS charges up to $10,000, 7% from $10,000 to $80,000, 5% from $80,000 to $250,000 and 3% above that (AWS Support pricing). For an account with $12,000 of monthly charges that is $900 plus $140, or $1,040 a month. Before January, list your accounts, decide which actually need round-the-clock cover from AWS, and leave sandboxes and archives on Basic. AWS’s end-of-support notice has the transition details.

Worked example: one AWS estate, six findings

Illustrative, not client work. The prices are AWS’s published examples quoted above; replace the quantities with yours.

FindingArithmeticMonthly effect
40 TB of gp2 volumes40,000 GB × ($0.10 − $0.08), after checking no volume needs more than the included IOPS−$800
30 public IPv4 addresses nobody uses30 × $0.005 × 730 hours−$109.50
20 TB a month of S3 traffic through a NAT gateway20,000 GB × $0.045, moved to a free gateway endpoint−$900
Two 8-vCPU RDS for MySQL 8.0 instances in Extended Support16 vCPU × $0.100 × 730 hours, removed by upgrading to 8.4−$1,168
$3,000 a month of non-production EC2 running all week$3,000 × (1 − 60 ÷ 168) on a 60-hour weekly schedule−$1,929
2 TB a month of debug logs on Standard ingestion2,000 GB × $0.50, halved on Infrequent Access−$500
Total before any commitmentAbout −$5,400 a month

None of these changes needs a commitment or a redesign. The Savings Plan comes afterwards, sized on the smaller bill. We do not put a percentage on it here, because the rate depends on instance family, Region, term and payment option.

Quick wins for the first week

  1. Day 1: export three months of cost by service and usage type, list the top fifteen lines and the untagged share, and name an owner for each line.
  2. Day 2: switch on anomaly detection and budgets with forecast alerts for every account or subscription, routed to the owning team.
  3. Day 3: find unattached disks, unused public IPs, idle load balancers and snapshots past retention; ask owners, then delete or tag with a reason.
  4. Day 4: schedule development and staging, convert gp2 to gp3 where IOPS allow, add S3 gateway endpoints, and set log retention everywhere.
  5. Day 5: list every database engine and Kubernetes version against its end-of-standard-support date, and put the upgrades on the roadmap with their monthly cost of delay.

Governance that stops the bill creeping back

  • Tags enforced at creation, not requested in a wiki: AWS tag policies or Azure Policy, plus a pipeline check on infrastructure code. The FinOps playbook lists the four tags we start with.
  • Guardrails in sandboxes: service control policies or Azure Policy that deny unapproved Regions and expensive instance families such as GPUs unless a ticket says otherwise.
  • A cost estimate in pull requests for infrastructure changes, so a new NAT gateway or a larger database is a reviewed decision.
  • Showback each month: every team sees its own spend and one unit cost, such as cost per customer or per thousand requests.
  • A calendar of expiries: commitments, engine support dates, Kubernetes versions and the support plan, reviewed each quarter.

When not to optimise

  • When the saving is smaller than the engineering time. A week of a senior engineer to save $40 a month is a loss.
  • During a launch, a migration or an incident. Change one thing at a time, and not then.
  • When the “idle” capacity is your recovery plan. Standby databases and warm disaster-recovery capacity look idle by design; tag them and leave them.
  • When a commitment would outlive the workload. If the architecture or the cloud may change within the term, stay on demand or buy small.
  • When reliability already suffers. Fix the outages first; a cheaper system that falls over is not cheaper.
  • Regulated workloads without a change process. Health and finance systems need the same review, testing and audit trail for a cost change as for any other change.
Services

Cloud cost optimization services

Four ways to engage. Effort figures are typical ranges, confirmed after a short scoping call and priced as a fixed fee per phase. Every change is made as code, through your review process, in your own accounts.

Fixed-scope cost review

The problem

The bill has grown, the console shows dozens of recommendations, and nobody knows which ones are safe or worth the effort.

How we do it

Read-only access to one cloud, three months of cost data, the native recommendation tools checked against real metrics, and an interview with each workload owner. You get a ranked list of findings, each with a monthly saving, a risk rating and an effort estimate. Typically 5–8 engineer-days for one cloud with up to about 20 accounts or subscriptions.

What you get

  • Cost baseline by service, team and environment
  • Ranked findings with saving, risk and effort
  • Extended-support and expiry calendar
  • Readout with engineering and finance
How it runs

How a review and sprint run

Typical for one cloud. Two clouds run in parallel with a second engineer.

Before

Scope and access

Agree which accounts or subscriptions are in scope. You create a named read-only role with billing view (AWS) or Reader and Cost Management Reader (Azure); no long-lived keys.

Days 1–3

Baseline and data

Three months of cost by service, usage type and tag; recommendations pulled from the native tools and checked against at least 30 days of metrics.

Days 4–8

Findings and readout

Ranked findings with saving, risk and effort, the expiry calendar and a commitment view, walked through with engineering and finance.

Weeks 2–5

Savings sprint

Agreed changes made as code through your reviews, with each saving measured against the baseline at the end.

Senior engineers firstYour first call is with someone who will do the work, not a sales script.
You own everythingCode, configuration, reports and documentation stay in your accounts and are handed over.
We run our own products15 live platforms we build and operate, so we feel production pain too.
Honest about fitIf a cheaper route or another provider suits you better, we will say so.
FAQ

Questions about reducing cloud costs

Start with changes that need no redesign: delete unattached EBS volumes and unused public IPv4 addresses, schedule development and staging outside working hours, convert gp2 volumes to gp3, add S3 gateway endpoints so traffic stops paying NAT processing, and set retention on CloudWatch log groups. Then check for RDS or EKS versions in extended support. Buy Savings Plans only after these, sized to the smaller bill.

Set Azure Advisor’s lookback to 30 days or more and act on its shut-down and resize recommendations, remembering its savings are at retail rates. Apply Azure Hybrid Benefit where you own eligible Windows Server or SQL Server licences, use Dev/Test subscriptions for non-production, delete unattached managed disks, and put batch work on Spot VMs. Then buy a savings plan or reservations for the steady baseline.

Ours start with a fixed-scope review of one cloud: a cost baseline, recommendations checked against real metrics, and a ranked list of findings with saving, risk and effort. A savings sprint then makes the changes as code through your reviews, a commitment plan sizes Savings Plans or reservations, and an optional monthly retainer keeps the savings from drifting back.

On AWS, Compute Savings Plans are the safer default because they follow you across instance families, Regions, Fargate and Lambda. EC2 Instance Savings Plans give a deeper discount but tie you to one family in one Region. Database Savings Plans are one-year, no-upfront and cover current instance generations. Reserved instances remain for services Savings Plans do not cover.

AWS will discontinue Developer Support and Business Support on 1 January 2027, except in GovCloud (US). The paid plans become Business Support+, priced from $29 a month per account or a sliding share of usage, Enterprise Support from $5,000 a month, and Unified Operations from $50,000. Review which accounts need paid support before January.

From 1 February 2027, newly bought reservations for services that Azure savings plans also cover, such as virtual machines, App Service and Azure SQL Database, cannot be exchanged. Reservations bought before then keep one final exchange. Refunds stay capped at USD 50,000 of cancelled commitment per rolling 12 months. Use reservations only where the VM series and region will not change.

We do not quote a percentage before seeing the bill, because it depends on how much is idle, oversized, on demand or in extended support. The vendors publish ceilings, such as up to 72% for some AWS Savings Plans and Azure reservations and up to 90% for Spot capacity, but those apply only to specific usage. The review puts a monthly figure on each finding instead.

Read-only access through a named role, yes: billing and resource read access on AWS, Reader and Cost Management Reader on Azure. We never ask for root credentials or long-lived keys. Changes in a savings sprint go through your own pull requests and change process, and access is removed when the engagement ends.

Last updated