Skip to main content
Cloud DevOps Solutions
Get a quote Talk to an engineer
Cloud Cost 2026-11-04 9 min read

How to Measure EKS Cost Savings Honestly: Baseline, Normalise, Verify

Saddique Khan

DevOps / SRE Engineer & CEO, Cloud DevOps Solutions

Most “we cut our Kubernetes bill by X%” claims compare this month’s invoice with last quarter’s invoice. That comparison is almost always wrong, because the two periods differ in traffic, headcount, pricing and the number of days. This guide sets out a method for measuring EKS savings that survives a sceptical finance reviewer. We publish no percentage here on purpose: your result depends on your baseline, and any figure quoted without one is marketing.

If you want the list of techniques first, read 12 ways to reduce an EKS bill. This article is about proving that they worked.

Why raw bill comparisons mislead

Four things move your bill that have nothing to do with optimisation:

  • Growth: more users mean more pods and more data transfer. A bill that is flat while traffic doubled is a large saving. A bill that fell while traffic fell is not.
  • Calendar effects: a 31-day month costs more than a 28-day month, and seasonal products have peaks.
  • Price and commitment changes: new instance generations, region price changes, Savings Plan purchases and expiring credits all move the invoice.
  • One-off items: support plan changes, marketplace fees, reserved instance upfront payments and refunds distort a single month.

If you do not control for these, you will either overclaim, which damages trust when someone checks, or underclaim and miss the case for further work.

Step 1: Fix the scope before you measure

Decide, in writing, what is inside the measurement:

  1. Which AWS accounts and EKS clusters are in scope.
  2. Which cost lines count: EC2 or Fargate compute, EBS and EFS storage, load balancers, NAT gateways, cross-AZ and internet data transfer, CloudWatch logs and metrics, ECR, and the EKS control plane fee.
  3. Whether shared platform costs (monitoring, CI runners) are allocated to workloads or excluded.

A common failure is optimising compute while log ingestion quietly grows. Keep every line that the cluster causes inside the scope, even if it appears under a different AWS service.

Step 2: Capture a baseline of at least four to six weeks

A single month hides weekly cycles and one-offs. Take the longest stable window you can, and record:

  • Cost by line item, from Cost Explorer or the Cost and Usage Report, grouped by cost allocation tags or by linked account.
  • Workload demand, such as requests per second, active users, jobs processed or whatever your product’s real unit of work is.
  • Resource efficiency, measured from the cluster: CPU and memory requests versus actual usage per namespace, node utilisation, and the share of nodes that are Spot or covered by commitments.
  • Prices in force, including your effective Savings Plan or Reserved Instance coverage during the window.

Store this in a document or spreadsheet that someone else can open in six months. A baseline nobody can find later is not a baseline.

Step 3: Choose a unit cost, not just a total

The most defensible metric is cost per unit of work:

  • cost per 1,000 requests,
  • cost per active customer,
  • cost per build or per job,
  • cost per environment for non-production.

Unit cost falls when you become more efficient, even when the total rises because the business grew. It is also the figure a finance team can compare month to month without arguing about volume.

If you cannot define a unit of work yet, start with cost per namespace or per team, and refine later. An imperfect unit is better than a raw total.

Step 4: Normalise before you compare

For each comparison period, adjust for the factors above:

  1. Days in the period: compare daily averages, not monthly totals.
  2. Demand: divide by your unit of work, or scale the baseline by the change in demand.
  3. Prices: re-price the new usage at baseline prices (or the baseline usage at current prices) so a commitment purchase is not counted as an engineering saving.
  4. One-offs: remove refunds, credits, upfront commitment payments and one-time migrations. List them separately so a reviewer can see you did not hide them.

What remains is the saving attributable to your changes. It is usually smaller than the headline number and much easier to defend.

Step 5: Attribute savings to changes

Where possible, tie each saving to the change that caused it. Make one change at a time, or at least record the date of each change. A table like this works:

ChangeDateMechanismEvidence
Lower CPU and memory requests on service Awhen deployedFewer nodes neededNode count and utilisation graphs
Spot for stateless workerswhen deployedLower price per vCPU hourMix of Spot and On-Demand hours
Scale non-production down overnightwhen deployedFewer idle hoursNode hours by environment
Reduce log retentionwhen deployedLess ingestion and storageCloudWatch usage by log group

This also protects you from regressions. If a saving disappears later, you know which change to look at.

Step 6: Separate recurring savings from one-offs

Report two numbers:

  • Recurring: changes that reduce cost every month, such as right-sizing, scheduling and storage tiering.
  • One-off: clean-up of orphaned volumes, snapshots and load balancers. These help once and then stop.

Finance teams budget on recurring savings. Mixing the two inflates a forecast that then fails to materialise.

Be careful with commitment purchases. A Savings Plan lowers your effective rate, but it also locks in spend. Count it separately, and check that you are not committing to capacity that your right-sizing is about to remove.

Step 7: Watch for rebound and trade-offs

Savings that harm reliability or delivery speed are not savings. After each change, check:

  • error rates, latency and saturation on affected services,
  • Spot interruption rates and how workloads handled them,
  • deployment frequency and on-call load,
  • whether teams re-inflated resource requests after being throttled.

Cost work that increases incidents usually gets reversed, and the reversal removes the saving.

Step 8: Report in a form others can check

A trustworthy report includes:

  1. The baseline window and what was in scope.
  2. The unit-cost metric and its definition.
  3. Normalised before and after figures, with the method.
  4. A list of changes with dates and expected mechanisms.
  5. Recurring versus one-off savings.
  6. Known caveats, such as an unusual traffic spike in the comparison period.
  7. The data sources, so someone else can reproduce the numbers.

If a provider or vendor gives you a savings figure with none of the above, ask for the baseline before you believe it.

A short checklist

  • Scope written down, including data transfer, logs and storage
  • Baseline of four to six weeks or more, stored where others can find it
  • A unit of work defined, or an interim proxy
  • Comparison periods normalised for days, demand, prices and one-offs
  • Each change dated and linked to a mechanism
  • Recurring and one-off savings reported separately
  • Reliability and delivery metrics checked after each change
  • Sources and method documented

Where this approach does not apply

If your workload is very small, the effort of a formal baseline may exceed the saving, and a simple before and after comparison with notes is enough. If your usage is changing fast, for example during a migration, pick a unit-cost measure and compare trends rather than periods. And if you rely heavily on credits or negotiated discounts, involve whoever owns those agreements before you report results.

Next steps

You can estimate your own starting point with the EKS cost calculator. It gives a rough model from the inputs you provide, not a measurement of your real invoice. If you would like a second pair of eyes on your baseline or want us to run the review with you, get in touch and we will tell you honestly whether an engagement is worthwhile.

#eks #finops #cost-optimisation #kubernetes #baseline

Related service

Need hands-on help? Our certified engineers deliver this as a managed service.

AWS EKS Consulting
Found this useful? Share it with your team or link to it from your documentation.

Need help implementing this?

Our certified engineers can handle the setup, migration, or ongoing operations for you.