Work the levers in payback order and your performance never suffers: delete what nobody owns, rightsize what is oversized, schedule what only runs in office hours, move cold data to a cold tier, and only then buy commitments. Teams do this backwards. They negotiate a three-year discount on an estate that is 30% idle, and then congratulate themselves on a cheaper mistake.
Find the waste before you negotiate anything
Cloud waste is not exotic. It is a staging environment from a project that shipped in March, three unattached disks, a load balancer with no targets, nine snapshots of a server that no longer exists, and a database sized for a launch that did not happen. Flexera’s 2026 survey work puts the estimate at 29% of IaaS and PaaS spend, with 71% of organisations now running a cloud centre of excellence specifically to keep this under control.
Start with an inventory you can defend: every resource, its monthly cost, its tags and a named owner. Anything without an owner gets a two-week notice in a shared channel and is then stopped, not deleted. Stopped resources are reversible, and the complaints arrive within a day if you were wrong.
- Unattached block storage volumes and orphaned snapshots.
- Load balancers and public IP addresses with nothing behind them.
- Non-production environments running 168 hours a week.
- Oversized managed databases on their default instance class.
- Log retention set to forever on a high-volume service.
Lever one: switch things off
Non-production is the easiest money on the table. A development and test environment used between 8am and 7pm on weekdays needs about 55 hours a week, not 168. Shut it down the rest of the time and you are paying for roughly a third of the hours, with zero effect on production performance because there is none to affect.
Use the platform’s own scheduler rather than a cron job on a server that itself gets shut down. Tag anything that must stay up with an exemption tag, and review that list monthly, because exemptions breed.
Lever two: rightsize, then consider spot
Rightsizing is dull and it works. Pull 30 days of CPU, memory and disk throughput percentiles, and compare the 95th percentile with the instance you are paying for. If your peak is 18% of a four-vCPU machine, you are buying three vCPUs of headroom you will never use. Step down one size, watch for a week, repeat.
Capacity model and operating system licence often matter more than size. The same Azure D4s v5 in East US is listed at $0.192 an hour for Linux, $0.376 an hour for Windows, and $0.040531 an hour on Spot capacity, read from the Azure price list in October 2026. Spot is roughly 79% cheaper than on-demand because the platform can evict it. That is perfect for batch jobs, video encoding, CI runners and anything that can retry. It is wrong for your payment API.
Lever three: put storage in the right tier
Storage is where the savings are safest, because moving a file to a cooler tier does not slow down your application unless the application reads that file constantly. The rates below are AWS list prices for the N. Virginia region, read in October 2026, with the monthly cost of 1 TB worked out alongside.
| Tier | Price per GB-month | 1 TB per month | Retrieval cost | Use it for |
|---|---|---|---|---|
| S3 Standard | $0.023 | $23.55 | None | Active files, website assets, working data |
| S3 Standard-Infrequent Access | $0.0125 | $12.80 | Per-GB retrieval fee | Monthly reports, older project files |
| S3 One Zone-Infrequent Access | $0.01 | $10.24 | Per-GB retrieval fee | Re-creatable data you can afford to lose |
| S3 Glacier Instant Retrieval | $0.004 | $4.10 | Per-GB retrieval fee | Archives you read a few times a year, instantly |
| S3 Glacier Flexible Retrieval | $0.0036 | $3.69 | Minutes to 12 hours | Compliance copies, finished projects |
| Deep archive tier | $0.00099 | $1.01 | $0.02 per GB standard, $0.0025 bulk | Seven-year retention nobody expects to read |
Two cautions. Cold tiers have minimum storage durations, so moving a file down and back up within a month can cost more than leaving it alone. And retrieval fees are real: pulling a full terabyte out of a deep archive at $0.02 per GB is about $20, which is fine for a genuine recovery and painful if it becomes routine. Set a lifecycle policy, then leave it alone.
Lever four: stop paying the egress tax
Every gigabyte leaving the platform is billed. Azure’s bandwidth pricing page lists the first 100 GB a month free from North America and Europe, then $0.087 per GB for the next 10 TB, dropping to $0.05 per GB at high volume. AWS uses $0.09 per GB in its own S3 pricing example for Ireland to the internet. Serve 5 TB of media a month and you are into three figures before you count storage.
- Put a CDN in front of static assets. Cache hits are served from the edge, and CDN egress is usually cheaper than origin egress.
- Keep chatty services in the same region and zone where you can. Cross-region replication is a real cost, not a free safety blanket.
- Compress and resize before you ship bytes. Modern image formats routinely halve payloads, which halves this line of the bill.
- Consider zero-egress object storage for public downloads. Cloudflare R2 lists $0.015 per GB-month with no egress charge, which changes the maths for media libraries and software distribution.
Egress is not a billing accident. It is a design decision you made three sprints ago without noticing.
Lever five: commit, but only once you know your shape
Commitments are the biggest single discount and the last lever you should pull. AWS documents savings of up to 72% off on-demand for Savings Plans and Reserved Instances, with EC2 Instance Savings Plans tied to a family and region, and Compute Savings Plans applying more broadly across EC2, Fargate and Lambda. Azure reservations and Google’s committed use discounts work on the same principle: a one or three-year promise in exchange for a lower rate.
| Purchase | Typical commitment | Flexibility | When it fits |
|---|---|---|---|
| On demand | None | Total | New workloads, anything you might delete this quarter |
| One-year commitment | 12 months of baseline spend | Moderate | Steady workloads you have run for at least two months |
| Three-year commitment | 36 months | Lowest, best rate | Core production you are confident about, up to 72% off on AWS |
| Spot or pre-emptible | None | Can be evicted | Batch, CI, encoding, anything that retries safely |
A practical rule: commit to no more than 70% of your observed steady-state baseline, and leave the peak on demand. You keep the discount on the predictable part without betting on a forecast.
The monthly routine that keeps the bill flat
Cost control is a habit, not a project. Thirty minutes a month, same day each month, with the same four questions.
- What changed? Compare this month’s cost by service against last month’s and explain any movement over 10%.
- What is idle? Anything under 10% utilisation for 30 days gets resized or stopped.
- What is untagged? Untagged resources get an owner or a shutdown date.
- What is about to expire? Commitments, support plans and certificates, so renewals are decisions rather than surprises.
If nobody has the time for that, we do it as a fixed monthly service for international clients, remotely, as part of our cloud solutions work. It pairs well with a migration done properly, because month two after a move is when the savings are largest. If you are weighing up spreading workloads across providers, read single cloud versus multi-cloud first, and make sure your backups are not the thing you economise on.
Frequently asked questions
How much can we realistically save?
On an estate that has never been reviewed, 20% to 35% without touching performance is normal, and it lines up with Flexera’s 2026 estimate of 29% wasted IaaS and PaaS spend. Most of it comes from stopping idle resources, rightsizing and storage tiering. Commitments then take another slice off what remains.
Will rightsizing slow our application down?
Not if you size to the 95th percentile of real usage and step down gradually. The risk is going two sizes at once on a workload with spiky demand. Change one thing, watch it for a week, then continue. Keep a note of the original size so rolling back is a single click.
Is spot capacity safe to use in production?
For the right workloads, yes. Spot can be reclaimed at short notice, so it suits queue workers, batch jobs, CI runners and encoding, where a retry costs nothing. At $0.040531 versus $0.192 an hour for an Azure D4s v5 in East US, the saving is large enough to design around. Keep user-facing APIs on normal capacity.
Should we buy a three-year commitment?
Only for the part of your estate you are sure about, and only after rightsizing. AWS documents up to 72% off on-demand for Savings Plans and Reserved Instances, which is excellent value on a stable baseline and an expensive way to lock in an oversized instance. Commit to about 70% of your baseline and keep the rest flexible.
Can you review our bill without access to our systems?
We can start from an exported cost and usage report, which is enough to spot the big items. Deeper work needs read-only access to your accounts. Everything is done remotely for international clients, with findings written up so your own team can act on them.
Want someone to find the 30% before your next invoice? Send us a cost export and we will come back with a prioritised list. Get in touch with Eudora Technology to talk about your project.
