Quick Summary
- Cloud bills rarely overspend on purpose. Waste hides in idle compute, forgotten environments, orphaned storage, and unused licenses that nobody’s job it is to catch.
- Flexera’s 2026 research puts wasted cloud spend at 29% industry-wide, the first increase in five years, driven largely by ungoverned AI workloads.
- You can audit the biggest leaks yourself this week: CPU under 10%, non-production with no shutdown schedule, storage marked “available,” licenses unused for 90+ days.
- AWS and Azure both ship native cost optimization tooling, but a recommendation nobody acts on saves nothing.
- Rightsize before committing to Reserved Instances or Savings Plans, since committing to the wrong size just locks in the waste.
- Cloud cost optimization services work because the looking never stops. Spend drifts back the moment nobody’s watching.
Nobody decides to overspend on cloud. The bill grows because every individual increment is defensible and no individual increment is anybody’s job to reverse.
A developer spins up a test environment for a migration. The migration finishes. The environment stays, because switching it off is nobody’s task and there is a small risk it is still needed. Multiply that by three years and a few dozen well-intentioned decisions, and you arrive at a monthly invoice where a meaningful share buys nothing at all. That share is not small. In Flexera’s 2026 State of the Cloud Report, organizations estimated that 29% of their cloud spend delivers no business value at all, the first increase in five years, and it’s a number worth starting from.
The useful question is not “how do we cut cloud costs.” It is “which specific things are we paying for that nothing depends on?” That question has findable answers.
Why Cloud Bills Only Move in One Direction
Three structural reasons, none of which are about technical skill.
Scaling up is automatic; scaling down is a decision. Autoscaling adds capacity when load demands it. Removing it usually requires someone to notice, decide it is safe, and act. One of those is instrumented and the other is a to-do.
Nobody owns the invoice line. Finance sees a total. Engineering sees a platform. The specific question (what is this particular resource for, and does anything still need it?) sits between the two, and questions between two teams do not get asked.
Deleting things carries asymmetric risk. If you remove a resource nothing needed, nobody notices. If you remove one something needed, it is your fault. Under that asymmetry the rational individual choice is to leave everything running, and that choice compounds.
This is why cloud cost work is an ownership problem before it is a technical one. The tooling to find waste already exists in every major platform. What is usually missing is anyone whose job it is to look, and the authority to act on what they find.
The Eight Places Cloud Spend Leaks
1 Idle and oversized compute
Instances provisioned for a peak that never arrived, or sized once and never revisited. Look for: instances averaging under ~10% CPU across a month. Both major platforms surface this natively, see AWS’s own cost optimization guidance and Microsoft’s Azure cost optimization guidance in the Well-Architected Framework. Why it matters: rightsizing is the least disruptive change available. Same workload, smaller allocation, no architectural work.
2 Environments nobody switched off
Dev, test, staging and migration environments running 24/7 for workloads used during office hours, if at all. Look for: non-production resources with no scheduled shutdown. Why it matters: a non-production environment that runs nights and weekends is paying for roughly three times the hours it is used.
3 Orphaned storage and forgotten snapshots
Volumes detached from any instance, snapshots of systems long decommissioned, backups of environments that no longer exist. Look for: unattached volumes, and snapshots older than your actual retention requirement. Why it matters: storage is individually cheap, which is exactly why it accumulates unnoticed for years.
4 Unused and over-tiered software licenses
Seats assigned to people who left, and premium tiers assigned to people who need the basic one. Look for: accounts with no sign-in activity in 90 days, and a comparison of license tier against actual feature use. Why it matters: this is usually the fastest single item to resolve, because it needs no technical change at all, and it is a security control as much as a cost one, since a licensed account is usually an active account, the kind of overlap our Security as a Service work also flags.
5 Data transfer nobody budgeted for
Egress between regions, between availability zones, or out to the internet, rarely modeled at design time and easy to make expensive with a chatty architecture. Look for: transfer as a share of total spend, and any single service disproportionately responsible.
6 On-demand pricing for entirely predictable workloads
Baseline workloads that run continuously, billed at on-demand rates instead of committed pricing such as AWS Reserved Instances or Savings Plans, or Azure Reserved VM Instances. Look for: consistently-running production instances with no commitment applied. Why it matters: this is a purely commercial change, same infrastructure, different purchasing model, but it requires confidence the workload is genuinely stable, which is an argument for doing it after rightsizing rather than before.
7 Logs and backups with no retention policy
Log retention set to default-forever, and backups kept well beyond any regulatory or operational need. Look for: log groups with no expiry set. Why it matters: retention is a compliance question as well as a cost one, and “keep everything indefinitely” is not usually what a regulator asked for.
8 Duplicated tooling across teams
Two teams solving the same problem with different paid services, or paying for monitoring the platform already provides. Look for: overlapping line items on the invoice.
A ninth leak, new in 2026
Wasted cloud spend had been falling for five straight years. In Flexera’s 2026 report, it rose again, and the report ties the reversal directly to generative AI workloads moving into production faster than anyone built cost governance around them. If your organization has started running AI workloads in the cloud this past year, treat that spend as a ninth item on this list: it needs the same ownership and review cadence as everything else, not a pass because it’s new. This is one reason cost governance for AI workloads is now part of our AI as a Service engagements rather than a separate conversation.
How to Audit Your Own Environment This Week
You do not need a consultant to find the first tranche. In roughly a day, in your existing console:
- Filter instances by average CPU under 10% over 30 days. Everything on that list is a rightsizing or shutdown candidate.
- List every non-production environment and check for a shutdown schedule. Anything without one is running about three times longer than it is used.
- Filter storage volumes for status “available,” meaning attached to nothing. Check the snapshot list against your actual retention requirement.
- Export your user license list and sort by last sign-in. Anything over 90 days is a candidate, and worth checking against your leavers list.
- Sort the invoice by line item, largest first, and ask “what depends on this?” for the top ten. If nobody can answer within a day, you have found something.
- Look at data transfer as a share of total. If it is material and nobody has an explanation, it is worth investigating.
Get a Free IT Assessment → · We respond within 24 hours.
What Cloud Cost Optimization Services Actually Do
The audit above is a snapshot. Spend drifts back, because the same three structural reasons remain. What a managed service adds is that the looking never stops and the acting is automated where it safely can be. This is cloud cost management as an ongoing discipline, not a one-time cut.
Concretely, within our managed hosting and cloud services:
Capacity that scales in both directions Sustained high load spins up a new virtual machine; when load drops, it shuts down again. The down direction is the one that changes an invoice, and automating it removes the human decision that otherwise never gets made.
A live inventory rather than a periodic export. You cannot question a resource you do not know exists. The estate inventory covers cloud and on-premises together, with end-of-life and unsupported versions flagged as standing items.
License position reviewed as routine, rather than as a scramble before a true-up.
A monthly report covering what was found and whether flagged items were actually closed. This is the accountability layer that the “nobody owns the invoice line” problem is missing.
One thing worth being direct about: the major platforms sell their own monitoring and scaling services separately, and you then need someone who can configure and act on them. Our monitoring and automation come bundled with the managed service. You are not buying a second product to watch the first.
Definition: rightsizing.
Matching the resources allocated to a workload to what it actually consumes, rather than what was provisioned when it was first deployed. Usually the lowest-risk optimization available, because the workload itself does not change.
Cloud vs. On-Premises: Is It Actually Cheaper to Move Back?
Sometimes, and it is asked more often than it is acted on, usually after a bill shock.
| Workload shape | Tends to favor | Why |
|---|---|---|
| Steady, predictable, always-on, no seasonality | Owned hardware, on a multi-year view | You’re paying for capacity you use every hour anyway |
| Variable, seasonal, or growing | Cloud | The entire value of cloud is not paying for the peak year-round |
The honest answer is that it depends on workload shape. Steady, predictable, always-on workloads with no seasonality can favor owned hardware on a multi-year view. Variable, seasonal or growing workloads generally do not, because the entire value of cloud is not paying for the peak year-round.
What makes the comparison unreliable is that the on-premises side is routinely undercounted. A fair comparison includes hardware refresh cycles, data-center or hosting costs, power, the staff time to run it, backup and disaster-recovery capability, and the capital tied up. It also has to price the thing cloud gives you that hardware does not: the ability to be wrong about capacity and correct it in an afternoon.
In practice, for most of the organizations we work with (legal and financial firms of 15 to 50 users, fully cloud-based, with no in-house cloud expertise), the answer is not repatriation. It is that their existing cloud environment has never been actively managed by anyone.
What Good Looks Like After Six Months
Not a one-off saving. A different operating pattern:
- Non-production environments shut down outside working hours by schedule, not by memory.
- Rightsizing reviewed on a cadence, so drift is caught in weeks rather than years. Quarterly is a reasonable starting point for most organizations, more often for automated elements like scaling schedules, which should run continuously.
- Storage and snapshot retention governed by an actual policy tied to your regulatory requirements.
- License position reviewed routinely against leavers and actual usage.
- A named owner for the invoice, internal or your provider, whose job includes asking what each line is for.
- Spend that tracks what the business is doing, rather than only ever rising.
The measure of success is not a single dramatic reduction. It is that the bill becomes explainable: every material line has an owner and a reason. That is also, not coincidentally, what makes the next budget conversation straightforward. It fits alongside our full range of managed services rather than being a one-time engagement.
How much can we expect to save?
It depends entirely on how long the environment has run unmanaged and how much of it is non-production. We are not going to quote you a percentage. Anyone who does before seeing your environment is guessing, and the honest answer comes from the audit.
Isn't this what the cloud provider's own cost tools are for?
Partly. Both major platforms have capable native cost tooling, and the recommendations they produce are genuinely useful. Two gaps remain: those services are usually billed separately, and, more importantly, a recommendation nobody acts on saves nothing. The tooling identifies; someone still has to decide and execute.
Will optimizing cost hurt performance?
It should not, if sequenced properly. Rightsizing matches allocation to actual consumption, so a correctly rightsized workload performs identically. The risk comes from cutting without measuring first. This is why the audit precedes the changes, and why commitment-based pricing should come after rightsizing rather than before.
We only have about 30 users. Is this worth doing?
Often yes, proportionally more than expected, because smaller organizations are least likely to have anyone whose job includes looking. The license and idle-environment items in particular tend to be worth finding at any size, and both are among the fastest to resolve.
Does this include Microsoft 365 licensing?
Yes. License position is one of the leak points above, and unused or over-tiered seats are typically the quickest item to resolve because they require no technical change. It is also a security improvement, since a licensed account is usually still an active one.
