For five years the cloud waste figure went down. In 2026 it went back up. That is the single most useful thing to know about cloud cost optimisation right now, because it means the problem everyone assumed was being slowly solved has started growing again.
It has not gone up because teams stopped trying. Waste reduction is still the top current priority for the people doing this work. It has gone up because the thing generating the waste moved, and most cost programmes are still pointed at where it used to be.
Transputec has managed cloud estates for UK organisations for 40 years, and 2026 is the first year the cost conversation has had to start somewhere other than the infrastructure. What follows is where the waste now sits, which of the familiar levers still reach it, and what to measure before cutting anything.
The classic levers still work. They just reach a shrinking share of the bill, because the fastest growing line on most cloud invoices is now consumption you trigger rather than infrastructure you provision.
What Actually Changed in 2026
Two pieces of research published this year tell the same story from different ends.
Flexera’s 2026 State of the Cloud Report, drawn from 753 cloud decision makers, puts estimated wasted cloud spend at 29%. The number itself is unremarkable. Its direction is not. Flexera’s own wording is that cloud-based AI workloads are “causing an increase in wasted cloud spend (29%) for the first time in five years”. Five consecutive years of improvement, reversed.
From the practitioner side, the State of FinOps 2026 surveyed 1,192 people responsible for more than 83 billion dollars of annual cloud spend. The finding that matters here: 98% of them now manage AI spend, up from 63% a year earlier. That is not a trend, it is a step change in what the job covers.
Put the two together and the picture is clear enough. The people doing cloud cost optimisation have not become worse at it. Their remit grew by a third in twelve months, and the new territory does not respond to the old tools.
Why the Old Playbook Stopped Reaching It
Everything in the standard cost playbook shares one assumption: that waste is something sitting idle, which you can find and switch off.
Right-sizing finds a machine that is too big. Reserved instances and savings plans reward you for committing to capacity you were going to use anyway. Storage lifecycle policies move data nobody has opened in ninety days somewhere cheaper. Scheduling turns off development environments at the weekend. Tagging tells you whose idle resource it is. Every one of those is an inventory problem with an inventory answer, and every one of them still works.
Consumption spend has no idle state to find. A model call that nobody needed costs exactly the same as one that produced something valuable, and it leaves no running resource behind to be discovered next month. There is no instance to resize, no commitment to true up, no snapshot to delete. By the time it appears on a bill it has already happened and the money has gone.
This is the structural reason cloud cost optimisation is harder than it was in 2024. The old waste was a thing you owned. The new waste is an event that occurred.
Can You See What Your AI Spend Is Attached To?
Most organisations can account for their compute and not for their consumption. Book an initial assessment with a Transputec cloud architect and we will map where both sit before anyone proposes a cut.
Book a consultationWhere Cloud Cost Optimisation Still Works
None of what follows is new, and that is the point. It applies to the 71% of spend that is still infrastructure, and it is where Transputec starts every cost engagement, because it is the part that pays back fastest and argues least.
| Lever | What it catches | Typical first finding |
|---|---|---|
| Right-sizing | Machines provisioned for a peak that never arrived | Sustained utilisation in single figures |
| Scheduling | Non-production environments running outside working hours | Development estates billed 168 hours a week for 40 hours of use |
| Commitment cover | Steady-state workloads paying on-demand rates | Predictable baseline with no reservation against it |
| Storage lifecycle | Data nobody has read in months on premium tiers | Snapshots and orphaned volumes with no owner |
| Tagging and allocation | Spend nobody can attribute to a team or product | A large unattributed bucket that no one will claim |
The last row is the one that makes the others possible. You cannot ask a team to justify a cost until you can tell them it is theirs. For a worked example of this on a real estate, the IQPC cost and efficiency work is ours, and it followed exactly this order.
You Cannot Right-Size a Prompt
So what does control look like when the thing you are controlling is an event rather than an asset?
It moves from sizing to limits. Where infrastructure cost is managed by choosing the right capacity, consumption cost is managed by setting a ceiling before anyone starts and watching the rate of burn against it. That is a different muscle and usually a different team.
Three practical consequences for cloud cost optimisation in an estate that now includes AI.
Set the limit before you enable the feature, not after the first bill. Defaults in this area are set by vendors who benefit from consumption, and they are rarely set low. Whatever platform you are turning on, find the spending cap in the admin centre and decide the number deliberately.
Decide who is allowed to switch things on. Consumption features are frequently enabled per user or per team, which means your cost exposure can change without anyone raising a change request. That is a governance question before it is a finance one.
Measure cost per outcome, not cost per month. A rising AI bill attached to rising output is not waste, and cutting it would be the expensive mistake. A rising bill attached to nothing in particular is the thing to find. Our note on measuring Copilot ROI works through that calculation for one specific case.
The Scope Problem Nobody Budgeted For
AI is the headline, but the State of FinOps numbers show the remit widened in every direction at once. Among those 1,192 practitioners, the share now managing each area, with last year’s figure alongside it, as set out in the Linux Foundation’s announcement of the survey:
| Area | 2026 | 2025 |
|---|---|---|
| AI spend | 98% | 63% |
| SaaS | 90% | 65% |
| Licensing | 64% | 49% |
| Private cloud | 57% | 39% |
| Data centre | 48% | 36% |
If your cost programme covers public cloud infrastructure and nothing else, it now addresses a minority of what the discipline is expected to cover. The SaaS line is the one most often missed, because SaaS renewals sit with procurement rather than with whoever watches the cloud bill, and nobody owns the overlap.
This is also why hybrid estates complicate the picture. Private cloud and data centre costs behave nothing like metered public cloud, and a single blended number tells you very little. Where Transputec runs both sides of a hybrid estate, we report them separately for exactly that reason.
What to Measure Before You Cut Anything
The most expensive version of cloud cost optimisation is the one that starts with a target percentage and works backwards. It finds savings, annoys engineers, and quietly removes headroom that turns into an incident two quarters later.
Four things are worth knowing before any number gets cut.
- What proportion of spend you can attribute to a team, product or customer. Below about 80%, every savings conversation is a guess.
- Your unit cost. Spend per customer, per transaction, per tenant. Total spend rising while unit cost falls is a business working correctly, and a monthly total alone cannot tell you that.
- Which spend is committed and which is elastic. Cutting committed spend usually costs more than leaving it.
- What changed. Most unexplained increases trace to a specific deployment, migration or feature enablement in the preceding fortnight, not to gradual drift.
Only after those does the question “where can we save” have a safe answer.
Six Questions That Tell You Whether You Have a Problem
Put these to whoever owns the bill. The answers take ten minutes and are more diagnostic than any tool.
- What percentage of last month’s cloud spend can you attribute to a named owner?
- What is your spend per unit of whatever your business actually sells, and is it rising or falling?
- Which AI or consumption features are enabled in your tenant, and who turned them on?
- What spending limits are set on those, and did somebody choose the number or accept a default?
- When spend rose last month, how long did it take to find out why?
- Does anyone look at SaaS and licensing alongside infrastructure, or are they different people who never meet?
Two or more vague answers is not a crisis. It does mean the next cost exercise should start with visibility rather than with cuts, because cutting what you cannot see is how availability gets damaged in pursuit of a number.
How Transputec Approaches Cloud Cost Optimisation
Transputec has run cloud and infrastructure for UK organisations for 40 years, with more than 120 specialists and a London headquarters, and cost work sits inside cloud management rather than being sold as a one-off audit. The reason is that a cost report ages badly. The engineering practice that keeps the number down is what holds.
Engagements start with an assessment rather than a proposal, and in order: attribute the spend, establish the unit cost, then act. The first two usually change what the third turns out to be.
Across the estate that means Azure and AWS managed services under the same accountability, which matters for cloud cost optimisation specifically, because multi-cloud spend is where attribution breaks down first and where two separate suppliers each report a tidy number that does not add up to the invoice.
If your estate is AWS-heavy, the platform-specific detail is in our guide to AWS cost optimisation, and the operating model behind it in AWS FinOps. This article is the vendor-neutral version of the same argument.
Where to Start
Start with the attribution number, because everything else depends on it. If you cannot say who owns 80% of last month’s bill, that is the first project, and it is unglamorous enough that it usually gets skipped in favour of something that looks faster.
Then find every consumption feature that is switched on and check whether anyone chose its spending limit. That single pass is the highest-value hour available in most estates this quarter, because it is the only control that works before the money is spent rather than after.
Then, and only then, run the classic levers across the infrastructure. They still work, they still pay back, and they are still where the quick wins are. They are simply no longer the whole job, which is the honest summary of what changed in 2026.
If you want a second pair of eyes on any of that, talk to our team about the cloud cost optimisation work behind Transputec’s cloud services.
Frequently Asked Questions
What is cloud cost optimisation?
It is the practice of reducing what you pay for cloud services without reducing what they deliver. In practice it covers right-sizing resources to actual demand, covering steady workloads with commitments, moving cold data to cheaper storage, switching off what is not in use, and attributing every pound of spend to an owner. Since 2026 it also has to cover consumption spend from AI features, which behaves differently and needs spending limits rather than resizing.
Why did cloud waste increase in 2026?
Because AI workloads arrived faster than cost controls for them. Flexera’s 2026 report found estimated wasted cloud spend rose to 29%, the first increase in five years, and attributes it to the surge in cloud-based AI. The State of FinOps survey shows why that caught people out: 98% of practitioners now manage AI spend, against 63% a year earlier, so the remit grew by a third in twelve months while the tooling and the habits stayed aimed at infrastructure.
How much cloud spend is typically wasted?
Around 29% of cloud spend is estimated to be wasted, according to 753 cloud decision makers surveyed for Flexera’s 2026 State of the Cloud Report. Treat that as an order of magnitude rather than a target: it is self-reported, it varies enormously by estate, and a mature, well-attributed environment will sit well below it. The useful question is not whether you match the average but whether you can see enough of your own spend to know.
What is the difference between cloud cost optimisation and FinOps?
Cost optimisation is the activity, FinOps is the operating model around it. Optimisation is the set of technical levers that reduce a bill. FinOps is the practice of making cost a shared responsibility between engineering, finance and the business, with allocation, forecasting and unit economics behind it, so that the savings hold instead of reappearing six months later. You can optimise without FinOps. The results tend not to last.
Where should we start if we have never done this?
With attribution, not with cuts. Work out what proportion of last month’s bill you can assign to a named team, product or customer, and get that above 80% before you change anything. Then audit which consumption and AI features are enabled and whether their spending limits were chosen or inherited. Only then run right-sizing, scheduling and commitment cover across the infrastructure. Starting with cuts in an estate you cannot see is how teams remove headroom and discover the cost of it during an incident.
This article was drafted with AI assistance and reviewed by the Transputec team. Featured image: AI-generated.
