Cloud was always an economic argument before it was a technical one. Leasing compute makes better sense than buying depreciating hardware that needs ongoing support, and that argument still holds. What has changed is that the saving is no longer automatic. It is now the return on a management discipline, and the organizations that skipped that discipline during adoption are paying for the omission every month.
The evidence is not subtle. Flexera’s 2026 State of the Cloud Report puts self-estimated waste at 29 percent of Infrastructure-as-a-Service and Platform-as-a-Service spend. That is the first increase in five years, and Flexera attributes the reversal to AI workloads. Broadcom’s Private Cloud Outlook 2025, a survey of 1,800 senior IT decision-makers, found that 94 percent report at least some public cloud waste, and 49 percent believe more than a quarter of their public cloud spend is wasted.
Scale turns a percentage into a board-level number. Flexera reports that 76 percent of large enterprises now spend more than $5 million per month on public cloud, and Gartner forecasts worldwide Infrastructure-as-a-Service end-user spending of $287 billion in 2026, up 29.3 percent from $222 billion in 2025. Waste that was tolerable at pilot scale is a material line item at production scale.
The underlying failure is usually the same. Many organizations never established a cloud cost-management framework during adoption, and a common rule of thumb holds that a company can spend up to a third of its cloud investment on services and resources that deliver less value to the business. Five operational disciplines recover most of that. None of them require re-platforming, and all five are available to any organization already running in the cloud.
Cloud spending is growing faster than cloud discipline. Flexera’s 2026 State of the Cloud Report puts self-estimated waste at 29 percent of IaaS and PaaS spend, the first increase in five years, and attributes the reversal to AI workloads.
Five operational levers recover most of it: retire idle resources, autoscale by default, automate repetitive work, use serverless where demand is intermittent, and commit deliberately through reserved and spot capacity. None require re-platforming.
What they do require is an owner. Harness found that waste takes 31 days on average to identify and eliminate, and that only 32 percent of developers work with fully automated cost enforcement. Accountability, not tooling, is the binding constraint.
By the numbers
29%
of IaaS and PaaS spend is estimated by organizations to be wasted, the first increase in five years, attributed to AI workloads
– Flexera 2026 State of the Cloud Report
94%
of senior IT decision-makers report at least some public cloud waste, and 49 percent believe more than a quarter of their spend is wasted
– Broadcom Private Cloud Outlook, 2025
76%
of large enterprises now spend more than $5 million per month on public cloud, which turns small waste percentages into material sums
– Flexera 2026 State of the Cloud Report
63%
of organizations have an established, dedicated FinOps team, up from 59 percent the prior year
– Flexera 2026 State of the Cloud Report
98%
of FinOps practitioners now manage AI spend, up from 63 percent a year earlier, across 1,192 respondents representing over $83 billion in annual cloud spend
– FinOps Foundation, State of FinOps 2026
31 days
average time to identify and eliminate cloud waste, while only 32 percent of developers have fully automated cost-enforcement practices
– Harness, FinOps in Focus 2025
Why Is Cloud Waste Rising Again After Five Years of Progress?
For five consecutive years, the self-reported waste number moved in the right direction. Organizations built tagging standards, bought visibility tools, and hired specialists, and the trend line rewarded them. Flexera’s 2026 State of the Cloud Report breaks that run. Self-estimated waste now sits at 29 percent of IaaS and PaaS spend, and Flexera attributes the increase to AI workloads.
The practitioner data tells the same story from the inside. The FinOps Foundation’s State of FinOps 2026, covering 1,192 respondents who collectively represent more than $83 billion in annual cloud spend, found that 98 percent of practitioners now manage AI spend, up from 63 percent a year earlier. In a single year, managing AI cost went from a specialist concern to a near-universal part of the job. The top reported priority for those teams is workload optimization and waste reduction.
Growth compounds the problem. Gartner’s July 2026 forecast puts worldwide Infrastructure-as-a-Service end-user spending at $287 billion for 2026, up 29.3 percent from $222 billion in 2025. Even a flat waste rate applied to a base growing at that pace produces a larger absolute loss each year. In 2026 the rate is not flat either.
Nor is this a problem confined to laggards. Broadcom’s Private Cloud Outlook 2025, which surveyed 1,800 senior IT decision-makers, found that 94 percent report at least some public cloud waste, and that 49 percent believe more than 25 percent of their public cloud spend is wasted. When almost half of senior technology leaders think a quarter of the bill is going nowhere, the issue is structural rather than local.
The response does not need to be exotic. The five steps that follow were practical when cloud adoption was the story, and they are more valuable now that consumption is the story. What has changed is the cost of neglecting them.
Steps 1 to 3: Stop Paying for Capacity Nobody Is Using
The first three steps address money already committed and quietly going nowhere. They are the cheapest to implement, the fastest to show a result, and the least disruptive to delivery teams. Start here.
Step 1. Retire unused cloud resources and instances
Virtual machines remain expensive even when idle or entirely unused. Around 15 percent of cloud cost is commonly associated with unattended and inactive services, because those services continue leasing compute power whether or not anyone touches them. The best practice is unglamorous: terminate what is not being used.
Automated scheduling is the most cost-effective way to do it. It is a one-time setup that requires no ongoing human monitoring. Configure instances to run only during specific windows, weekdays only for a development environment, for example, and group worker nodes so the schedule applies cleanly. The saving recurs every month without anyone remembering to claim it.
Step 2. Always autoscale cloud resources
Without an auto-scaling mechanism, an organization forgoes one of the defining benefits of cloud. Auto-scaling purchases capacity as it is needed, scaling up and down with demand. That is the exact opposite of the common mistake, which is buying excess capacity upfront and treating a leased platform like an owned data centre.
Implementation is well established. When system thresholds are reached, additional instances activate automatically to distribute the traffic load. During low-traffic periods, the scaled instances shut down automatically and the expense stops with them. The business outcome is a cost curve that follows demand instead of anticipating it.
Step 3. Automate the repetitive work
Cloud automation tools should own the repetitive tasks: virtual machine backups, storage unit backups, source code deployments, unit test execution, and security and compliance monitoring. Automating them minimizes human involvement and allows teams to focus on strategic business processes rather than on maintenance that a script performs more reliably.
Terraform is the standard example. It automates infrastructure-level deployment using source code, enabling single-cloud to multi-cloud configuration through scripts and significantly accelerating provisioning. The cost benefit is indirect but durable: infrastructure defined in code is infrastructure that can be reviewed, versioned, and decommissioned deliberately rather than forgotten.
Idle compute is not an accident. It is a decision someone made once and nobody revisited.
Every unattended instance was provisioned deliberately, for a reason that was valid at the time. Waste accumulates because there is no scheduled moment at which that reason is re-examined. Automated scheduling and scaling policies exist to make that re-examination continuous rather than annual.
Steps 4 and 5: Buy Compute the Way the Business Actually Consumes It
The first three steps remove spend that produces nothing. The last two change the price of the spend that remains. Both depend on knowing how predictable and how interruption tolerant each workload really is, which is a business classification exercise as much as a technical one.
Step 4. Serverless computing and Function as a Service
Serverless computing, delivered as Function as a Service, charges only for actual usage time. Services trigger automatically when they are needed and shut down when the work completes, using precise resources and avoiding unnecessary modules running continuously in the background.
For a finance leader, the significance is the shape of the cost rather than its size. Serverless converts a fixed infrastructure commitment into a variable one that moves with transaction volume. For workloads that are intermittent, event-driven, or seasonal, that alignment between cost and activity is worth more than a negotiated discount on capacity that sits idle between peaks.
Step 5. Utilize reserved and spot cloud instances
Reserved instances suit steady demand. Committing to fixed resources over time proves cost-effective when the workload is predictable, and the commonly quoted range is savings of up to 40 percent for a one-year commitment and up to 60 percent over three years. Amazon Web Services publishes a maximum of up to 66 percent for its Compute Savings Plans against On-Demand pricing. Treat that as a vendor ceiling on an ideal commitment, not as an average an enterprise should expect to achieve.
Spot instances work differently. They involve purchasing a provider’s unused capacity at a steep discount, typically cited at 70 to 90 percent below the price of an equivalent-sized regular instance, with Amazon Web Services publishing a maximum of up to 90 percent off On-Demand for EC2 Spot Instances. That figure is also a ceiling rather than a norm. The trade-off is real: spot capacity can be reclaimed and terminated with little warning.
Two mitigations make the risk manageable. Limit spot usage to non-critical tasks such as batch processing and test environments, or use auto-scaling to maintain the required instance count so that a reclaimed instance is replaced rather than simply lost. The second mitigation is why Step 2 and Step 5 reinforce each other.
| Lever | What it recovers |
|---|---|
| 1. Retire idle resources | Spend on unattended and inactive services, commonly put at around 15 percent of cloud cost, which continues to accrue whether or not anyone is using them |
| 2. Autoscale by default | The gap between capacity purchased in advance and capacity actually demanded, corrected continuously in both directions |
| 3. Automate repetitive work | Engineering hours lost to backups, deployments, test execution, and compliance monitoring, plus the drift that manual provisioning introduces |
| 4. Serverless and Function as a Service | Charges for idle runtime, because billing applies only to actual execution time on intermittent and event-driven workloads |
| 5. Reserved and spot instances | The premium paid for on-demand flexibility a workload does not need. AWS publishes ceilings of up to 66 percent for Compute Savings Plans and up to 90 percent for EC2 Spot Instances |
Why the Discipline Fails Without a Named Owner
None of the five steps is a secret. Every cloud platform documents them, every provider sells tooling for them, and most engineering teams can describe them accurately. Organizations still fail to execute them consistently, which suggests the constraint is not knowledge. It is accountability.
Harness, in its FinOps in Focus 2025 report, quantified the gap. The average time to identify and eliminate cloud waste is 31 days. More than half of engineering leaders, 52 percent, say a disconnect between FinOps and developers drives waste. Only 32 percent of developers have fully automated cost-enforcement practices in place.
The 31-day figure is the one to put in front of a leadership team. A month is long enough for a misconfigured environment to consume a meaningful share of a quarterly budget before anyone notices, and long enough that the engineer who provisioned it has moved on to other work by the time the invoice raises the question.
Cloud cost management fails at the handoff, not at the tooling.
Finance can see the invoice but cannot change the architecture. Engineering can change the architecture but rarely sees the invoice in time to act on it. The 31-day detection lag Harness reports is the measurable shape of that gap, and no dashboard closes it on its own.
ML arteka treats cloud cost as an engineering control rather than a reporting exercise, placing enforcement inside the delivery pipeline where provisioning decisions are actually made.
The organizational answer is beginning to take hold. Flexera reports that 63 percent of organizations now have an established, dedicated FinOps team, up from 59 percent the prior year. The FinOps Foundation reports that 78 percent of FinOps teams report to the CTO or the CIO, which places cost accountability inside the technology function rather than alongside it in finance.
That reporting line matters more than the org chart suggests. A team that reports into technology can require a scaling policy before a workload reaches production. A team that reports into finance can only describe what already happened. Given that 98 percent of FinOps practitioners now manage AI spend, the difference between prevention and description is about to get expensive.
How Do You Know Whether Cloud Cost Management Is Working?
A falling cloud bill is not proof of success, and a rising one is not proof of failure. Growth, migration, and AI adoption all move the total in ways that have nothing to do with discipline. Leadership needs measures that separate the spend that produces value from the spend that does not.
| Metric | What it tells leadership |
|---|---|
| Waste rate | The share of IaaS and PaaS spend not attributable to a running, used workload. Flexera’s 2026 self-reported benchmark is 29 percent, so anything materially above that is a priority rather than a peer-group norm |
| Time to detect and eliminate waste | How long an idle or oversized resource survives before someone acts. Harness reports an average of 31 days for 2025. Weeks is common. It should not be acceptable |
| Commitment coverage | The share of steady-state compute running on reserved capacity or savings plans rather than on-demand pricing, and whether that share tracks the predictability of the workload |
| Spot eligibility and use | How many workloads have been formally classified as interruption tolerant, and how much of that eligible capacity actually runs on spot instances |
| Autoscaling coverage | The share of production workloads governed by active scaling policies rather than fixed provisioning decided at launch |
| Automated cost enforcement | The share of developers working in pipelines that apply cost policy automatically. Harness put this at 32 percent in 2025, which sets a low bar to clear |
| Unit economics | Cloud cost per transaction, per customer, or per product, so growth in spend can be distinguished from waste in spend |
Report these together, quarterly, to both the technology leadership team and the finance function. A single set of numbers reviewed by both groups removes the argument about whose problem the bill is.
Five Leadership Takeaways
Before your next infrastructure budget review. Before your next AI platform approval.
- The saving is a discipline, not a property of the platform. Cloud makes lower cost possible. Only management makes it happen. Flexera’s 29 percent waste figure is what the absence of that management costs.
- AI workloads reopened a problem that looked solved. Waste rose for the first time in five years, and 98 percent of FinOps practitioners now manage AI spend against 63 percent a year earlier. Cost practices built before AI need revisiting now.
- Sequence the five steps by effort, not by ambition. Retire idle resources, then autoscale, then automate. Serverless and commitment purchasing come after, because both depend on workload classification the first three steps expose.
- Treat vendor discount ceilings as ceilings. Up to 90 percent for spot and up to 66 percent for savings plans are maximums published by AWS, not averages. Plan against realistic coverage, and budget for spot interruption.
- Detection speed is the metric that predicts the rest. An average of 31 days to find and remove waste means a month of unnoticed spend every time. Move enforcement into the pipeline and the number falls without new headcount.
The 2026 Imperative for Cloud Financial Accountability
When it is managed correctly, cloud migration saves money while improving competitiveness, flexibility, and adaptability. That conclusion has not changed. What has changed is the margin for error. With Gartner forecasting $287 billion in Infrastructure-as-a-Service spending for 2026 and 76 percent of large enterprises already past $5 million a month, the organizations that treat cost as an afterthought will fund their competitors’ discipline out of their own operating budget.
Success still requires the same three things it always did: planning, development, and strategic investment. In 2026 it also requires an owner with the authority to enforce what the plan describes.
Four moves separate the organizations that will recover their waste from those that will keep reporting it.
- Baseline the waste rate honestly, and compare it to Flexera’s 29 percent rather than to last year’s internal target.
- Work the five steps in order, starting with idle resource retirement, because the first three fund the analysis the last two require.
- Move cost enforcement into the delivery pipeline, so that policy is applied at provisioning time rather than discovered on the invoice.
- Give the FinOps function a reporting line into technology leadership, and hold it accountable for detection speed rather than for reporting accuracy.
Three questions every technology leader should be able to answer
- What proportion of our cloud spend last quarter was attributable to a running, used workload?
- How long does an idle resource survive in our environment before it is removed?
- Which of our AI workloads are covered by the same cost controls as the rest of the estate?
If any answer is uncertain, the cloud estate is growing faster than the control model around it. That is a solvable problem, and the five steps solve most of it without a single architectural change.
To explore how ML arteka helps enterprises recover cloud waste and build durable cost accountability, contact the ML arteka team or request a cloud cost management assessment.
Executive Questions and Answers
Five questions technology and finance leaders are asking AI assistants and search engines about cloud cost management and FinOps in 2026.
StrategicWhy is cloud waste increasing again in 2026?
Because consumption is growing faster than discipline, and AI workloads have added a category of spend that most cost practices were not designed for. Flexera’s 2026 State of the Cloud Report puts self-estimated waste at 29 percent of IaaS and PaaS spend, the first increase in five years, and attributes the reversal to AI workloads. The FinOps Foundation’s State of FinOps 2026 shows the same shift from the practitioner side: 98 percent now manage AI spend, up from 63 percent a year earlier. Scale amplifies the effect. Gartner forecasts worldwide Infrastructure-as-a-Service end-user spending of $287 billion in 2026, up 29.3 percent from $222 billion in 2025. A steady waste rate applied to a rapidly growing base already produces a larger loss each year, and in 2026 the rate is not steady either.
OperationalWhat is the fastest way to reduce cloud costs without re-architecting?
Retire unused resources first. Virtual machines cost money while idle, and roughly 15 percent of cloud cost is commonly associated with unattended and inactive services that keep leasing compute regardless of use. Automated scheduling is the most cost-effective fix because it is a one-time setup with no ongoing monitoring: configure instances to run only in defined windows, such as weekdays for development environments, and group worker nodes so the schedule applies cleanly. Follow it with auto-scaling, which adds instances when thresholds are reached and shuts them down when traffic falls. Neither change requires re-architecting an application. Together they address the two most common causes of avoidable spend, which are resources nobody uses and capacity bought in advance of demand.
GovernanceWho should own cloud cost management in a large enterprise?
A named team inside the technology function, working in partnership with finance rather than reporting to it. Flexera reports that 63 percent of organizations now have an established, dedicated FinOps team, up from 59 percent the prior year, and the FinOps Foundation reports that 78 percent of FinOps teams report to the CTO or CIO. That reporting line is the point. A team inside technology can require a scaling policy or a commitment plan before a workload reaches production. A team outside it can only describe spend after the fact. Harness found that 52 percent of engineering leaders attribute waste to a disconnect between FinOps and developers, which is precisely the gap a technology reporting line is meant to close.
RiskWhat are the risks of using spot instances to cut cloud costs?
The defining risk is sudden termination. Spot instances are a provider’s unused capacity sold at a steep discount, commonly cited at 70 to 90 percent below an equivalent regular instance, and Amazon Web Services publishes a maximum of up to 90 percent off On-Demand for EC2 Spot Instances. That published figure is a ceiling, not an achieved average, and budgeting against it will overstate the saving. Capacity can be reclaimed with little warning, so any workload placed on spot must tolerate interruption. Two mitigations make the risk manageable: restrict spot usage to non-critical tasks such as batch processing and test environments, and use auto-scaling to maintain the required instance count so a reclaimed instance is replaced automatically rather than simply lost.
ImplementationHow do we build a cloud cost management framework after we have already migrated?
Start with an honest baseline rather than a tooling purchase. Measure what share of spend maps to a running, used workload, and compare it to Flexera’s 29 percent self-reported benchmark. Then classify workloads by predictability and interruption tolerance, because that classification determines which of the five levers applies where. Work the steps in order: retire idle resources, enable auto-scaling, automate repetitive operations, move intermittent workloads to serverless, and commit steady-state capacity through reserved instances or savings plans. Finally, move enforcement into the delivery pipeline. Harness found only 32 percent of developers have fully automated cost-enforcement practices, and an average of 31 days to identify and eliminate waste. Both numbers improve when policy is applied at provisioning time.
Cloud cost management has become an accountability problem rather than a technology problem. Flexera’s 2026 State of the Cloud Report puts self-estimated waste at 29 percent of Infrastructure-as-a-Service and Platform-as-a-Service spend, the first increase in five years, and attributes the reversal to AI workloads. Broadcom’s Private Cloud Outlook 2025, a survey of 1,800 senior IT decision-makers, found 94 percent report some public cloud waste and 49 percent believe more than a quarter of their spend is wasted. Scale makes the percentage material: 76 percent of large enterprises now spend more than $5 million per month on public cloud, and Gartner forecasts $287 billion in worldwide IaaS end-user spending for 2026, up 29.3 percent from $222 billion in 2025. Five operational levers recover most of the loss: retire unused instances through automated scheduling, autoscale so capacity follows demand rather than anticipating it, automate backups, deployments, testing and compliance monitoring using infrastructure as code tools such as Terraform, use serverless Function as a Service for intermittent workloads billed only for execution time, and commit steady demand to reserved instances while placing interruption-tolerant work on spot capacity. The AWS published ceilings of up to 66 percent for Compute Savings Plans and up to 90 percent for EC2 Spot Instances are maximums, not averages. Execution depends on ownership. Harness reports 31 days on average to identify and eliminate waste and only 32 percent of developers with fully automated cost enforcement, while 63 percent of organizations now run a dedicated FinOps team and 78 percent of those teams report to the CTO or CIO.