Cloud computing has changed the way companies purchase and consume IT resources. However, the ability to provision capacity quickly and scale it as needed has an inevitable consequence: not everything purchased is used efficiently.
This is where cloud waste comes in, a topic central to optimization strategies and FinOps practices for years. In 2026, the problem is certainly better understood and managed than in the past, but it is far from being solved. On the contrary, the transition of AI from experimentation to production is introducing new cost variables and new complexities to manage. Let's take a closer look.
Key points
- Cloud waste is cloud spending that does not translate into proportional resource utilization. It can result from overprovisioning or purchasing models that do not match the actual workload profile.
- The problem is growing, and AI introduces additional layers of complexity. Training, inference, and agentic systems make it harder to accurately predict consumption.
- Optimization must start with solution architecture. Sizing, inference methods, autoscaling, resource utilization, and consumption models are all decisive factors.
Cloud waste at 29% in 2026: what does this really mean?
According to the Flexera State of the Cloud 2026, 29% of organizations' IaaS and PaaS spending is wasted.
Before interpreting this figure, however, it is worth clarifying what cloud waste actually encompasses. It is not just about forgotten virtual machines left running or storage that nobody uses. Waste can arise, for example, from:
- Resources that are oversized relative to the actual workload;
- Duplicated services;
- Test environments that remain active longer than necessary;
- Idle or unused resources that continue to generate costs;
- Purchasing models that are not cost-effective given the actual needs of workloads.
Overprovisioning, a phenomenon that predates cloud computing, remains one of the most evident manifestations of cloud waste. CAST AI's 2026 State of Kubernetes Optimization Report reports average CPU utilization of 8% and memory utilization of 20% across the clusters analyzed, while CPU overprovisioning has risen to 69%. A similar pattern emerges from Datadog's State of Cloud Cost, which indicates that a very substantial share of container costs (more than 80%) is associated with idle resources or capacity requests that exceed actual usage.
It is important to emphasize that some excess capacity serves a specific purpose: absorbing spikes, ensuring performance, maintaining safety margins, or allowing autoscaling to respond quickly. Not every gap between purchased and utilized capacity is therefore waste; the problem arises when that margin is not intentional, measured, and justified by the value it delivers.
The 29% cloud waste figure mentioned above should therefore be interpreted as the sum of normal operational overhead and many different forms of inefficiency, which may relate to provisioning, actual utilization, and purchasing models.
The close relationship between AI and resource waste
The race to adopt AI adds a new layer of complexity to cloud cost management. Not so much because AI is inherently inefficient, but because it introduces variable workloads, expensive infrastructure, and consumption patterns that are extremely difficult to predict.
The first factor is the capacity required. Gartner estimates that in 2026, global spending on AI-optimized IaaS will grow by 96.4%, driven by both training and inference. The latter is what changes the picture most significantly: training is concentrated in specific phases, whereas inference consumes resources every time a model is used in production. More users, requests, agents, and automations therefore mean greater ongoing consumption of cloud capacity.
To ensure performance and response times, companies may acquire more GPU capacity than necessary, choose oversized instances, or keep resources available even when utilization levels do not justify them.
How to reduce cloud waste in the age of AI
With AI, spending visibility and consumption monitoring are not enough. The real leap forward comes from designing workloads so that they use expensive capacity only when needed.
Optimization must therefore be embedded in architectural decisions: inference methods, instance sizing, autoscaling, request batching, model selection, and intelligent data pipeline management. Designing complex systems while relegating economic sustainability to a secondary concern can lead to unpleasant surprises.
Choose inference methods and capacity based on actual demand
Not all AI requests need to be handled in the same way. Some require real-time responses and low latency; others can be processed asynchronously or in batches. Similarly, not all workloads justify keeping resources constantly active or using frontier models.
- The first optimization lever is therefore to align inference methods, active capacity, and the actual demand profile. For intermittent workloads, it may make sense to adopt services that can scale to zero when there are no requests; for deferrable processing, compute resources can be activated only when a job starts.
- Even when real-time processing is required, sizing the environment for the theoretical maximum peak risks leaving GPUs and other resources underutilized for extended periods. Autoscaling and capacity planning should therefore be based on indicators that closely reflect actual workload behavior, such as concurrent requests, queues, throughput, and tokens processed.
Measure cost per unit of work
With AI, looking only at total spending can be misleading. An increase in costs may simply reflect the fact that more users, applications, or processes are using the service. In this case, spending more does not mean being less efficient.
A new dimension needs to be introduced: cost per unit of work. Depending on the use case, this could mean the cost per request processed, document analyzed, conversation handled, ticket resolved, or task completed.
This comparison makes it possible to understand whether the workload is scaling effectively and at the right cost. If spending doubles because the amount of work processed also doubles, efficiency may have remained unchanged. If, on the other hand, costs grow faster than the volume of work produced, it means that something in the architecture, the model being used, or resource consumption is becoming less efficient.
Build a sustainable consumption model
In the cloud, costs depend not only on how many resources are used, but also on how they are purchased. Keeping everything on demand provides maximum flexibility, but it can be very expensive when part of the AI workload has sufficiently stable and predictable usage.
This is where the ability to build an intelligent mix of elastic consumption and committed capacity comes into play. The variable portion, subject to spikes or difficult to predict, can remain on demand; the more stable portion can instead benefit from options such as Savings Plans, Reserved Instances, Committed Use Discounts, or equivalent mechanisms offered by cloud providers.
Get value from AI without losing control of costs
At Kirey, we work to ensure that digital transformation becomes a tangible driver of competitive advantage, capable of generating long-term value, including through the economic sustainability of projects.
With AI, this balance becomes even more important. It is easy to be guided by the technology's potential and relegate cost optimization to the background. This applies to individual workloads, but even more so to complex architectures and agentic systems, in which models, components, and agents interact continuously, generating new cloud resource consumption.
Our experience allows us to address this challenge from the very earliest architectural and design decisions, because the cost of an AI solution depends not only on the resources purchased and used, but also on how they are sized, orchestrated, and integrated into the overall solution.
Contact us to design a roadmap together that brings innovation, control, and business results into alignment.
