Report: Utilization of Kubernetes Infrastructure Remains Abysmal
An analysis of tens of thousands of Kubernetes clusters deployed on cloud services published today by CAST AI finds average CPU utilization stood at just 8% in 2025, while memory utilization was 20%.
More troubling still, utilization rates of expensive graphical processor units (GPUs) stood at 5%, the analysis finds.
CAST AI president Laurent Gil said the 2026 edition of the company’s State of Kubernetes Optimization Report shows that instead of improving utilization rates in the past year things are actually getting worse because the amount of IT infrastructure being wasted has slightly increased. In 2024, CPU utilization rates stood at 10%, while memory utilization stood at 23%.
Overall, the analysis finds CPU overprovisioning stands at 69%, up from 40% a year ago. Memory overprovisioning, meanwhile, stands at 79%.
Much of that waste can be attributed to a historic tendency to overprovision IT infrastructure resources and a lack of appreciation of how cloud-native application workloads change and evolve, said Gil. Workloads change, traffic patterns shift, and the configuration that was accurate six months ago is unlikely to remain accurate, he added.
IT teams rarely revisit deployments to optimize utilization, said Gil. Worse yet, resource definitions propagate through Helm charts and shared manifests, spreading conservative estimates across new deployments. Cluster autoscalers respond to requests, not actual usage, so in the absence of an automation framework there is no means to rightsize clusters, noted Gil.
IT teams that have deployed the CAST AI platform, for example, have on average seen a ~50% reduction in provisioned CPU footprint.
That issue becomes especially acute when deploying AI applications on GPUs, noted Gil. Cloud service providers have raised GPU pricing. There is no such thing as a spot instance of GPU cloud service so the only way to minimize the impact of higher pricing is to rely more on autonomous optimization to increase utilization rates, said Gil.
As cloud infrastructure costs continue to rise there is little doubt IT teams are considering their options. In fact, the CAST AI report notes that since the second quarter of 2024, adoption of ARM processors has grown at 3.5× the rate of x86 processors. Arm processors now account for 9% of the total CPU fleet.
Whether IT teams embrace automation to rightsize Kubernetes clusters remains to be seen. However, during uncertain economic times, there is always increased pressure on IT teams to find ways to manage infrastructure more efficiently as organizations look to balance competing priorities.
In the meantime, however, it would appear that utilization rates in the age of AI might even wind up being lower than they already are. The issue then becomes, given the cost of AI, how long it will be before business and IT leaders decide they are no longer going to tolerate that level of waste at a time when many organizations are still trying to determine the actual level of return on investment (ROI) being derived from their AI investments.
