Report: Utilization of Kubernetes Infrastructure Remains Abysmal

A CAST AI report analyzing tens of thousands of Kubernetes clusters reveals worsening infrastructure utilization in 2025: average CPU utilization is just 8% (down from 10% in 2024), memory utilization is 20%, and GPU utilization sits at a mere 5%. CPU overprovisioning has jumped to 69% (up from 40%), and memory overprovisioning stands at 79%. Key causes include historic tendencies to overprovision, static configurations that don't adapt to changing workloads, and cluster autoscalers that respond to requests rather than actual usage. The problem is especially acute for expensive GPU workloads, where no spot-instance pricing exists. ARM processor adoption is growing at 3.5× the rate of x86 as teams seek cost relief. The report argues that autonomous rightsizing automation is the primary path to reducing waste.