GKE Is Built on Google’s Kubernetes Expertise – So Why is it Still Underutilized?

Google created Kubernetes, but it still struggles with underutilization. The 2026 State of Kubernetes Optimization Report shows average CPU utilization at just 8%, memory utilization at 20%, and GPU utilization at 6% on GKE. CPU overprovisioning reached 69% while memory overprovisioning reached 79%. The issue lies in workload requests, which determine whether resources are truly needed.

In Autopilot, oversized requests directly increase costs, while in Standard, they affect scheduling and can lead to more node capacity being provisioned than what the application consumes. Rightsizing resource requests is crucial for cost optimization, but it requires understanding actual workload behavior, not just choosing conservative numbers.

The report recommends starting with a small group of workloads, establishing their actual usage, and adjusting requests based on observed behavior. This approach helps avoid changing hundreds of deployments at once. Rightsizing doesn't necessarily mean sacrificing reliability; it can even reduce over-provisioning and OOM kills. However, not all workloads are suitable for rightsizing, such as those that require certain resources due to production variability.

Cluster Autoscaler and Node Auto Provisioning can help optimize the underlying Compute Engine fleet, but they cannot compensate for inflated workload requirements. Spot VMs can provide cost savings for workloads that can tolerate interruption, but they require careful consideration and appropriate candidates. Ultimately, the key is to base decisions on production data rather than assumptions, and to find the right balance between cost and reliability.