Report: Utilization of Kubernetes Infrastructure Remains Abysmal

Why It Matters

Chronic under‑utilization inflates cloud spend, eroding margins for enterprises that rely on Kubernetes for AI and other workloads. Improving utilization through automation is now a cost‑control imperative as cloud pricing, especially for GPUs, continues to rise.

Key Takeaways

• Average CPU utilization in Kubernetes clusters fell to 8% in 2025
• Memory usage averaged only 20%, with GPU utilization at 5%
• CPU over‑provisioning rose to 69%, memory over‑provisioning to 79%
• CAST AI users cut provisioned CPU by roughly 50% after automation
• ARM processor adoption grew 3.5× faster than x86, now 9% of fleet

Pulse Analysis

The latest CAST AI analysis underscores a paradox in modern cloud operations: despite the explosive growth of Kubernetes‑driven microservices, the underlying infrastructure sits largely idle. With CPU usage averaging a single‑digit 8% and memory barely reaching one‑fifth of capacity, enterprises are paying for resources they rarely consume. This inefficiency is magnified in AI workloads, where GPU utilization hovers at a meager 5%, even as providers hike prices and offer no spot‑market relief.

Root causes trace back to entrenched over‑provisioning habits and a lack of dynamic right‑sizing. Legacy Helm charts and shared manifests embed conservative resource requests that propagate across new deployments, while cluster autoscalers react to demand spikes rather than sustained usage patterns. The report shows CPU over‑provisioning at 69% and memory at 79%, indicating that most clusters are oversized by a wide margin. Automation platforms like CAST AI can halve the provisioned CPU footprint by continuously reconciling actual consumption with declared limits, a benefit that becomes critical as GPU costs surge and organizations scramble to justify AI ROI.

For decision‑makers, the data signals a clear mandate: invest in continuous optimization tooling or face escalating cloud bills with diminishing returns. The accelerating shift toward ARM processors—now 9% of the fleet and growing 3.5‑times faster than x86—offers a secondary lever for cost reduction, given their lower power and licensing expenses. As economic uncertainty persists, the pressure to extract value from every compute cycle will drive broader adoption of autonomous scaling solutions, reshaping how enterprises manage Kubernetes at scale.