Scale-to-zero for GPU-heavy workloads like model-serving endpoints and batch jobs reduces idle GPU spend by releasing resources at the pod level, rather than waiting for node drains.

The rollout carries inherent risks: beta features remain default-enabled but may affect upgrades for existing HPAs and depend on vendors delivering updated DRA drivers for true cost benefits.

Looking ahead, expect broader partitionable-device support, faster movement of scale-to-zero to stable, quicker DRA driver releases, clearer EKS timelines, and new FinOps tooling tailored to scale-to-zero.

Kubernetes 1.37, released late August 2026 (Garhwal), introduces scale-to-zero for HPA, enabling exact zero pods when demand drops and automatic resumption when traffic returns.

Dynamic Resource Allocation for extended resources, including GPUs, reaches general availability, letting DRA drivers satisfy traditional requests directly without rewriting pod specs.

The release brings security and identity improvements, with Pod Certificates and ClusterTrustBundles stabilizing, plus a new ulimits field in SecurityContext for pod-level resource-limit policies.

Native scale-to-zero covers basic cases, but KEDA still adds value for broader event-driven autoscaling; DRA complements node-level autoscaling by enabling fine-grained device release before node decisions.

DRA features introduce device taints, NUMA awareness, and derivedAttributes to standardize hardware placement data across vendors, reducing vendor lock-in.

GA of DRA and scale-to-zero shifts FinOps from detecting idle GPU usage to preventing idle costs, though cold-start latency remains a consideration for latency-sensitive workloads.

Configuring scale-to-zero involves setting minReplicas: 0 in the existing HPA and pairing with a DRA GPU claim to ensure the accelerator is released when pods terminate.

Cloud provider rollout timelines vary: AKS preview in September 2026 with GA in October 2026; GKE in rapid/early-access by late September 2026; EKS 1.37 GA date not yet confirmed.

Even though scale-to-zero is in beta, default-enabled, it makes idle GPU-backed workloads highly cost-effective by freeing GPU claims immediately as pods scale to zero.