Google announced several major GKE enhancements at Cloud Next '26 targeting AI and agentic workloads. Key releases include: GKE Agent Sandbox using gVisor isolation for secure, high-throughput agent execution (300 sandboxes/sec at sub-second latency); GKE hypercluster enabling a single Kubernetes control plane to manage 1 million chips across 256,000 nodes spanning multiple regions; inference performance improvements via ML-driven predictive latency routing (up to 70% TTFT reduction) and automatic KV Cache tiering; new reinforcement learning capabilities including RL Scheduler, RL Sandbox, and observability dashboards; and intent-based autoscaling with native custom metrics support for HPA, cutting reaction times from 25s to 5s.
Table of contents
GKE Agent Sandbox: Accelerating the agentic eraGKE hypercluster redefines the scalability ceilingSupercharging state-of-the-art inferenceEliminating RL compute bottlenecksIntent-based autoscaling on custom metricsNew workloads, same missionSort: