The Stack Observer
Cloud Native / Kubernetes

Kubernetes v1.36 and the AI Infrastructure Revolution: What You Need to Know

•Stackxx•Cloud Native, Kubernetes

Kubernetes has long been the orchestration backbone of cloud-native infrastructure, but May 2026 has delivered a wave of updates that underscore a critical inflection point: Kubernetes is no longer just the platform for containers — it is becoming the operating system for AI. From the upstream Kubernetes v1.36 release bringing production-grade control plane safety features, to Google Cloud Next ’26 unveiling massive-scale AI infrastructure on GKE, to AWS warning of a major container image ecosystem shift, the Kubernetes landscape is evolving at an unprecedented pace.

Kubernetes v1.36: Safer Upgrades and Smarter Cloud Integration

The Kubernetes project released v1.36 this month, and while it may not grab headlines with splashy new user-facing features, the improvements under the hood are exactly the kind of evolutionary progress that keeps the platform reliable at scale. Two features in particular deserve attention from cluster operators.

Mixed Version Proxy Graduates to Beta

First introduced as an Alpha feature in Kubernetes 1.28, the Mixed Version Proxy (MVP) has graduated to Beta in v1.36 and is now enabled by default. This feature solves a deceptively simple problem that has plagued highly available control planes for years: during rolling upgrades, different API servers run different versions, and a client request landing on an older server for a resource it does not yet know about returns an incorrect 404 Not Found. The consequences can be severe — mistaken garbage collection, blocked namespace deletions, and confused controllers.

MVP fixes this by transparently proxying requests to a peer API server that can serve the requested resource. In v1.36, the implementation has been modernized significantly. The Alpha version relied on the StorageVersion API to discover peer capabilities, which had a notable limitation: it did not support CRDs or aggregated APIs. The Beta version replaces this with Aggregated Discovery, allowing API servers to dynamically understand what their peers can serve. Additionally, v1.36 introduces Peer-Aggregated Discovery, meaning discovery requests now return a unified view of all APIs available across the entire cluster, regardless of which API server a client happens to connect to.

For operators running multi-master clusters, the practical implication is clear: rolling upgrades just became safer. The feature does require proper configuration — specifically the --peer-ca-file flag must be set to establish secure TLS communication between API servers — but the default enablement means most clusters will benefit automatically once flags are configured.

Watch-Based Route Reconciliation Gets Observable

Kubernetes v1.36 also introduces a new alpha counter metric, route_controller_route_sync_total, in the Cloud Controller Manager (CCM). This metric was added to help operators validate the watch-based route reconciliation feature gate introduced in v1.35, which switches the route controller from a fixed-interval loop to a watch-based approach that only reconciles when nodes actually change.

The impact is tangible for cloud operators: in stable clusters where nodes rarely change, the old fixed-interval loop could generate 60 or more unnecessary sync calls every ten minutes. With watch-based reconciliation enabled, that number drops to one — or zero — dramatically reducing pressure on rate-limited cloud provider APIs. The new metric makes it straightforward to A/B test this behavior and prove the efficiency gains.

GKE at Next ’26: Kubernetes as the AI Operating System

If upstream Kubernetes is the steady foundation, Google Kubernetes Engine is where the platform’s AI future is being aggressively built. At Google Cloud Next ’26, GKE announcements made it unmistakably clear that Google views Kubernetes as the operating system for the AI era — and the numbers back that up. GKE now powers AI workloads for all of Google’s top 50 platform customers, including the largest frontier model builders.

GKE Hypercluster: A Million Accelerators, One Control Plane

The headline announcement was GKE hypercluster, entering private general availability. It allows a single, Kubernetes-conformant GKE control plane to manage up to one million accelerators distributed across 256,000 nodes spanning multiple Google Cloud regions. For organizations that have been fracturing their infrastructure into hundreds of disconnected clusters to meet AI scale demands, hypercluster promises to collapse that operational complexity into a single, unified capacity reserve.

Security at this scale is addressed through Google’s Titanium Intelligence Enclave, a software-hardened security engine that delivers hardware-attested, pod-level isolation. The model ensures proprietary weights and prompts remain cryptographically sealed from platform administrators — a critical requirement for frontier model builders.

GKE Agent Sandbox: The Infrastructure for Autonomous Agents

As AI evolves from simple conversational chatbots to ecosystems of proactive, autonomous agents collaborating on complex tasks, infrastructure must handle hundreds or thousands of isolated execution environments with minimal latency. GKE Agent Sandbox, built on gVisor kernel isolation — the same technology securing Gemini — delivers 300 sandboxes per second at sub-second latency, with up to 30% better price-performance on Axion processors compared to other hyperscalers.

What makes Agent Sandbox particularly compelling for Kubernetes operators is that it addresses a genuine operational gap: running untrusted code, third-party tools, and entire agent workflows safely without sacrificing cluster performance. In traditional container environments, achieving this level of isolation often meant accepting significant latency overhead or complex sidecar architectures. Agent Sandbox collapses that trade-off, making it feasible to host multi-tenant agent workloads on shared GKE clusters without compromising security boundaries.

Companies like Lovable, which sees over 200,000 new AI-generated projects daily, are already running on GKE Agent Sandboxes precisely for the fast startup, rapid scaling, and secure isolation it provides. Lovable’s co-founder Fabian Hedin noted that GKE’s sandboxing capabilities allow them to reliably scale to hundreds of secure sandboxes per second, ensuring seamless builder experiences even during massive, unpredictable demand spikes.

llm-d Joins the CNCF Sandbox

A significant open-source development announced at Next ’26 is llm-d, officially accepted as a CNCF Sandbox project. Founded by Google Cloud alongside Red Hat, IBM Research, CoreWeave, and NVIDIA, llm-d aims to make Kubernetes the universal orchestrator for distributed AI inference under the banner “any model, any accelerator, any cloud.” The project’s acceptance into the CNCF represents a broader industry consensus that open standards, rather than vendor-specific walled gardens, should define the future of AI infrastructure.

llm-d powers GKE Inference Gateway’s intelligent routing, using real-time KV-cache hit rates, inflight request counts, and queue depth to route each request to the optimal backend. Google Cloud’s Vertex AI team validated this in production, demonstrating over 35% Time-to-First-Token latency reduction for context-heavy coding workloads, and a 52% improvement in P95 tail latency for bursty chat traffic. The gateway also doubled prefix cache hit rates from 35% to 70%, directly lowering re-computation overhead and cost-per-token.

Complementing this routing intelligence, Google leads development of the Kubernetes LeaderWorkerSet (LWS) API, which enables llm-d to orchestrate wide expert parallelism and disaggregate compute-heavy prefill and memory-heavy decode phases into independently scalable pods. Google has also extended vLLM natively for Cloud TPUs with a unified PyTorch and JAX backend, delivering up to 5x throughput gains compared to earlier releases. Together, these advancements help ensure that whether you are scaling on Google Cloud TPUs or NVIDIA GPUs, state-of-the-art AI serving remains a highly optimized, accelerator-agnostic capability.

The Bitnami ECR Public Removal: A Supply Chain Wake-Up Call

Not all Kubernetes news this month was about new features. AWS published an urgent advisory that Bitnami container images will be removed from Amazon ECR Public Gallery on June 10, 2026. With 317 repositories currently hosted, this affects a significant swath of containerized workloads running on Amazon ECS, EKS, and any CI/CD pipeline pulling from public.ecr.aws/bitnami/.

The risk is real but manageable: running containers with locally cached images will continue to operate, but any event triggering a fresh pull — container crashes, scaling events, rolling updates, node replacements — will fail. AWS recommends identifying affected workloads, mirroring required images to private ECR repositories, and implementing image caching strategies to insulate against future upstream changes.

For the Kubernetes community, this is a reminder that free public image registries are not guaranteed infrastructure. Organizations running production Kubernetes workloads should treat upstream image availability as a dependency requiring active management, mirroring, and fallback planning.

Red Hat’s Virtualization Play: OpenShift as the Universal Platform

Red Hat’s Kubernetes story this month centers on OpenShift Virtualization, which the company is positioning as the strategic platform for managing VMs, containers, and AI workloads on a unified infrastructure. With the broader virtualization industry undergoing a period of significant change and cost pressure, Red Hat is betting that organizations want to consolidate rather than fragment their infrastructure.

The company also highlighted substantial partner ecosystem growth, adding over 1,500 RHEL-certified applications as validated for OpenShift Virtualization, alongside improvements in storage, networking, backup, and disaster recovery integrations. The message is clear: Kubernetes is not just for greenfield cloud-native applications anymore — it is the consolidation layer for all enterprise compute.

What This Means for Platform Engineers

The convergence of these developments paints a clear picture of where Kubernetes is headed in mid-2026:

  • Control plane safety is improving. The Mixed Version Proxy and watch-based reconciliation in v1.36 make upgrades and cloud integrations safer and more efficient.
  • AI workloads are driving Kubernetes innovation. From llm-d to GKE hypercluster to Agent Sandbox, the platform is being fundamentally retooled for inference, agents, and reinforcement learning at massive scale.
  • Supply chain resilience requires active management. The Bitnami removal is a wake-up call: treat public image dependencies as risks, not conveniences.
  • Kubernetes is the consolidation layer. Red Hat’s OpenShift Virtualization push, alongside cloud provider AI infrastructure, shows Kubernetes becoming the universal substrate for VMs, containers, and AI.

For platform engineers and infrastructure leaders, the takeaway is straightforward: Kubernetes is not standing still. The investments made in upstream reliability, cloud-native AI orchestration, and supply chain hardening today will determine whether your platform can handle the workloads of tomorrow.

Sources

Next signal