What New Changes in Kubernetes 1.36 Should Platform Teams Watch in the AI Era?

Kubernetes 1.36 Signals a New Direction

Kubernetes is evolving from a general‑purpose container orchestrator to a unified infrastructure layer for complex resources, workloads, and organizational environments. The release (code‑named Haru) emphasizes that AI‑era platform construction must consider not only models and compute but also holistic resource, scheduling, and governance capabilities.

1. Dynamic Resource Allocation (DRA) Brings Complex Resources into Native Governance

AI workloads require managing GPUs, DPUs, NPUs, high‑performance networking, and specialized storage alongside CPU and memory. DRA enables Kubernetes to describe and allocate these resources dynamically, allowing platform teams to answer questions such as which teams may use which accelerators, how tasks request specific resource specs, and how idle resources are reclaimed.

2. Workload‑Aware Scheduling for AI Training and Batch Processing

Traditional scheduling treats each Pod independently, but AI training, batch jobs, and large‑scale data processing often involve groups of Pods that must be scheduled together, satisfy combined resource constraints, and respect fairness across tenants. The new Workload‑Aware Scheduling feature pushes the scheduler to consider an entire workload as a unit, evaluate resource sufficiency for the whole task, and enforce multi‑tenant isolation and priority.

3. Security and Resource Governance for Multi‑Tenant Platforms

User Namespaces reach GA, mapping container‑side users to host‑side identities, reducing the risk of container escape in shared clusters. Fine‑Grained Kubelet API Authorization also reaches GA, extending security controls to the node side where Pod lifecycle, logs, exec, and port‑forward operations occur. Together with Memory QoS and in‑place vertical scaling, these features shift resource governance from per‑container request/limit to workload‑level quality of service and dynamic adjustment.

4. Control‑Plane Scalability for Large‑Scale Clusters

The enhancement of Server‑Side Sharded List and Watch addresses API Server pressure when object counts and client numbers grow. List and Watch are core mechanisms for controllers, operators, schedulers, and monitoring systems; scaling them is essential for stable, automated operation of massive AI‑focused clusters.

5. Three Takeaways for Platform Teams

Complex Resource Governance: DRA shows that GPUs, accelerators, and specialized hardware will become first‑class citizens in Kubernetes.

Complex Workload Scheduling: Features like workload‑aware scheduling and vertical scaling indicate a move from single‑Pod placement to holistic task orchestration.

Complex Organizational Governance: User Namespaces and fine‑grained Kubelet authorization reflect growing attention to multi‑tenant security, permissions, and isolation.

In the AI era, building infrastructure is not just about adding AI toolchains; it requires integrating AI workloads into a unified cloud‑native governance stack.

References: Kubernetes blog (2026) and the official 1.36 release page.