Kubernetes, cloud-native computing's engine, is getting turbocharged for AI (www.zdnet.com)

CNCF and the Kubernetes community announced a Kubernetes AI Conformance program plus a suite of runtime improvements aimed at making Kubernetes a first‑class, production-ready platform for AI. The conformance program defines shared test criteria, reference architectures and validated GPU/accelerator integrations so AI and ML workloads become portable across public cloud, private and hybrid environments — reducing vendor lock‑in and fragmentation. Major providers (e.g., Google Cloud) have already signed on, and with roughly 58% of organizations running AI on Kubernetes today the program targets predictable, secure, and interoperable deployments at scale. Complementing the standard, Kubernetes is being rearchitected with AI-specific features: safe minor‑version rollbacks and the ability to skip updates to lower upgrade risk; finer-grain hardware controls for GPUs/TPUs/custom accelerators; dynamic GPU provisioning and scheduler optimizations; and new APIs and features like Agent Sandbox and Multi‑Tier Checkpointing. Agent Sandbox offers strongly isolated, declarative sandboxes for stateful agent workloads with fast provisioning and pod snapshot resume, while Multi‑Tier Checkpointing layers local fast storage, node replication, and cloud backup for resilient, low‑latency checkpointing across distributed training. Together these changes make upgrades safer, multi‑tenant clusters more secure and efficient, and large-scale training and agentic inference far more robust and portable across vendors.