Kubernetes 1.34 dropped Dynamic Resource Allocation (DRA) - think persistent volumes, but for GPUs and custom hardware. Vendors can now plug in drivers and schedulers for their devices, and workloads can pick exactly what they need.
Coming in 1.35: a new workload abstraction that speaks the language of multinode jobs. It adds topological awareness and atomic scheduling, making big, messy AI workloads play nice with the cluster.
