# GKE dynamic slicing for TPUs is now generally available

published: 2026-11-09T13:05:38.952Z
topic: infrastructure
source: https://cloud.google.com/blog/products/containers-kubernetes/gke-dynamic-slicing-on-tpus/
url: https://aiagentnews.io/stories/gke-dynamic-slicing-for-tpus-is-now-generally-available

Google Cloud has made dynamic slicing generally available on Google Kubernetes Engine for TPU workloads. The feature forms TPU slices on demand instead of relying on fixed nodepools, which Google says cuts job startup time by up to 5x and workload recovery time by up to 4.5x.

Dynamic slicing separates logical TPU slice formation from physical nodepool provisioning. Previously, each slice topology needed its own fixed nodepool, so changing workload needs meant deleting and recreating pools, and a single hardware fault could take down an entire slice. Google says the feature reduces startup times by up to 5x and recovery by 4.5x.

The feature comes in two forms. Super-slicing uses TPU optical circuit switching to join 16-VM cubes into larger topologies when a workload is scheduled, while sub-slicing partitions one cube into smaller independent slices that can run different jobs at the same time. When a job finishes, its slice is torn down and the partitions return to the pool.

Setup begins by pre-provisioning TPU capacity into 16-VM nodepools using a new incremental provisioning type, which can start with healthy hosts even when some are faulty. Workloads request dynamic slices through annotations, and the slice controller creates a Slice custom resource to form them. The feature is offered in all regions where TPUs are available, but only under TPU All Capacity reservation mode.

## Why it matters if you build agents

Teams running agent workloads on TPUs can request the topology a job needs without keeping a fixed nodepool for each shape. Inference workloads can share a cube across smaller jobs, and partially healthy hardware can still be used instead of losing a whole slice. Check that your reservations use TPU All Capacity mode, since the feature is not supported under other modes.