AI developers want a GPU the way they want a kind cluster: kubectl apply and it's there, kubectl delete and it's gone, no console in between. The cloud default punishes that: provisioning latency, sticky nodes, per-tenant security setup. We give developers a kubectl context with an auto-provisioning GPU under it: a T4 or L4 in ~3 minutes, gone in ~30 seconds, zero infra to touch.

Live demo: a dev creates a GPU-template tenant cluster, applies a pod, and a node auto-provisions in GCP. The pod runs, is deleted, the node disappears 30s later. Under the hood is a four-piece contract: tenant Kubernetes for isolated RBAC and node pools; Karpenter for auto-provisioning; an idle policy (consolidateAfter 30s) collapses nodes on idle, not on cluster delete; and a tunnel keeps the experience kubectl-native, not console-native.

We close with the failure modes (Karpenter template/provider drift, NodeProvider creds silently disabled) and the GitOps boundary that keeps multi-team setups intact.