Piotr Zaniewski

AI development has a cluster problem: laptop Kubernetes like kind can't host real GPUs, so engineers juggle Docker Compose, full clusters, and scripts that drift from production. We show a pattern where one laptop kubectl context spans local Docker workers and a cloud NVIDIA T4, joined by a single curl over a WireGuard tunnel, so the same workload runs locally and on cloud GPU without two separate setups. The idea is Kubernetes-native: any Linux host (KVM, GCE, bare metal) becomes a node by pasting a one-line curl that installs kubelet and connects over the tunnel, in one tenant cluster. We demo it live: laptop Docker workers, a local KVM node, and a GCE T4 in a single context. We pull llama3.2 onto the T4, deploy a service that talks to it, expose it through a built-in LoadBalancer, and pause/resume the cluster like a laptop. We close with the failure modes (tunnel reconnect after sleep, egress cost on model pulls, GPU eviction) and the GitOps that keeps the topology declarative.