A Polite, Container-Native Bursting Fabric for AI Workloads on Shared Clusters

Turning a shared, multi-tenant GPU cluster into elastic capacity means fighting , YAML, and container cold starts, with no feedback loop into running code. We present a that makes cluster GPUs appear as elastic capacity . The fabric, , makes cluster pods first-class subscribers on the same NATS bus as the workstation---as ephemeral Kubernetes Jobs or persistent queue-group pools---behind a Go control plane whose probes cluster state and backs off rather than crowding tenants. Its leaf-node topology surfaces orchestration issues (JetStream vs.\ core-NATS delivery, asymmetric subject interest). Two components make bursting practical: compresses on-wire payloads, and right-sizes each pod's GPU batch. We measure it on real infrastructure---burst cold-start median 7.3\,s (n=20), warm pool 76\,ms, up to 8.4 compression speedup over a 2\,Mbps link, 94\ where each supporting layer helps and where it does not. The rule-based politeness gate defers to co-tenants under structured or slack load rather than crowding them; the control-theoretic analysis of feedback admission---including where that politeness is regime-limited---is a companion contribution. Target venue: CANOPIE-HPC / SC26. Author preprint deposited for archival and citation. Draft — pending author review.