Shared GPU resources in scientific clouds are increasingly contended, yet static project allocations cannot share them fairly among competing tenants. We present a multi-tenant SLURM platform-as-a-service (PaaS) built on the HUN-REN Cloud, extending our prior cloud deployment with a schedulingpolicy layer and an elastic scaling path defined as code. Codified per-user quotas and quality-of-service (QoS) classes form the service contract, and worker capacity is added with a single command. We characterize the automated scale-out from measured data. Provisioning dominates the rollout cost and scales sublinearly, configuration grows only mildly, and registration is negligible. Adding sixteen nodes completes in under seven minutes. The rollout is bounded and predictable, with variance concentrated in the cloud-provisioning substrate. We quantify the configuration effort as a fixed set of idempotent tasks delivered by one command and reused unchanged for node recovery. These results characterize the deployment and maintenance of a SLURM PaaS on a national scientific cloud.