From zero to self-service GPU: Building an internal ML compute platform on bare metal

Most platform teams do not have a GPU strategy. Data science teams submit tickets, wait for hardware, and configure environments manually, while platform engineers field requests they were never set up to handle at scale. This workshop walks through the core architecture decisions behind a self-service ML compute platform built on bare metal: API-driven provisioning, environment bootstrapping with cloud-init, Kubernetes integration for workload scheduling, and the internal portal layer that abstracts it all for end users. Attendees leave with a reference architecture they can adapt to their own stack and a clear picture of the sharp edges before they hit them in production.