We (Saturn Cloud) published a write-up on the architecture and setup process for running our platform on Nebius. The goal was to make it straightforward for teams to use GPU-backed environments without having to spend much time dealing with Kubernetes operations.
The article walks through how resources are provisioned, how environments are created, how GPU jobs are scheduled, and what abstractions we use to keep the system flexible while hiding most of the complexity of the underlying cluster. It also includes some of the design decisions we made along the way and a few of the tradeoffs we ran into.
Since we built the system, I’m happy to answer questions about the architecture, decisions, limitations, or areas we are still iterating on.