Serve AI models on GKE with the right hardware.
Recent revisions are kept. Restoring one creates a new revision.
