They're not all tooootally atomic, I have clustered things like libation+audiobookshelf, ollama+ds4+litellm, but I like that I can manage things almost entirely at the network and host level, generally atomically per service, and especially between services (eg homeassistant and the local ai).
It's DEFINITIVELY overkill, and fills up fast, but nix making everything near fully declarative to roll fwd/back or shuffle between proxmox hosts at any time makes for a clean and easy mental model. Both for me and the AI, because it can infinitely retry a nix flake config until it gets it perfect.
Sounds like a poster child for LXC containerization with Proxmox — all the host-level management of a non-ephemeral VM (including snapshots!), without the every-machine-has-its-own-kernel overhead.