Yeah, weekly backup to B2 or similar should be fine. Postponed setting it up myself - I paid for a year of Ente and currently sync everything both to premium Ente and my local Immich, so backup is fine for now.
For the heat: I am running it on a Pi4 with 8GB RAM and it runs pretty decently even without cooling. For the large import, I set up only the ML-container on a more potent desktop (still fiddling with GPU HW-acceleration) and added it as remote ML to Immich, so I can offload heavy ML load when e.g. importing many files.
The normal upload stream is fine with the local ML, though