Most of my systems rarely go much beyond 30 days between reboots. I run update playbooks on them ~weekly so they usually reboot to load updated kernel modules and whatnot every 2-4 weeks.

I do it because “all of the above”.

If you need it fast, or have unpredictable load spikes, then cloud makes sense. If not, then it costs more over time than the on-prem investment,the end.

Almost never since getting a whole home generator.

I did pretty much that with Retroarch on a Nvidia Shield at one time. Worked great.