The guide mentions:

Your ISP will give you the first 64 bits, and your host machine will have the last 64 bits.

This isn't correct. While some ISPs do give you the first 64 bit (a /64 prefix), this isn't recommended and not terribly common either. An ISP should give its users prefixes with less than 64 bit. Typically a residential user will get a /56 and commercial users usually get a /48. With such a prefix the user can then generate multiple /64 networks which can be used on the local network as desired.

While you certainly can run AI models that require such a beefy GPU, there are plenty of models that run fine even on a CPU-only system. So it really depends on what exactly Ollama is going to be used for.

I wouldn't run it as a router due to its high power consumption, but it would be a fine computer for retro gaming for games up until ~2005 if you add a graphics card.

Edit: 75 LXC containers, 22VMs.

That's a lot of power draw for so few VMs and containers. Any particular applications running that justify such a setup?

I found both whoBIRD and Birdnet-Pi to give good results, as long as you dismiss the low confidence results. For results with a confidence of 80 % or higher I very rarely have incorrect results. Every once in a while it confuses one kind of thrush with another, but they do sound similar to my human ears as well.