I think you sort of answer your own question when you wonder what the value of having tons of smaller instances versus a smaller number of larger ones as you discuss the stress of federation from the larger instances soon after.

I'm not sure I have. I was saying you must have/will end up with few large instances PERIOD. There is just no way around it. Otherwise you have a discovery/usability problem. Absolutely 0 non-techies will setup their own Lemmy instance. It's not even reasonable to assume they could. All I do for work is setup cloud deployments of applications and build tools to maintain and monitor them, and I have 0 interest in setting up a Lemmy or Mastodon instance. I like Mastodon in theory, if the main instance maintainers shit the bed, create an account on another instance. But I'm not setting up my own instance.

For example, I'm into woodworking, selfhosting, programming, and cooking. I'm not setting an instance for any of that. I just want to know what are the popular instances that have woodworking, programming, and cooking communities and just subscribe to them. I'm not gonna go looking for hundreds of small instances all over the internet, nor am I gonna host my own.

As for moderation, everything does not federate everywhere at once, there needs to be some interaction between the instances first (like subscriptions to communities) for things to really start transferring over. Yes, moderation will still take work, but you don’t have to address the entire fediverse at once.

Yes, that makes sense if the problematic aspects of another instance is just their content or culture. Like if there is a say ultra-religious instance, or a toxic incel instance, or a Nazi instance, etc. Sure, you never federate with them and you never have to worry about their culture.

This doesn't work for spammers/malicious actors instances though. As a spammer or malicious actor, all I have to do is setup dozens and dozens of random instances that are just constantly spamming every instance that's on a list somewhere. Forcing it to federate with my spam instances and putting load on them. As a bad actor, I actively come to you, not the other way around.

Then you're back to the allowlist, which breaks the "federation" idea in general or introduce a lot more mod/support cost needing to manually allow each instance that you "think" is not spam/malicious.

So finally in that a world of few inevitable large instances, and hundreds of thousands of small ones that are all polling/pushing activity to the big ones, that's just significantly more load on the system as a whole. There is no magic around cache, latency, inconsistency, slowness, etc except to be on the instance where you want to converse with people.

I tried googling for how ActivityPub scales to see if there is some write up and this came up for example https://lucumr.pocoo.org/2022/11/14/scaling-mastodon/

from that article:

One thing seems relatively certain: if Mastodon wants to host a sizable community where some people have followers from most other instances, then the size of an individual instance will matter a lot and I'm pretty sure that the only sensible approach will be to either not permit small instances to participate at all, or for those to come with some other restrictions that will require special handling.

Many developers don't want to accept the problem of back-pressure. (A topic I wrote about quite a bit incidentally). Unfortunately some bad servers can really break you, and you will have to avoid federating to them. In general too many small servers will likely cause issues for very popular accounts on popular servers.