I do really think that a form of (don't kill me!) somewhat centralized archive of posts that is integrated with fediverse servers might help take some load off of instances. While also shifting away from the "every post needs to stay there forever" mode. Maybe your server only keeps a week or a month of posts (and assets) around and then they go to "the collective archive" (if people want their posts archived). You could even use the "login via OAuth" workflow to allow people to delete their own posts from the archive. So I think it wouldn't be massively hard to build. It would cost quite some money to run long-term but I think it's easier to collect that money as that form of organization. Also it would be a great resource for research.
I do think the distributed social web needs more infrastructures like that.
https://tldr.nettime.org/@tante/113395669754192558
last time I heard about the server load from mastodon being expensive to operate, I heard it's the act of pushing that puts load on the server and takes it down. Not the act of serving, typically. It sounds like serving the static content is relatively easy cpu-wise. But it sounds like the pushing code is still based on a one-thread-per-request model that thrashes the CPU w/ context switches and doesn't work very well unless you have a whole lot of cores.
Something like gotosocial that has proper non-blocking IO would be a lot cheaper to operate on a smaller computer, but bandwidth will always be an issue I guess.
It sounds like you are talking about making a CDN, kind of like how PeerTube can have videos "seeded" by random volunteers on the net ?
Such a thing could even be layered on top with JavaScript, without needing to modify the server
