I think you're both over- and under-thinking this.

First: if you go with a cloud provider, your provider is your hypervisor. If they don't directly and clearly offer the features (import, export, Windows, etc.) that you need, they're not the right choice.

On the subject of exporting images from the cloud to use locally, eh, that's not the most common feature. It does exist, but I've never seen a provider that does it in a way that's not a complete pain in the ass and/or is simple to use one you have the image, which sounds like something you might actually want. Someone giving you a qcow image of your disk does not necessarily make importing it and getting it back up on a local KVM system easy, for example.

It sounds like half of what you need is a decent laptop (Windows in the cloud kinda sucks for any application that's got any sort of GPU-accelerated component, still.) and the other half is just a Linux server somewhere running docker, be it VM, real hardware, or some box sitting in your closet at home.

(Edit: clarification: you can get a Windows VM that has a proper GPU passed thru to it, but they're expensive and if you don't get a GPU, then you just have basically software rendering, which limits what kind of software you can reasonably use with reasonable performance.)

Nothing you mentioned about your "future home server environment" can't be run on a VM, but I would comment that you're looking at a pretty expensive route if you go the VM path. Nextcloud, for example, will want ~3-4gb of ram to not utterly suck and be useless. A full *arr stack with something like qbittorrent can happily eat another 4-5gb. Add in Plex/Jellyfin? That's another 2-4gb there, too. And so on.

You might actually end up spending less on a dedicated box than you would on a VM, simply because the pricing for things like RAM are usually substantially better for real, actual, physical hardware once you get past like, 4 or 8gb.

I'd also recommend going with a containerized install for all this stuff: I bounce stuff around between servers and it's just a matter of lifting the data out of the bind mount and moving it to it's new home, and then a docker compose up and we're done.

TLDR: Cloud maybe fine, bare metal maybe fine: it comes down to budget and comfort with what you're willing to deal with.