You can also use rump kernels to make an application that runs directly on Xen, KVM, or hardware or that e.g. uses NetBSD file system code on Linux (or Cygwin) userland. Much of this still need some work to do what you want, but less than writing the drivers from scratch, much less maintining them. rumpkernel.org is starting to have packages for some applications. This doesn't require otherwise using NetBSD.
(I remembered that Haiku got network drivers from some WhatsitBSD, so I thought those are maybe there too, but it seems it was FreeBSD actually, so just some confusion of mine. So no microkernel there, it seems. Bummer.)
As to running on Xen etc, that was what I seemed to vaguely remeber as what rump kernels mean, so I was really confused and surprised to slowly find out that the NetBSD usage of the term seems to diverge quite substantially, as far as I can (or can't?) understand. Hmph.
There is a FAQ with more information at: https://github.com/rumpkernel/wiki/wiki/Info:-FAQ
I'm pretty sure that Antti created these terms, so it is an authoritative definition :).
For "b", rump kernels don't support "compiling out". Rather, the idea is to "link in" only what you need, as in "don't need file systems in that particular application? well don't link in VFS then".
I'm not sure if you're right or wrong, but maybe the rump kernel FAQ at http://wiki.rumpkernel.org/Info:-FAQ will straighten your thoughts. In the regrettable case that it doesn't, you can always suggestion additional questions.
My current idea is that I can take the current NetBSD kernel and:
a) for some (any? or only some?) hardware subsystems, ask the kernel (by some switch/config? where?) to run drivers in userspace, in "microkernel-like" way (and thus they reboot transparently in case of failure/bug/crash?), at some performance cost;
b) compile out/disable (? by some preprocessor directive?) some subsystems of the kernel, so that I get a "smaller" kernel e.g. for embedded machines which don't have video-card/GPU hardware;
and probably also combine the both above approaches together - i.e. build a reduced kernel like in (b), and then optionally run any missing subsystems/drivers like in (a).
Is that correct? Or am I wrong?