Yes. We are at the same path. Memory is of no concern, at least if we are talking about a few objects.
I have got the advice to keep the polys down by using instances. A good example is the bridge which I talk about in my blog.
If I do the whole bridge in my modeler the result is about 1900 polys. By creating some parts in the modeler and assemble them
to a bridge in the editor the poly count is reduced to 10% (=190 polys). Looks good on the paper, but what is realy gained by
doing this. Obviously some memory is saved, but as you say, memory is of no or at least less concern these days. Now if doing
this doesn't end up in a better framerate speed, I cant see why there should be any reason to bother at all. And as all probably
knows. Assembling pieces together in the editor is possible in theory, but not i practice. So I don't really see any meaning of
this poly count hunt for such objects that we are talking about.
One thing that I CAN understand is the case where a building has many materials. Then there may be some speed gain by
dividing separating the pieces with same material into a separate gmf. The gain would be that same shader could be used
for the complete gmf file. I guess that switching shaders is a thing that might bring speed down a bit.