Most graphics cards dont need to reserve the entire memory pool, they simply reserve an overlay "windows" so the cpu / os can copy data to the GPU. That visible spot can be moved around so that windows can see all the graphics cards ram, but without wasting the memory address space of the computer as a whole.
Thats why you can SLI a pair of 1GB graphics cards and still have more than 2GB available ram for windows.
Before I switched to 64bit, I had 3.5GB available to windows with a 768mb 8800GTX.
Look how it used to work with AGP, you could specifiy how big the windows was in bios, and it was just fine to have a 64mb window with a 256mb card. The only problem these days is although PCI Express still uses a "window" into the GPU's ram, but its no longer possible to manually configure the window size, so its all down to the design of the GPU and drivers.