The naive interpretation of multisampling would imply that, for instance, 8x MSAA would require a framebuffer that takes 8 times the space of a non-multisampled framebuffer, for all the duplicated samples. Since the latest video cards support even 32x MSAA, that would mean that just the color buffer of a 1600x1200 output would use 1600·1200·4·32 = ~245 MB.
Is this actually the case? I mean, I realize that potential memory optimizations are likely to be implementation-dependent, but is there any information on this? Should I be extremely conscious of, for instance, allocating multisampled textures? (This is my main question.)
I'm asking in the context of OpenGL, but I don't reckon this would be different between DirectX and OpenGL.