There’s a config option that lets you scale down the texture automatically. For example, passing it 0.5 will cause all 1024x1024 textures to be loaded as 512x512. While this might not be desirable for ALL textures (thinking, UI/fonts), it would be a quick way to reduce memory load and test that.
Notice that analyze is probably an estimate. It knows that an A by B texture at C bitdepth will take up ABC memory, but that doesn’t account for pixel packing modes of your video card. So, textures might actually take up more memory, when converted into the card’s native format.
Another trick I’ve seen for reducing GPU memory load is to only store the X/Y offsets for a tangent space normalmap and then compute Z at runtime. That gives you an extra channel in your normalmap that you can use for other things, like gloss or transparency. (The Moho engine for Supreme Commander uses this trick – Nice way to save some texture RAM when you’ve got hundreds, possibly thousands, of normal-mapped objects running around ingame.)
Edit: I said ‘two extra channels’ but I meant one. 