By compressing the dynamic range you are necessarily compressing the range of values your data can hold. That IS data compression. Kind of stupid data compression because you cannot then recreate the original data from it. You essentially lose all of the mapping between frequencies outside of your normalization envelope and map them to some other sound that is in that place(if there is sound occupying that space). That's what all that noise is--data that has been mapped to the same frequency. Maybe it doesn't make much of a difference when there are 2-3 instruments and they don't get in each other's way when normalized, but when you have sound and fury all compressed in a small space that's when you will really hear it.

As a side note it might be worth noting that it may be a way of making bad artists sound good. In a way you make their job easier with a smaller sound space they need to master. If you haven't noticed how some artists sound good on CD and they suck live this is the reason...not to mention other neat tricks they do in the studio but I digress...