Normalization is great to be forced to use while you learn, but after you have learned, then you are free to (and should) break out of the mold when necessary and proper.

Some things are just not meant to be normalized (such as OLAP databases), and some things would just run too slowly normalized. Plus, once you factor in partitioning (if your database is just that big), then denormalization can make a lot of sense.

Often the best mix of things involves a little denormalization and storing "expensive" calculated data sets using memcached.