I still don't buy the "page split" argument. It is purely physical, and actually a "red herring" at that because if you design the schema correctly so that you use the randomness of the GUIDs to minimalize the problem with insert hot spots which are implied by sequential ID values then using GUIDs results in simpler maintenance and a net INCREASE in performance.

Concventional wisdom is a wonderful thing as long as it is correct. In this case the conventional wisdom applies to poorly designed systems (ones that are predicated on sequential inserts), but falls apart for systems designed to work with truly large amounts of incoming data. If you are willing to trade some disk for a thousand fold increase in INSERT performance, the pseudo-randomness of GUIDs are a blessing.

-PatP