I have used GUIDs (uniqueidentifier) as PK columns, and use them for the clustered index. Thus, newid() is not a good default for the column, since it isn't sequential, and new rows will fragment my clustered index. Instead we use newsequentialid(), but it is not guaranteed to be globally unique.

Fortunately, this database is not yet distributed, so I'm not worried about creating conflicting PK values in separated instances, but I'm curious to know what the ideal solution is.