Probably has to do with rebuilding a decent working set in the database buffer pool. My guess is your buffer hit rate kept improving with each subsequent backup.