The issue is that they had multiple threads generating objects and they wanted to use unique time-based UUID's because they wanted to efficiently fill a btree-indexed database. So if the Most significant bytes are constant, then all of the new objects will land in the same leaf node, and you can use certain hueristics such as header compression and assuming that you should fill a leaf node before adding a new node (instead of splitting the node equally) because you know that you are inserting keys into the b-tree in monotonically increasing sort order.

If you use random number generator to generate random UUID's, then you will be inserting the objects all over the b-tree and once the tree is too big to fit into memory, performance goes down the toilet.