Looks like he’s doing 1M INSERTs per transaction.

In his 2nd example he’s doing the same thing, except he first sorts the data for each of those 1M INSERTs, leading to a 2-3x speedup.