No idea (and no 2014 to play with). Could have to do with the way he defined the indexes, which indexes he defined. There's a fair bit more that goes into index design for these tables compared to normal ones. Not enough available info (and not enough experience with Hekaton)

In my tests, inserts were around 0%-5% slower at low concurrency levels, I just don't have the hardware to test out thousands of concurrent users.