For a simple example: 1234 + 5678 + 9012. You get: 15 thousands, 8 hundreds, 11 tens, 14 ones. Now adjust for overflow (small to big): 4 ones (1 overflow moved), 2 tens (1 overflow moved), 9 hundreds, 15 thousands. Final calculation: 15,924. Notably the final 'adjustment' phase does not need to be done in 300ms, so all he's demonstrating is being able to repeatedly add 0-9 to 4 small numbers in 300ms. That's certainly an achievement and one that would require a lot of training, but nothing beyond that.
You can also see this in the video by his timing. The numbers start at 22 seconds and he finishes at 59 seconds. So he spent 30 seconds on the numbers (100 numbers at 0.3 seconds), and then around 6 seconds to input his answer.