The sensors which have a big value for AC1, don't work, because AC1 x 4 will overflow in the calculation of B3, because it is performing a 16 bit calculation there, unless you tell it otherwise.
But if the sensor happens to have small value of AC1, you don't get this problem, because it still fits in 16 bits.
You will observe that in Bosch's calculation, AC1 is 408 and Leo's example ha 408 also, but Jacob and Erhja have 8000 and 9000 there, which will overflow when multiplied by 4.