I was a bit confused about the read and write sorry ! I understand what you mean...
But do you think that the IO cost (of only one page) needed to handle the index writing is superior than 300ms ? Because each insert in any of these tables is that slow.
NB: between my "small" and my "big" tests the only difference is the number of table (10 and 3000), there is almost the same amount of on-disk data