Nothing is truly zero overhead. C++ templates are neither, because they create some overhead in code size which is obviously non-free (both space-wise and time-wise). Zero-overhead for me = optimized to the point that noone can notice. There is no problem for a VM to generate 10 versions of a tight loop, every one optimised for particular type, with a switch for the type right before entering the loop, or the whole method, or even a whole call-tree.