The theory of mconcat is that it should handle monoids that need to be summed in some special way. I don't know if anyone actually uses it so, however. Still, Reid is right that the circular dependency sets a very high bar.
I guess the [a] monoid is a good example where using mconcat can make a difference.
What seems to be an omission is that Dual has no implementation of mconcat. It'd make sense to define 'mconcat = mconcat . reverse' - if the original monoid benefits from a certain order of operations, we should keep the order.