Indeed, it is the argmax of the likelihood, but the likelihood is not invariant under coordinate transformations. The quantity p(x)dx is invariant, not p(x). By picking a suitable coordinate transformation you can put the MLE on any value where the likelihood is not zero.