Previous work has shown that reasoning with real-time temporal logics is often simpler when restricted to models with bounded variability---where no more than v events may occur every V time units, for given v, V. When reasoning about formulas with intrinsic bounded variability, one can employ the simpler techniques that rely on bounded variability, without any loss of generality. What is then the complexity of algorithmically deciding which formulas have intrinsic bounded variability? In this paper, we study the problem with reference to Metric Temporal Logic (MTL). We prove that deciding bounded variability of MTL formulas is undecidable over dense-time models, but with a undecidability degree lower than generic dense-time MTL satisfiability. Over discrete-time models, instead, deciding MTL bounded variability has the same exponential-space complexity as satisfiability. To complement these negative results, we also briefly discuss small fragments of MTL that are more amenable to reasoning about bounded variability.