Under the production approach to markup estimation, any flexible input should recover the markup. I test this implication using manufacturing datasets from Chile, Colombia, India, Indonesia, the US, and Southern Europe, as well as store-level data from a major US retailer, and overwhelmingly reject that markups estimated using labor and materials have the same distribution. For every dataset, markups estimated using labor are negatively correlated with markups estimated using materials, exhibit greater dispersion, and have opposite time trends. I continue to find stark differences in markups estimated using energy and non-energy raw materials. Non-neutral productivity differences across firms can explain these findings.