The p-median problem is a classic discrete location problem with several applications. It aims to open p sites while minimizing the sum of the distances of each client to its nearest open site. We study a Benders decomposition of the most efficient formulation in the literature. We prove that the Benders cuts can be separated by a polynomial time algorithm. The Benders decomposition also leads to a new compact formulation for the p-median problem. We implement a branch-and-Benders-cut approach that outperforms by an order of magnitude state-of-the-art methods on benchmark instances.