We propose a nested decomposition scheme for infinite-horizon stochastic linear programs. Our approach can be seen as a provably convergent extension of stochastic dual dynamic programming to the infinite-horizon setting: we explore a sequence of finite-horizon problems of increasing length until we can prove convergence with a given confidence level. The methodology alternates between a forward pass to explore sample paths and determine trial solutions, and a backward to generate a polyhedral approximation of the optimal value function by computing subgradients from the dual of the scenario subproblems. A computational study on a large set of randomly generated instances for two classes of problems shows that the proposed algorithm is able to effectively solve instances of moderate size to high precision, provided that the instance structure allows the construction of stationary policies with satisfactory objective function value.