A class of network topological optimization problems is formulated as a nonlinear mixed integer programming model, which can be used to design transportation and computer communication networks subject to a budget constraint. The approach proposed for selecting an optimal network consists of separating the continuous part of the model from the discrete part by generalized Benders decomposition. One then solves a sequence of master and subproblems. The subproblems of the minimal convex cost multicommodity flow type are used to generate cutting planes for choosing potential topologies by means of the master problems. Computational techniques suited to solving the master and subproblems are suggested, and very encouraging experimental results are reported.