We construct an exact RKBS model for neural networks with arbitrary width, depth and topology and use this model to derive tight bounds on Rademacher complexity.