We prove that neural collapse is approximately optimal in deep regularized ResNets and transformers end-to-end.