We identify a smoothness-generalization dilemma in message passing that limits GNN universality across varying homophily and propose the Inceptive Graph Neural Network (IGNN), a universal framework to address the dilemma.