We present a principled recipe for building graph foundation models that generalize across arbitrary graphs, features, and label spaces.