Foundational machine learning interatomic potentials (MLIPs) are poised to revolutionize atomistic modeling by enabling accurate and scalable simulations across the periodic table, including systems with arbitrarily complex compositions. This talk introduces the graph atomic cluster expansion (GRACE), a framework from which equivariant message-passing neural networks can be derived. I will demonstrate the application of GRACE in diverse materials science and chemistry simulations, highlighting its efficiency and accuracy through a comparative analysis with other prominent MLIPs. The talk will conclude with a brief discussion of extensions to the GRACE framework, including the incorporation of magnetism, charge transfer, and other atomic properties.