“An optimal AI model architecture should adapt its connectivity to the physical hardware topology, with dense connections within a chip and increasingly sparse connections across larger physical distances.”