AI inference chips are diverging from training chips, and the watershed is not compute power but the physics of memory bandwidth. Two AI ...