The chip giant is preparing new "Rubin" and "Feynman" architectures to help companies address generative AI's skyrocketing costs and energy demands.