Enfabrica’s new Ethernet-based AI memory fabric system drops AI inference cost per user per token by up to 50%.