I’ve wondered if it makes sense to buy Intel along with Cerebrus in order to use Intels newest nodes while under development to fab the Cerebrus wafer level inference chips which are more tolerant of defects. Overall that seems like the cheapest way to perform inference - if you have $100B.