Especially when all major labs are building their custom inference chips.