Why do you need huge economies of scale though? For 90%? of queries a local model will be good enough?