According to @_avichawla, consolidating 85+ models in SIE cuts GPU costs ~75% by pooling memory and evicting LRU models across one server.