is this inference cost optimization?