However, this all assumes realtime requirements. For batching, you can smooth over the demand curve, and you don't care about latency.