I’d guess the optimal strategy is “get that thing turned on as fast as possible and keep as much of the cluster running 24/7 as you possibly can before you get leapfrogged.”