The most effective way is to be able to dynamically distribute load across a large array of servers, turning some off at times of lower load. Anyone who runs a large datacenter probably already does this or something equivalent, either to save money or to service more customers with the same amount of hardware. It's more effective to optimize the hardware architecture, but it's worth noting that improvements in efficiency usually don't translate to reductions in power consumption. Instead, it causes induced demand and usage increases to fill available capacity.