I am sometimes wondering wouldn't the most economic thing to be to geographically distribute AI computing. Latencies can be large, but power could be cheaper if the loads are just moved around the globe as needed. This would mean that hardware sits powered off or in very low power state for larger chunk of day, but maybe that is acceptable in some scenario.