So the model is buy some of the worlds most expensive hardware and let it sit idle for half the time? If I want to save the same throughput I need to buy at least twice the hardware!

Load throttling is one of those ideas that seems great as long as someone else is doing it.

Ha, that's a great point. I guess if you have some latency sensitive inference workload the capacity will effectively be dynamic, but that is likely uncorrelated with local energy prices I imagine.