Or under clocking and under volting for even better performance to price/power/longevity ratios

For a single rack, you really don’t have too many choices for power. You make a choice to provision and pay, I never had anyone check how much of that I used and give me money back. Maybe things have changed though.

No doubt. Especially for GPU inference at scale. We overclock/overvolt for training and tune way down for inference.