When you deploy a datacenter at this scale you contract for an amount of power and you're basically obligated to use it. The trick is to get more compute out of the same amount of power.
By the way if you still have AC UPS systems that is probably your #1 problem and you definitely will not benefit from some fancy machine learning thing. Just get rid of most of your AC systems. For example you may benefit by adopting Open Rack 48V-to-point-of-load scheme that uses an in-rack DC UPS.
Total, more the company inhouse data center type affairs - big, but not your cloud or google size affairs. Then doing one AC to DC conversion instead of all that overhead per system along with the associated heat and centralising - pays for itself.
The cut-off for that level of work I'd say is if you design your own servers over just speccing from a vendor level is when you would be doing this. Upto that, it's still vendor off shelf. Though been a while and an option some vendors may now offer at lower scales these days.
By the way if you still have AC UPS systems that is probably your #1 problem and you definitely will not benefit from some fancy machine learning thing. Just get rid of most of your AC systems. For example you may benefit by adopting Open Rack 48V-to-point-of-load scheme that uses an in-rack DC UPS.
https://www.opencompute.org/files/External-2018-OCP-Summit-G...