Any insight into the actual methodology? I couldn't find specifics, but I would be curious what their baseline condition is.

I wonder if the baseline case is "no control optimization" or if it was based on current control best-practices. For example, one article claims it produces cooler water temperature than normal based on outside conditions. This is a best practice in good energy management through wet-bulb outdoor air temperature reset strategies without using ML. If their 40% savings was above and beyond these best practices, that's a pretty big accomplishment. If it's based on the static temperature setpoint scenario (i.e. non best practice), it's less so.

Edit: after skimming [1], it seems like their baseline condition was the naive/non-best practice approach. I'm not discounting the potential for ML, but I think a more accurate comparison should use traditional "best practice" control strategies, not a naive baseline condition. In some cases, it seems like the ML approach identified would be less advantageous than current non-ML best-practices (e.g., increasing cooling tower water by a static 3deg rather than tracking with a wet-bulb temperature offset)

[1] https://research.google/pubs/pub42542/

I read/heard somewhere that this isn't actually used in practice, but I can't find a source. Anyone at Google willing to shine some light on this?

is that the one where "AI" told them to just turn off unused cloud instances?

"In fact, the model’s first recommendation for achieving maximum energy conservation was to shut down the entire facility, which, strictly speaking, wasn’t inaccurate but wasn’t particularly helpful either."

https://sustainability.google/progress/projects/machine-lear...