Also omitted - the amount of speculative buying for planned capacity that never actually happens. I worked for one of the big tech companies for several years, and more specialized hardware especially (ML accelerators) were spun up with the notion of "we don't know who will need these, but we don't want to not have them if they're needed". Cue massive amounts of expensive hardware sitting plugged in and idle for months as dev teams scramble to adopt their stuff to new hardware that has just enough difference in behavior and requirements as to make it hard to migrate over.

Also also, there's a bunch of "when in doubt, throw it out" - automated systems detecting hardware failure that automatically decommission it after a couple strikes. False positive signals were common, so a lot gets thrown out despite being perfectly fine.

I agree with this. Efficiency vs cooling the infrastructure and updating hardware after a maximum of 5 years. Still, I'm not 100% sure about statistics. Do you know of any comparative studies or the like?

Just one fitting side note. We had an interview with a local data centre manager and during the discussion, we somehow started talking about alternative setups, like a raspberry pi server. The interviewee reminded us of the efficiency of their virtual servers. He even gave us a tour through their digital dashboards and showcased the 1 watt used by a server (vs roughly 4 watts of a Pi, with much less performance).

This is not to say that low-tech is not the way to go. Less mining and hazardous work conditions are always good and need no measurement for emphasis.

There are actually quite a few projects to use waste heat from data centers to heat homes. With aquifere storage that can even be used in a seasonale fashion.

Did I read that you host this on renewable energy?