I'm not sure how many old xeons I could buy with 1 grand (including associated memory sticks, HDDs and racks) but I'm pretty confident it's going to assemble more than 10 servers. If we get mobos with extra sockets that pushes the core count up even more.
But then didn't I even give more optimistic numbers than what you're suggesting? Sure, old chips might draw a bit less, but not by a lot. By them being significantly cheaper than the latest and greatest Ryzen, you can get a ton more of them at the same cost of a specific ASIC, so even if they draw a tiny bit less, in total the consumption will go even higher than the 600W/$1k I suggested.
This page suggests a Xeon E5-2620 pulls over 50W in idle and it only has 6c/12t. But of course significantly cheaper than a 5950X.
https://www.servethehome.com/dual-intel-xeon-e5-2620-v1-v2-v3-compared/intel-xeon-e5-2620-v1-v2-v3-power-consumption/That's a good point. For a new algo to get widely adopted, there has to be a resource surplus that can be used. With PoW it used to be various ASIC models and electricity supplies.
You wouldn't want to utilize an already scarce resource such as GPUs, because that will cause vendor backlash (nvidia crypto mining throttling comes to mind).
Nobody has a need for old CPUs and servers so there's a surplus of them, and that's the only reason why I suggested them.
For something like this to work, there would have to be a new system design where all the parts use less power. But this drifts into ASIC-like territory, and particularly with things such as CPU dies, are hard to manufacture because foundries such as TSMC impose quotas on how much you can order (built-in scarcity => less adoption).