So far the conclusions are that watercooling is not really practical outside of experimental systems (such as the K supercomputer), air cooling tends to have high running costs, and oil is a mess.
My design, inspired by heatpipe coolers aims to dissipate heat using standard heatpipes from chips to chassis, and then dissipate the chassis heat using a cold wall on both sides of the rack. The cold wall is actually just a rack-wall sized waterblock with bottom connectors for leak protection.
The result is passive cooling at chassis level (with the exception of light airflow to keep the rest of the board cool), and extremely cheap standard scale watercooling.
The hot water from the rack gets cooled by another cold wall which swims in an artificial river located just below the compute floor, with a slight slope and a single set of redundant pumps to keep the river flowing.
It may be possible to further improve the efficiency by using huge heatpipes instead of cold walls, but then that would position the “artificial river” above the racks…
So far, I’ve gotten my hands on some very nice heatpipes, but haven’t begun testing the cold wall part.