The CDU is inside the datacenter and strictly liquid to liquid exchange. It transfers heat from the rack block's coolant to the facility coolant. The facility then provides outdoor heat exchange for the facility coolant, which is sometimes accomplished using open-loop evaporative cooling (spraying down the cooling towers). All datacenters have some form of facility cooling, whether there's a CDU and local water cooling or not, so it's not particularly relevant.
The whole AI-water conversation is sort of tiring, since water just moves to more or less efficient parts or locations in the water cycle - I think a "total runtime energy consumption" metric would be much more useful if it were possible to accurately price in water-related externalities (ie - is a massive amount of energy spent moving water because a datacenter evaporates it? or is it no big deal?). And the whole thing really just shows how inefficient and inaccurately priced the market for water is, especially in the US where water rights, price, and the actual utility of water in a given location are often shockingly uncorrelated.
The whole AI-water conversation is sort of tiring, since water just moves to more or less efficient parts or locations in the water cycle - I think a "total runtime energy consumption" metric would be much more useful if it were possible to accurately price in water-related externalities (ie - is a massive amount of energy spent moving water because a datacenter evaporates it? or is it no big deal?). And the whole thing really just shows how inefficient and inaccurately priced the market for water is, especially in the US where water rights, price, and the actual utility of water in a given location are often shockingly uncorrelated.