If you have water for the racks:
we've all gotta have water for the chillers. (compressors pull too much power, gotta use cooling towers outside.)
http://www.knuerr.com/web/en/index_e.html?products/miracel/cooltherm/coolthe...
i love knuerr's stuff. and with mainframes or blade servers or any other specialized equipment that has to come all the way down when it's maintained, it's a fine solution. but if you need a tech to work on the rack for an hour, because the rack is full of general purpose 1U's, and you can't do it because you can't leave the door open that long, then internal heat exchangers are the wrong solution. knuerr also makes what they call a "CPU cooler" which adds a top-to-bottom liquid manifold system for cold and return water, and offers connections to multiple devices in the rack. by collecting the heat directly through paste and aluminum and liquid, and not depending on moving-air, huge efficiency gains are possible. and you can dispatch a tech for hours on end without having to power off anything in the rack except whatever's being serviced. note that by "CPU" they mean "rackmount server" in nanog terminology. CPU's are not the only source of heat, by a long shot. knuerr's stuff is expensive and there's no standard for it so you need knuerr-compatible servers so far. i envision a stage in the development of 19-inch rack mount stuff, where in addition to console (serial for me, KVM for everybody else), power, ethernet, and IPMI or ILO or whatever, there are two new standard connectors on the back of every server, and we've all got boxes of standard pigtails to connect them to the rack. one will be cold water, the other will be return water. note that when i rang this bell at MFN in 2001, there was no standard nor any hope of a standard. today there's still no standard but there IS hope for one.
(there are other vendors too, of course)
somehow we've got standards for power, ethernet, serial, and KVM. we need a standard for cold and return water. then server vendors can use conduction and direct transfer rather than forced air and convection. between all the fans in the boxes and all the motors in the chillers and condensers and compressors, we probably cause 60% of datacenter related carbon for cooling. with just cooling towers and pumps it ought to be more like 15%. maybe google will decide that a 50% savings on their power bill (or 50% more computes per hydroelectric dam) is worth sinking some leverage into this.
that's just creepy. safe, i'm sure, but i must be old, because it's creepy.