I don't think 17 kW in a comparatively small volume is that difficult to dissipate. For one, there are no hot spots. We're talking about 4 W peak dissipated over a 27 mm x 27 mm surface as atomic unit. The chips are glued to perforated (to interconnect the layers, unfortunately the designers used 8 bit wide links instead of 1 GBit serial ones) sheet copper with some distance in between. I would think the chips would be perfectly happy to operate ~60 deg C, this should be manageable with a pretty strong airflow as the geometry is already strongly suggestive of a radiator. Alternatively, we can incorporate cooling channels into sheet copper and carry the heat away by liquid, which then can be dissipated by a large radiator elsewhere.
Alternative solution would be setting the CPU-studded copper sheets vertically into ~60 deg C boiling point fluorocarbon (ozone layer? Who needs the ozone layer!?). We would have excellent convection, and the heat would be carried away by evaporation enthalpy. It would condensate elsewhere, preferably in a large radiator which could be cooled passively. Of course we could simply use the heat to prepare hot water in the building, resulting in a slight plus instead of a big minus on the air conditioning bill.
The chiefest trouble is making an efficient power source (2.2 and 3.3 V) for that beastie. I guess the power source would dissipate at least as much as the CPU box itself. Then there is the small matter of interfacing a disk array/RAM bank to the beastie (could be handled by a Beowulf, though).