Recently I was asked to estimate the power consumption of a rack we have to decide if we could put more machines in or not. After some time poking around I got to the conclusion that our 26 servers in that rack suck an avg of ~5KW. It's quite a lot, I wasn't really expecting such number, hopefully the measured one will be lower, but that's what I got from some math and paper work.

Theoretically speaking the maximum nominal power draw can be calculated as PSU * efficiency coefficient, where latter is generally around 80%. In my case all boxes are shipped with a 450W PSU, which based on the above would mean 360W power consumption per box. Taking a look at the various components' specifications can be useful to confirm the above number, being the PSU's wattage purely a nominal value. I also found this website that helps you calculating what sort of PSU you need, and by doing so offers a nice panoramic on standard hw's power draw. In my case:

AMD Opteron	85W
ATI Radeon	30W
4 sticks DDR/DDR2	40W/30W
2 SATA 15kRPM HD	80W (40W each)
4 PCI cards	20W (5W each)
Various fans	5W

That sums up to 260W, plus you want to account for the motherboard and other bits an pieces. All together is still less than the 360W estimated above, probably closer to a real measurement. It should also kept in mind that all these numbers indicate nominal powers, and during the normal operations of a system the power draw is most likely inferior (it's also true the opposite when the box is switched on, surges have to be expected).
So back to our rack, considering all the above and that some machines are bigger, assuming an average power draw of 300W under normal condition, 26 servers consume about 7.8KW.
There are some racks, like ours, that employ an AC-DC converter, which are supposed to conserve energy and be more efficient (you need a PSU that can take DC, tho). Specs for those racks claim a 30% saving, so our 7.8KW are now down to ~5.5KW.
On top of this I need to add 4 switches, the hp procurve we got consume 30W each, so another 120W.
In the following week I'll pay a visit to the ISP and take some real measurements with a clamp meter, I'm really curious to see how big is the difference with the theoretical calculation I just went through.