Scott35: You might want to take a look at the following waveform. Scroll down to figure 8 on the following web page:
http://www.zlan.com/waveforms.htm This is actually a very severe level of harmonics, and rather typical of what I have seen a couple times I've put a scope on a switching mode power supply.. Notice that each spike is conducting for about 1/3 of the time, and 2/3 is spent not-conducting. Next, imagine superimposing the same waveform over itself for 120 degrees phase shift and 240 degrees phase shift. The spikes essentially won't overlap. And don't forget that these spikes are actually about 3 times higher in current than the peaks of a normal linear load with the same average current.
Of course if you have a mix of loads, and computers are just a small portion of those loads, then you're not really going to see much of an issue with triplens on your neutral, as long as the various overcapacity requirements in the code are followed.
However, in the case of a room full of computers (and I've worked in rooms of as many as 2500 computers), you end up with essentially no loads other than these very spiky switch mode loads. If enough computers are pushing the average current load to 12 amps on a 15 amp rated circuit, then you're going to see 36 amps average on the neutral. Those "double neutral" cable assemblies, panels, and transformers are really going to get more than just double the load.
I read a study from Square-D where an analysis suggested that in the typical office building, triplen currents would not exceed double, and rarely even reach it. But it appears that was a study of the overall office environment or building, and it appeared to have been done many years ago. In the age of the internet and a web based business presence, there will be more and more computer rooms running multiple servers and requiring dedicated circuits. A 4-wire 208Y/120 branch circuit to that room is simply going to be unsafe (36 amps on a #14 neutral). If you up the cable to have a #10 neutral, you can probably get away with it, though I think #8 should be required on the neutral for this example case.
Among the many things I do is specify data center designs for internet providers or business going online in a big way. I always include in my design a specification that all 120 volt circuits must have individual neutral conductors regardless of the system phasing. No shared neutrals should ever be allowed for this kind of load (even for single phase systems, but that involves some other more complex issues).
I've also considered connecting computers L-L at 208 volts. Most can handle it just fine. But this shifts the problem around a bit. What happens is you get 2 sets of spike currents on each line conductor instead of 1 in the L-N case. So that case of 3 sets of 12 amp average switch mode loads is going to be putting as much as 200% load on each line, instead of 173% as would be expected in the case of 3 sets of linear loads (so 15.47% more). Those 3 sets of 12 amp loads are now going to be 3 sets of 6.93 amp loads at 208 volts, so that means 13.86 amps on each line conductor (not 12). That's too much for the 80% margin of a 15 amp circuit. Up it to a 20 amp circuit (or spread the loads among more circuits) and this might work.
A good computer room design will involve UPSes. I've read that some UPSes (of the double conversion always online tyoe) will have less harmonic current than a regular switch mode power supply. That might be a way to avoid or mitigate most of these issues. But there is always the risk that someone could bypass the UPS in the event of failures (and it seems some UPSes do fail).
If the computer loads are a major portion of the whole building load (e.g. a building just for a large data center), then this issue has to be dealt with upstream, as well, likely even involving the POCO.