Bit of all of the above - the power requirements are a compelling green argument (don't know if you've seen the suggestion, but people have the idea that the UK should move its datacentres to the Shetland Isles and turn them into, more or less, gigantic windfarms), but the big driver I see is cost/flexibility. At the moment, researchers, if they've got money, have to build (or use) some compromise grid rather than having the architecture which is best suited to the problem they're facing - and they have to find uses for it the whole time or have it sit idle. That's no use, really; what you want is to be able to get the resources you need as and when you need them. (To put it nerdily - you want to avoid premature optimization! It's the same story as just-in-time supply in manufacturing.)

So it's about having the most appropriate performance for the best cost at the least impact, and the only way to do that, I reckon, is economies of scale at a larger scale than the research group, university, or even the national research council. Hence the cloud...