Adding to David's quote of $65,415 per hour here's a quick analysis of non-EC2 attacks (i.e. buy and run ALL the hardware to sustain a 51% attack)
Since power efficiency will be HUGE for this scale of an operation, let's use the most power efficient card we can find. This lovely pastebin doc tells us that is the Radeon 5850 at 1.595 MH/W.
At present you need 12.42 TH/s to pull off a 51% attack, which would require 51,585 of these cards. Assuming we put four cards to a rig it also requires 12,897 computers to run them all. Assuming the PCs take a scant 200 watts, each rig requires 804 watts of power, 10,365,735 watts in all. Assuming a fairly high efficiency cooling system, we can bump that up to about 14,619,916 watts. Before anyone claims that this alone is impossible, keep in mind it's about 0.7% of the peak output of Hoover Dam. Speaking of Hoover Dam, I'm going to use my local energy rate of 11.28 cents per kW/h (I live in Vegas) to find that it would cost $39,579.04 to run and cool these rigs for a single day.
Since these cards cost $192 at the cheapest outlet I could locate, we'd spend $9,904,320 on cards and at if we spent a scant $500 on each rig's other hardware they'd cost us $6,448,120 as well for a total hardware cost of $16,352,440. If we amortize that cost over a year of operation before the parts die or become too obsolete to pull off our attacks any more, we can estimate $44,770.54 per day in hardware costs.
Total cost in amortized hardware and electricity to run a 51% attack for one day on the bitcoin network: $84,349.58 - but also recall that our total startup costs were 16.35 million dollars, and all that would buy you is the ability to double-spend for a time, so no it wouldn't really be worth it.