Yeah, that's the bit I learned a week after doing the calculations. When people first started, they were using CPUs. When GPU usage became available, CPU-miners became so horribly inefficient that electricity cost alone killed them.
Now, granted, I can't imagine anything being that much better than GPUs, but new GPUs are usually much more efficient than old ones, and they come out all the time. It could quickly became an arms race just to break even.
On top of that, apparently there's also a built-in cycle that makes it harder even if the machines don't change. The linked article touches on that briefly as part of his decision.
FPGAs and ASICs are in use by a few people/companies. They have the opposite cost structure of GPUs. The two main measures are MegaHash per $ and MegaHash per Watt, roughly analogous to fixed expenses and operating expenses though the MH/Watt also influences investment in power and cooling infrastructure. GPUs have a high MH/$, but a rather poor MH/W.
I own 2 Radeon 5830s, $110 each, that still spit out about $5 a day at current prices making for a rather impressive rate of return in spite of high power requirements. From what I have read, FPGAs require about an order of magnitude higher capital investment with the benefit of a drastically reduced power consumption. Last I checked about 6 weeks ago, a decent GPU rig could pay for itself in a month or two, while an FPGA setup could pay for itself in about a year. The tradeoff is that the GPU rig's payout may dip below power costs before it pays itself out (likely the case for any rigs purchased in the past month) essentially betting on the short term difficulty increases, while FPGAs are betting on the long term health of the bitcoin system itself.
FWIW, the newer ATI GPUs aren't as good for mining as the older 5800 series.
I was a PoliSci/CompSci/Econ mega-nerd in Uni, so Bitcoin is my fantasy come true.
