I'm surprised that BFL didn't get better efficiency than they did. Just going from 90nm to 65nm should double efficiency, right? Then where's the additional advantage of using the full custom approach? I'm comparing with the current power estimate of the bASIC.
If you think the 54 GH/s ASIC is going to pull only 100w, I have a bridge to sell in Brooklyn as well. The 27 GH/s units might be around 100w, though.