@Gomeler
Agreed on both points. Design matters and we won't be moving past 28nm for a while.
As for design there is something horribly broke in BFL design and I can't see them doing a significant redesign at this point. Bitfury and BFL aren't the only two with vastly different efficiencies at the same scale. Granted both of these are vaporware at this point but I find it interesting that Hashfast is reporting ~0.6W/GH (at the chip) and KNC is reporting >2W/GH (at this chip) and both of them are on a 28nm process. We will see when it gets to real silicon but if true that is significant.
As for next gen. We won't be seeing anything smaller than 28nm. 22nm is cost prohibitive and you are looking at huge NRE. By the time it becomes possible the market will be flooded. I do think there will be incremental improvements on the same process (the "tock" in Intel tick tock strategy). There is also a troubling trend which Nvidia highlighted a while back in that the cost per transistor isn't going down much even when processes become mature so the smart money may be on optimization.
The 110/130nm tech is going to be obsolete very soon. I don't mean people won't be able to mine with them I just mean nobody is going to be able to sell new units. If AsicMiner and Avalon are going to stay in the game they will be forced to make a move to 55nm if not 28nm. When everyone is on the same process it is going to come down to design, system optimization, larger runs to reduce incremental cost, and having a solid supply chain. It will be interesting. I don't think every company that is around today will still be around a year from now.
I agree, it sounds like BFL there is something wrong with BFL's chip. Perhaps it has a chunk of the chip that is unusable but they can't disable. We don't know the finer details but it might not be out of the question that they are taking their 65nm design, correcting the issues and shrinking it to 28nm. I'm not completely familiar with the process but I suspect this would involve less engineering effort especially if they aren't hand-placing the design.
I'm not so sure there will be that many incremental changes made to chips once they start to hit the process limit. I say this as I expect the margins on chips to become so slim that it would be financial ruin to invest in the NRE for another 28nm chip once you already have one. Unless of course we discover a major optimization to SHA256 which would yield sufficient sales to recover the NRE and turn a profit. I fully expect we'll hit a lower-bound on energy efficiency for a given process (let's say 0.25W/GH for sake of argument) and then it'll come down to the GH/wafer as everybody competes on maximizing profit/wafer and GH/USD.