I think the important thing for AM is "hash density", since deploying new data centers is costly and time consuming. With a 40nm design they could potentially multiply their hashing power by an order of magnitude or more in the same space (cue to the potential need for special cooling solutions).
They don't depend on self-mining though. There's also franchising as well as sales.
You point out the price per GH/s as being similar, but is that on a per wafer basis, i.e. without the NRE's, or do you somehow extrapolate the production numbers and include them as well?
I didn't say the price is similar... Look, I don't know exact numbers of course, but what I wanted to say is that if there were, say, 20x more hashpower per die, but the die costed 5x as much, then the chip would only have 4x greater price efficiency, not 20x. And yes, the costs are greater for both fixed AND recurring costs, so I'm not extrapolating anything.
Similar ~= less than an order of magnitude. I think we both agree 28nm is not what AM needs.
They wouldn't need more space for self-mining with 40nm and I'm sure the franchisees would appreciate paying less in electricity (and it would be easier to deploy in DC's). You couldn't really scale to 1PH with the 130nm design for less than the migration to 40nm, so 40nm makes sense.
I don't know the exact costs for each node, but it strikes me as useless to go for an expensive node, capable of producing vast quantities of chips per wafer, when you will really only use a few chips (or risk killing the network). 28nm is really just a marketing spoof right now. AM needs to be ready for the future, but 28nm will probably not be needed for some time (the others will be stuck at 28nm until the next node becomes affordable at, say 2-5M$ NRE).