That is why we are looking at pre-orders.  A large portion of the initial 6 PH/s in 2013 has ALREADY been paid.  Now maybe miners overpaid and will never see a return but that hashing power will come online regardless because it is a sunk cost.   Take Avalon chips as an example.  There is a very good chance those that bought them will never see a positive return because they aren't online yet however once manufactured they are a sunk cost.  Not mining just means a -100% ROI.  Mining means something better.  They may still lose but a -20% ROI is better than -100% so they will be used ... by somebody.

Now I agree with some of the numbers it is tough to estimate how many sales have occured.  Cointerra indicated they are selling 2 PH/s.  How much of that 2 PH/s have they sold.  The good news is most of the "may not be able to sell" are further out.  I would consider the 2013 numbers to be of a higher confidence.  I would also point out that the amount of capital depends on the selling price.  The cost per GH (excluding NRE) of 28nm silicon is probably less than $0.25.  An entire system is probably about $1 (people forget how much power supplies including DC to DC PSU cost).

So if we assume 6 PH/s has already been paid the question becomes how much more capacity could be sold at prices declining to $2 per GH/s.  

Simple version:
Price reflects supply and demand not cost.  If people are buying $20 per GH/s and you are selling out why would you lower the price.  You lower the price when it is higher than your marginal cost of production AND you are no longer able to sustain sales at the current price.  I think we can get to 100 PH/s eventually just by cutting prices however that is harder to project.