Regarding mining, I don't think people are very cost-efficient in how they setup their mining rigs. There's too much "I want to get the best there is" mentality floating around which increases CAPEX cost significantly (for less OPEX of course - but IMO it's not worth it unless the electric costs are too high). The fact that people are buying brand new 280s/290s and pay thousands of dollars from the get go is a bit disturbing.
Clocked 5850s and 6870s are pulling like 1.3 MHs* and people can find used ones at around 100$ per card instead of 500-600$ for a brand new 290. Will a 290 give 5-6x performance? Nope.
* For some reason, series 58xx and 68xx seem to have a factor of 3.6-3.8x compared to Scrypt.
Not sure how you're getting 1.3mhs on a 5850, mine start being unstable and risk crashing the comp when trying to push over 1mhs, but if you want to ignore electric costs and look at just hardware costs, you also need to inclue motherboards, ram, etc, since you'll need twice as many pcie slots to get the same hash power.
Regardless, electric costs are not negligible when you run them 24/7 for moths at a time, and although I agree with you about the 280x's and 290s not being worth the premium, I wouldn't buy any 5xxx or 6xxx cards at all at this point, as there is a significant efficiency step up to gcn, so 7xxx is the minimum.
My preference is up to dual slot mobos with celerons and cheap PSUs - like 20-25 euros at most - where you can make a system with like less than 150$.
If one opts for 4x cards like 250-300w each, then one requires very expensive mobos + very expensive PSUs (which could equal the cost of one further 290). Furthermore, the 5xxx and 6xxx can be resold for near the same money at a later date (very small reduction in price if bought used / resold as used) versus the 290s who may have lost something like 30% of their value due to being priced as premium / top-end products (and they won't be on that position for long).
I agree with the larger efficiency, primarily due to the 28nm energy saving gains, but the price gap is so large that one has to be mining for a loooooong time to make up for the price difference.
For example, let's say you have 2x 5850 pulling 2x200w (I consider it clocked at max point where it's stable) vs 1x290x (optimized/clocked fully) pulling 1x290w (which is the max TDP and not above it).
If we go something like 0.6x for DRK power draw it goes down to 240w for the 2x5850 and 175w for the 290x.
After a year running 24/7 with the same hashrates, the costs will be:
2x5850: 0.24kw/hr x 24 hours x 365 days = 2.102 kwh x 0.2$ per kw = 420$ for the power + 200$ for the cards = 620$ total.
1x290x: 0.175w/hr x 24 hours x 365 days = 1.533 kwh x 0.2$ per kw = 306$ for the power +500-600$ for the card = 806-906$ total.
If we went for scrypt it would be like
2x5850: 0.4kw/hr x 24 hours x 365 days = 3.504 kwh x 0.2$ per kw = 700$ for the power + 200$ for the cards = 900$ total.
1x290x: 0.29kw/hr x 24 hours x 365 days = 2.540 kwh x 0.2$ per kw = 508$ for the power +500-600$ for the card = 1008-1108$ total.
And, if one tries to recover some of the costs, by that time I'd guess that there'd be significant price loss if one tried to resell the 290x vs the 2x5850s.
The premium pricing in 290s is inflating the CAPEX tremendously and the improved OPEX is not THAT better to make up for it over a reasonable amount of time. That's how I see it at least.
