MTWSD :
Basically, Scrypt ASIC miners might not be economically viable.
when you consider that the miners are causing AMD to lose market share in the PC gaming arena, then it might be in their interest to actually design a card with the unnecessary stuff gutted.
Making a new set of photo masks to make an R9-290X re-spin specifically for ASIC mining would not be cost-effective for AMD: the R&D, validation and photo masks would cost a few millions up-front, adding over $300 per chip in costs if they only sold ~10k of them.
Most of the cost of making chips (aside from the R&D that goes into the logic design itself) is die size and the die size is ultimately constrained by IO pin count and design size. At 512bits, you are looking at ~700 memory control and data pins and 1500+ power+ground pins for over 2200 total pins, which would be a 49x49 or larger grid (need to leave a few open spots in the grid to make fan-out through the substrate manageable) and with a 0.4mm ball pitch (~60 balls per inch), you would still end up with a die in the neighborhood of 390sqmm just because there are that many balls that need to fit under it, which is only 10% smaller than the stock R9-290X. Almost no savings to be had on the production end of things.
BTW, the main reason why CPUs have IGPs is exactly because of that: integrating the memory controller in the CPU considerably increased the die area required to fit all those extra balls so CPU manufacturers added IGPs and larger caches to do something with that otherwise wasted space.
So, if AMD made mining-specific chips, they would most likely end up significantly more expensive than their regular GPUs due to the unnecessary extra R&D cost that does saves next to nothing on production costs.
BTW, the HD7990 was not a 768bits chip; it was a pair of 384bits (HD7970) chips.