So at the risk of quoting everyone's favourite rumour monger, has anyone seen that article about process costs from the Common Platform conference?
http://semiaccurate.com/2012/03/19/global-foundries-shares-the-cost-of-doing-business/
Seems the costs of 32nm/28nm are a lot higher and the whole cost thing does seem to be spiraling out of control. (BTW, I like the comment asking: "Now, I wonder what Intels costs are?" and wondered the same thing myself after all even Intel's pockets are not bottomless...)
While I'd love to either AMD or Nvidia to give a great bargain 28nm GPU, with yields poor and costs high I don't see that happening any time soon.
It's not spiraling out of control, based on those numbers.
If it was spiraling out of control, the increase in costs would be higher than the benefits from the improvement.
Going from 45->22nm (~4x as many transistors per mm^2), costs less than 2x as much for fab costs and process development cost, which is what GF/TSMC etc pay for.
The cost increase for chip development is more of an increase, 2.5x, but you're still getting 4x the transistor density.
The BENEFIT of this is that older processes still remain somewhat competitive, since they will be refined to the end of life, and customers may feel "good enough" is acceptable, while the bleeding edge players will still require the bleeding edge.
Of course, it requires some consolidation to enable players to have the capital budgets for the initial expenditures required, and it won't necessarily lead to the dramatic price drops we've seen historically, but for the actual fabs it's not exactly all negative.