It doesnt just produce on the leading edge and retool, it sells the older capacity to many many other firms for years years later. Sure, Intel may produce chipsets at .13 or something, but TSMC is probably still cranking stuff out at .35.
Geeforce,
I think Voltron's remark is critical: you'd be amazed how many 180nm chips are still being produced at mass volume, with no end in sight, so depreciation can be spread over a much longer time.
AMD doesn't have that luxury. Your leasing example breaks down in that case: Nvidia can lease for a short time the cutting edge technology, and then move on to the next. TSMC can depreciate equipment over a longer period and lower initial prices to keep its customers competitive. Meanwhile AMD has to buy the fab, use it for a short time, and upgrade after that. It doesn't have products that it can milk for years to come.
Also, it may well be that TSMC customers can't request highly tuned custom processes, but the GPU business is so much more diverse and fast moving that the vast majority of logic remains in standard cell anyway. Once you're in that situation, fab tuning really won't help that much (with standard cell, wire delays are a much larger part of overall speed than with custom.) Given the amount of manpower required, it's very unlikely that AMD will transition to full custom for their GPUs.
With new fabs costing around $2B or more, I think it would be lunacy to build your own unless your revenues surpass that by an order of magnitude. At the end of the day, former ATI and Nvidia are still small players in the world of semiconductors.