aevm :
I'm guessing they've got some technical difficulties and had to go this way.
The 8800 GTX shows pretty fine what kind of problem they have. Nvidia was unable to put the actual GPU together with the IO unit on a single die. That was using 80nm i think.
The GPUs are getting more and more complicated and thus the transistor count is rising. GPUs have left CPUs in the dust regarding the transistor count. Now taking into account that a GPU is basically a big parallel processor, why not chop it up in smaller independent units.
I.e. take a 6600GT core, shrink it to 55nm and put three of them on a single die. Now manufacturing it, one of the cores is defective, gets deactivated and you still get a 6600GT SLI. Now place a neat interconnect on the same chip so you can easily link multiple of those chips in a row on a single PCB. Basically that is already happening - its just not called cores but shaders. It's a little more difficult than that, but any monkey can find the details on the net.
I think both companies are trying to get around the limitations of their manufacturing capabilities while trying to remain profitability. Neither ATI nor Nvidia could just invest a big chunk of money to get 45nm production running. If Nvidia would have had the ability to produce the 8800 GTX at a tested 45nm process at the time it was designed, they would have done that. The reality looks different though. NVida and former ATI too, are both fabless. They have their production outsourced to TSMC for example. The smallest process they can get is the one offered by those fabs.
To do it themself they would have needed to invest everything they had into a new technology that wasn't tested before, might have yielding problems with a new product that has the possibility of flopping? Dangerous course of action with a few parallels to what AMD has done with their Phenom. I'm not even sure if there is enough money in the entire GPU market to start a 45nm process from scratch or if Nvidia and ATI together could keep a 45nm fab profitable and running at full load for long enough to pay its cost.
Instead GPU manufacturers feed on the crumbs falling off the table of the big players. IBM, Intel, etc. that can keep multiple fabs running at load and have a wide palette of products they can manufacture at those fabs. Intel uses their older 90nm process for chipsets now, for example, until they switch it to a newer process.
The whole SLI/xfire is a stop-gap measure until either the companies can get their chips done on a single die OR until they have figured out a way to improve scaling like we see it with multicore CPUs today.
A neat example of the whole situation is S3, i think. They are not really competing in the GPU market anymore but they still produce chips and even though they are targeted at a different market segment, they know that they will have the same trouble as the big players. Just look at the S3 Chrome series. The newest one will be based on a 90nm and another one following later on 65 nm - a production technology that Nvidia or ATI don't even use anymore by then. Still, S3 developed their own form of SLI/xfire called multichrome.