Any card coming out within 6 months won't be using 65nm, and thus one can conclude the 80nm manufacturing process will be used.
Current cards are using 90nm, going to 80nm will gain them +25% to +30% in transistor count, (As in 90^2 divided by 80^2), and a little in clock speed, say another 10% or so.
If the 7900 GTX got (+(1.25 x 1.1)) 38% faster would anyone really care ?, I think not (as it is under a doubling of the performance which is what people want).
So it makes sense to wait (unless your current setup is too slow) for 65nm GPU/VPUs around March 2007. (As 90nm^2 divided by 65nm^2 = at least +91% increase, or twice the number of transistors as the current 7900 GTX), and then another +33% in clock speed, for a total performance leap of around 2.54 times as power than the existing 7900 GTX.
Chances are they'll add another layer to the GPU / VPU construction when moving to 65nm, so it will be more like a +105% (2.05 times) the number of transistors when compared to the existing 7900 GTX, and thus provide 2.75 times the performance in the final version (eg: Likely called the GeForce 8800 to 8950 GTX with Direct X 10 support in March 2007).
The only alternative if die shrinks are not available is to make 'larger' GPUs, with lower yields, which would give the parts a very high price...
or designing a new interconnect on the video cards themselves and running two (or more) GPUs with low power usage on one card (and I am not talking about SLI or Crossfire on one card which is just a basic hack), this would be far more efficient than existing SLI / Crossfire solutions.
Have a look at:
http://www.es.com for some ideas.