nVIDIA is aiming to release the Fermi architectures replacement called Kepler in 2011. The presentation slide at the GTC event also revealed that they are aiming to improve DP-FP GFLOPS per watt dramatically with each new architecture. Fermi is around 1.3 DP GFLOPS per Watt, they want Kepler (28nm) to be around 4-6 DP GFLOPS per Watt and the Maxwell (22nm) architecture coming in 2013 might end up around 14-16GFLOPS per Watt. Hard to accomplish ambitions, but if they succeed Maxwell could provide 12-16x times the double precision floating-point performance of the current Fermi architecture per Watt.
Image courtesy of The Tech Report.