Quantum efficiency isn't the only driver. Read noise can be a major factor as well. In some cases (some sensors, some ISO settings) driving read noise to zero could provide better than a 1-stop improvement. While QE at 100% is not possible, read noise at essentially zero is possible.
Yes, I hadn't considered lowering the read noise. Anyone know how many stops of noise improvement is practically and theoretically possible there? I don't know enough to make sense of the info at http://www.sensorgen.info . I want to get a number so I can spout off and say "noise performance cannot ever improve more than X stops from what we have today" every time someone expects a 2x improvement in the next model.