So the thing is that as long as Nvidia tapes out a new architecture every single year, the best you can really do is beat them by a factor of two or so by matching them without the constraint of fixed function graphics transistors along for the ride.

And I think you need to beat them by a factor of 5 to 10 against currently shipping hardware (not last year's model or worse) to get traction. Also you have to deliver perf, not just perf/W. Those two constraints alone will eliminate 80% or so of these companies I suspect.