It doesn’t matter if the market is small, since it is a high end, professional tier market where profit margins are dramatically higher. You are aware of the cost of DSPs, GPUs offer a tremendously better price/performance ratio. And it doesn’t only have to be audio DSP - every intensive number crunching scenario benefits dramatically from GPU compute. Image processing, video processing, simulations, encoding, even vector graphic rasterization. And it is all professional stuff, not your typical 5 dollar toy app.
Besides, adding GPU compute facilities to JUCE is not all that big effort, just a nice C++ wrapper for the OpenCL API to create, compile and execute the kernels and maybe some integration with already existing APIs. For a guy like Jules it should take a few days. The hard part is writing the kernels, which is up to the individual developer.
Just take a look at the implementation in Qt - it is not all that much code, most of it is documentation.