FPGA is still too cumbersome to make it big. It's too expensive for general appliances, talent is hard to find, and development process is still stuck where software was 20 years ago. FPGA vendors are still trying roll out their own everything-included non-standard solutions. Those solutions don't scale well. I've seen engineers struggling to trace where some signal ends up, it's complete insanity.

I find GPUs conceptually similar to FPGAs for most soft applications (video processing and similar number crunching). They also provide huge number of re-purposable blocks for programmable parallel computing. GPUs have won out because they became mainstream through gaming and they more readily opened up to general software practices and methodologies. It's no surprise machine learning community is avoiding FPGAs for most part.