In the coming years we will see a lot more applications powered by deep learning. If someone releases a chip that provides 10x performance compared to the current GPU-based method, they will sell a lot of chips.
No doubt -- the problem is the assumption that 10x performance is possible. Though if anyone can deliver it will (imo) be Intel: it will be a process war, and they seem to be able to do more transistors than anyone else.
Intel may have to find a different niche. Their Xeon Phi approach to parallelism is very interesting, speaking as someone dabbling in AGI, but is not comparably well suited for deep learning.
The answer to the question "How well suited are FPGAs for -insert field where GPUs or vanilla processors do no excel at-?" is always "Far better than the GPUs or proccessors but with an abysmal power usage". In this kind of new developments FPGAs are usually used as a test before moving to ASICs.