For now, they are the low volume option (<1-10m units) if you need special interfaces. If all you are doing is number crunching a Gpgu would likely be a more effective option.
FPGAs are not a silver bullet. If you can express your design using a GPGPU instruction set, then by all means please use that. Its going to be more efficient by an order of magnitude, from power consumption and especially cost.
High end FPGAs can easily run in the $30k for a single unit range (this is what NVidia is going to buy to simulate their latest design in hardware, its not likely going to end up in a shipping product). They are also relative power hogs. Their strength lies in highly concurrent systems, especially around data streaming from disparate interfaces to other interfaces, when you can't put enough I/O bandwidth into an existing processing solution.
GPUs are good at processing and will floor an FPGA there. They are relatively terrible at I/O.
Where’s the Beef? Why FPGAs Are So Fast -- http://research.microsoft.com/pubs/70636/tr-2008-130.pdf
BLAS Comparison on FPGA, CPU and GPU -- http://research.microsoft.com/pubs/130834/ISVLSI_FINAL.pdf
"Results show that FPGAs offer comparable performance as well as 2.7 to 293 times better energy efficiency for the test cases that we implemented on all three platforms."
Here's a talk about compiling legacy C/C++ code to FPGAs. With a couple of modifications to the code they claim 100x-300x performance/watt improvement in various benchmarks.
slides: http://msrvideo.vo.msecnd.net/rmcvideos/103696/dl/103696.pdf
video: http://research.microsoft.com/apps/video/default.aspx?id=103...
For example they have a Black Scholes simulation that runs 30x faster and uses 6W of power instead of 68W. That's >300x power efficiency.