One guess would be that GPUs address a wider market segment (everybody needs a display + hardware to drive it). More people buying GPUs, lower prices, more $$$ for R&D, repeat.
Another likely issue with FPGAs is that for everything but raw compute they still need support circuitry (physical I/O ports, memory interfaces, etc) that are different for every app but not really "reconfigurable" in the same sense as the FPGA fabric.
Finally, I'd wonder if reconfiguration time is part of the problem - until relatively recently, reconfiguring an FPGA was all-or-nothing and could take multiple milliseconds. Not a big deal when configuring a device once on boot, but serious headwind when trying to context-switch between different jobs that need FPGA assistance.
Another likely issue with FPGAs is that for everything but raw compute they still need support circuitry (physical I/O ports, memory interfaces, etc) that are different for every app but not really "reconfigurable" in the same sense as the FPGA fabric.
Finally, I'd wonder if reconfiguration time is part of the problem - until relatively recently, reconfiguring an FPGA was all-or-nothing and could take multiple milliseconds. Not a big deal when configuring a device once on boot, but serious headwind when trying to context-switch between different jobs that need FPGA assistance.