I think you are looking at it from the wrong perspective. It doesn't take 100x the devs to make FPGA compelling on the desktop or the server. Just like bespoke

I think you are looking at it from the wrong perspective. It doesn't take 100x the devs to make FPGA compelling on the desktop or the server. Just like bespoke accelerators in Apple Silicon are used behind a library, so too will the accelerators implemented via FPGAs. The program itself can be copied a billion times. Your argument can be made for GPUs as well, the users (end users) aren't the ones writing the shaders, but GPUs are used my hundreds of millions of people.