Well for a start, if it's an FPGA that the end user can download a bitstream to configure it is really easy to burn the chip and then Amazon has to replace it.

Well for a start, if it's an FPGA that the end user can download a bitstream to configure it is really easy to burn the chip and then Amazon has to replace it. If you sandbox it you lose some flexibility. Also, as you said, demand and cost are the problems I think of. The regular FPGA developer usually has no idea of how to interact with the cloud. I worked with FPGA for five years before going to python / backend. Most of FPGA people develop in windows and don't know web at all in the application layer. I think FPGA as a service can be something really interesting. But it's really hard to find the customers. GPU programmers are software developers, whereas FPGA people aren't.