I used to do FPGA designs and did a bit of that and ASIC design consultancy in the 1990s, and wondered just how far FPGAs can go before they become really marginal.
They were great when CPUs were slow (4MHz) and I used the X2000 and X3000 families effectively. But the tools (Viewlogic 4 and XACT5) were horrible, not generally backwards compatible, the company was arrogant about that (the salesman said "got to crack some eggs to make omelletes") so I had to archive the old tools (nowadays you could probably use VMWARE), had to get cracks for their dongles (if they broke they left you hanging, and I had two of them concurrently, and I doubt they would run under VMWARE).
Today you get a $7 ARM which runs at 168MHz and every one of the designs I did back then could be done as a state machine from an interrupt running at say 10-100kHz. And that's without doing anything clever.
I was really glad to get out of the FPGA business ~ 20 years ago. Since then I have seen their capacities go sky high, so big that nobody can design anything in them unless they either drop in large blocks (called "IP" these days), drop in RAM arrays, drop in a 128 bit barrel shifter
or is a wizard in a hardware description language (VHDL was used way back). And their prices went sky-high. Well, they were always high - except for the bottom-end devices. The bigger X4000 devices were £100+ and thus with very limited applications.
The remaining obvious applications were things like a 20 channel ARINC429 + 1553 interface card (with FIFOs etc) which would cost a fortune if built with chips from the likes of HOLTEK, and with an FPGA it costs a smaller fortune
And crypto (RSA and such) and packet routing for high speed links. These applications will always exist but the volumes aren't going to be great.