> The more application specific you get, the smaller the total volume of chips

Sure, and that is the definition of niche. FPGAs are niche.

You could have made a good point that an application specific design that already requires an FPGA might want to now also have a local LLM, so putting the LLM right on the FPGA might be the most expedient option in that case.

Other than that, I don't think people are reaching for FPGAs to do LLM training or inference in general because I don't think it can be cost effective vs other options.