1) Assuming system-level model
10,000 microphones * 8 kilosamples per second = 800 megasamples per second.
NVidia Tesla does 100 teraflops. I get about 125,000 FLOPS per sample. That feels adequate to me!
2) One filter per microphone per passenger. We need to divide by 250.
500 FLOPS per sample per passenger. That's more than enough for a very fancy IIR.
Of the two, I think #1 is more likely to work than #2, precisely because you want a coherent model. If you want to adjust your model for someone walking down the aisle (or any kind of system ID), that's a lot easier with 125,000 FLOPS per sample than 500 FLOPS per sample.