Recording is a use case where it seems like USB boxes are the preferred solution nowadays. MsMadLemon says that no matter how many inputs there are on the box, there never seems to be enough. Maybe what we really need is a low cost, high SNR, and large number of inputs digitizer box?
I can think of many professional high-channel count high-quality audio interfaces, from Burl, MOTU, Apogee, UA, and more. The problem is that you won't get bottom-scraping low cost here. But, perhaps you should just buy the Behringer X-32 Rack, that gives you 32 ins, 32 outs over High Speed USB, and oh yeah, it comes with a free mixer.
Probably would be good to base it around a Pocketbeagle (take advantage of the I/O accelerator hardware in the chip) or a low cost FPGA plus a FX2 for the USB interface. The FPGA solution would be better in terms of having very low latency and implementing antialiasing filters
Sure, you can use a Cypress FX2 part with the external memory interface (or whatever they call it) talking to an FPGA, and have the FPGA manage taking that bus and splitting it out to however-many DACs and however-many ADCs you need for your design. (High Speed USB has enough bandwidth for enough 24-bit/96 kHz channels.) But, you need to work out how many registers you need to do this, it ends up being a big chip.
And I have no idea what you're going on about with anti-aliasing filters. They're all in the converters; your FPGA sends and receives I2S (or TDM or other standard converter protocol) and the converter chips (from TI, Cirrus, AKM, ESS and others) do their thing.
Or maybe use Ethernet since it's inherently isolated to prevent ground loops?
Ethernet is widely used in professional audio applications using both Dante and AVB for high channel count applications. It works. You can buy a Dante interface module from Audinate and plug it into your design, or license their IP and roll your own. You can design your AVB product around the XMOS parts and they provide a free core.