You're right, but I think you underestimate the intersection of what's theoretically possible and desperation.

We'll probably see multiple different things. Some of the equipment will be repurposed for other cloud GPU workloads. We'll probably see some companies try to make adapters and drivers to get the server hardware compatible with consumer desktops (there's someone who did a writeup, either on hacker news or hack a day who got older server RAM working on a desktop with an adapter and Linux tomfoolery). Buy up the stuff being sold at a loss, strap their adapter on it, and sell for profit.

But on a low level, all this shit is turing complete. There will be performance losses and considerable overhead, but in theory translation layers and emulation should be able to be written to effectively convert GPU compute to whatever other format is needed.

If it's taking a total loss vs squeezing the last squirts of profit out, I think there's going to be attempts to squeeze.