from time to time I wonder about the cost benefit of using an eGPU vs just sshfs+EC2's G2xLarge and I don't think there's ever been a time in the past year where it was more worth it to have an eGPU to do ML dev work.
In isolation, it's probably not, but most of the compute-intensive researchers I know also have a need for a reasonably powerful laptop anyway. From there, it's not a particularly expensive hop to getting desktop-grade performance at your desk without needing an entire different machine. That, for me, is where the eGPU value proposition makes sense.