Exploring here using the eGPU Pro with ML. Conclusion so far is that as long as you take care to open / use the right monitor for ML it is possible to get fairly good results using the eGPU with multiple camera inputs.
In these tests I am using ML virtual camera output and importing that to Zoom - all driven by the eGPU and on virtual screen on the exteranl monitor driven by the eGPU.
However the CPU has to do a lot of work - I am not quite clear what type of work the CPU is doing and what type of work the GPU is doing.
Here is the result with 4 cameras and no layouts compositions or filters: