While officially impossible, theoretically it should be work, in the same way as laptop GPUs (the dedicated GPU gives the rendered image back to the integrated GPU for seamless switching instead of outputing it directly).
There's used Tesla M6 cards (top of the range Maxwell GM204 cores, 8GB of RAM) going for really cheap sometimes. They have no video output, but you can use them in laptops with hybrid graphics thanks to the output being routed through the CPUs IGP as you said.
Needs quite a bit of driver fuckery to have it recognized as either a GTX 980m or a Quadro M5000m, and you lose HDMI/DP output, but it's not a bad card for an upgrade if you only use the internal display.
I'm surprised there are no MXM to PCIe x16 adapters, these kinds of cards are cheaper than desktop ones (due to the market/pricing being totally screwed) while providing similar performance.
Depends on what NVidia takes out. If they remove the triangle fill units, they're unsuitable for graphics but OK for "mining". That leaves the more general purpose parts that let you do arithmetic in parallel.
Those might also be useful for machine learning.
