Pushing wafers through GloFo could have a lot to do with the WSA - if I understand it correctly, AMD have to pay for a certain amount of wafers regardless of whether they're used or not so they might as well use the capacity. As for why they're using Samsung too (
BTW, does anyone know where AMD actually state this?), it's likely due to it being the same node, meaning less costs for porting the design to a different fab and not having two different-performing versions of the same GPU (which could reduce demand for one or the other).
I do wonder about Vega though. After all, TSMC do have experience with HBM on the Fury...