also removing skip connections leads to a rougher loss landscape, hence it should be harder to find the optimal weights.
I appreciate the sibling comment perspective that memory pressure is a problem, but that can be mediated by using fewer/longer skip connections across blocks of layers.