Are GNNs/GCNs viable for graphs with no node features, with only the unique node IDs? Are they different from DeepWalk at that point?

I started to dig into GNNs for the first time and I have trouble understanding its advantages over NLP inspired embedding methods like DeepWalk and node2vec. Do GNNs only shine with node features? Or can they handle IDs/giant one-hot vectors as well? Does the usual input for GNNs only consist of a vector of handcrafted features? Are GNNs used directly for tasks like link prediction or they are just embedding generators for other models?