Is it better to have max power emitted spectral density everywhere and be flat, or have the same integrated energy but space out channels in the frequency span available? I think this is only a real question in the context of human laws on emission power rather than some information theory idea. It seems obvious that you want to fill in and use the entire frequency span to the absolute limit so it'd be even.

You can do that kind of thing with IQ based sampling and signal processing because it is designed for shifting information in frequency. But with something like 1gb ethernet (PAM5) that doesn't use a carrier and just changes the voltage level of a baseband pulse, one of 5 states -2, -1, 0, 1, 2, there's no way it's evenly filling the available frequency span, is it? In the frequency domain the very fast pulses are are the superposition of a bunch of very high frequency components. But does the PAM5 pulse shape result in an even and full filling of the available spectrum? I doubt it. It's probably a lot of power at the higher freqs and not much in the lower. It's a waste and it creates frequency requirements for the transmission line that are not proportional to actual summed frequency span used.

Over the same frequency span the, admittedly much more complex, RF-based solution gets much closer to the information theoretic capacity of the transmission line all things held the same. We could have 50 gigabit ethernet on the same transmission line tolerances that baseband modulated PAM5 ethernet requires for 1 gb. The transceivers would cost a lot more though. Ethernet is conceptually simple in ignoring the frequency domain. But that is just pushing more and more complexity into the magnetics of the transceivers and the quality of the transmission line.

Maybe it's time for RF ethernet again? It worked for Docsis. 70 ohm twisted pair isn't bad for RF over quite a range.