At the end points of that discussion you'll see the limits. VRAM and bandwidth are sort of what keeps the consoles kicking around for a long period of time. The more VRAM and the more bandwidth the more than can make up for the deficit in real-time compute power, in exchange for pre-computed assets, much like we saw with this generation. I would say what you saw with UE5 is only the beginning of this generation, 2021 aspirational stuff. Where games will be in 7 years time is the end point I'm looking at.
I do use the words textures, but you're right to include geometry, or shadow maps etc. All pre-baked assets.
One of the ways I like to look at things is to just remove the streaming portion out of the discussion. What you can render and see on the screen is going then be limited to what's in VRAM. Since you're not moving there is no strreaming lets say. So you would agree that if 1 system has say 2 GB of memory vs another with 16GB of memory. And you tailored the visuals to meet their VRAM sizes, you'd expect the 16GB VRAM to look much better. I guess in this example, as you have proven the PS4 over the 7870 etc, regardless of whatever it does, it just cannot run higher settings than PS4 because of that lack of memory. I'd bet if you found a stronger card with less VRAM in the same fashion would still produce inferior visuals to the PS4 for the same reason (using Death Stranding or HZD), these titles were designed with 6GB of VRAM in mind.
And so when we consider that concept, 32 would be better than 16 and so forth. The streaming element determines the rate of change we can handle in the game with respect to the size and quality of the assets. There are going to be bandwidth and compute limits so it's not like you can toss 64GB of memory on a shit video card and get great visuals. You still need the bandwidth and compute to render all those assets in the detail they deserve.