Even if you had no issues in rendering latency, you're overlooking the data aspect. Any open-world AR, is going to be continuously downloading and discovering new geometry, models, textures, mobile programs, as you walk around.
Let's say you walk up to a restaurant. As you look at it, you see a virtual waitress outside. The you can see the menu floating there, interact with virtual food models, interact with a fully animated and articulated intelligent agent. This all has to load dynamically in a an "AR Web"-like environment (no way you're going to download and install an APP per local brick and mortar).
For this to scale, and not have hideous load times, there's got to be some kind of quad-tree/oct-tree like index (or other spatial acceleration structure) that's streams in, LOD streaming, and for this to work, the "sidelinking" stuff they talk about will be a huge boon.
Because the "internet of things" in an AR world isn't an "internet of PHYSICAL things" like door locks and web cams, but also an "internet of VIRTUAL things" to control, like virtual vending machines to order food. That means somewhere nearby, there's a next-gen router-like device, that offers side-linking, and has an IP service in it serving up assets to you. Now, it could go up through your phone carrier, through the internet backbone, into a cloud provider, or it could avoid that round trip and talk directly to the restaurant's "virtual POS" 5g device.
Point is, we don't know how this is all going to work yet, and it's very premature for people to pooh pooh the use cases for higher bandwidth or lower latency.