I'm pessimistic
AI/ML as we know it could be useful in choosing how and where to send packets. However, the number of levels in the AI model will directly influence the latency of that decision, as will typical overhead processing, and simple semiconductor switching time. As a user moves around in a big city, they will be approaching, passing, and departing access points/antennae, hence those decisions will have to be made quite frequently. If you are streaming 3D hi-res VR needing that Terabyte per second, can you really do that without hiccoughs? With lots of buffering, you might, but then you won't be in real-time - there will be time deltas among people who are physically close, simply because the routing decisions for them differed. For people far apart, e.g. continents apart, those deltas could be very noticeable - simple speed-of-light will ensure that.
Its all a nice dream, but I expect physics will get in the way.
Oh, and how is your reliability? My understanding is that as semiconductor feature size decreases, overall reliability, especially over the long term, decreases. 10 times as many transistors of 1/10 the size will get, what - 100 times more observable bit corruptions?