How the hell would foveated streaming even work, it seems physically impossible. Tracking where your eye is looking then sending that information to a server, it processing it and then streaming that back seems impossible.

Eye tracking hardware and software specifically focus on low latency, e.g. FPGA close to the sensor. The resulting packets they send is also ridiculously small (e.g 2 numbers as x,y positions of the pupils) so ... I can see that happening.

Sure eyes move very VERY fast but if you do relatively small compute on dedicated hardware it can also go quite fast while remaining affordable.