Recently I saw a talk where John Carmack said he is working with camera manufacturers to find ways to reduce firmware overhead on the image coming out of the sensor for VR/AR applications to reduce latency. A lot can happen in a few decades, but this is a very hard problem to solve.
The trick is to render the pictures using the latest position of the headset. Include depth maps to make the rendering more accurate. Use ML to in-paint the gaps.
Reprojection for VR is typically done on the currently rendered frame, so the difference that the additional camera movement the reprojection is trying to account for is between 1/60th and 1/120th of a second. Here you would be trying to compensate for 30+ms, so the result is going to be a lot lower in quality.
Also for a moving scene, you'd need to send a depth or position buffer and a velocity buffer for each frame, meaning you'd need something like twice the bandwidth for your video. Probably more, since I can't imagine how you would compress that information: any artifacts are going to give weird results in the reprojected image.
Viable interactive VR will not happen over streaming (video sure). It's already tough to reduce latency to an acceptable range when games are played locally.
John Carmack has some good blog posts about this if you're interested in the details.
Removing the end user capital expense of a gaming rig sounds like a huge change. Would drastically reduce the cost of VR setups as well.