Comment by solid_fuel

They've doing inside-out SLAM to map the environment and tracking your eyes for foveated rendering at the same time. That's fairly heavy - I think the compute required would necessitate a much higher speed connection with a desktop and a fair bit of CPU time, so it's probably cheaper and easier to put the compute right in the headset instead of scaling up the connection.

Also, they probably landed on this rough design a few years ago - before the AI hyperscalers bought up all the RAM - so the economics changed a lot under their feet.