For AI training, latency is one of the limiting factors, which needs to be kept in the nanoseconds. And a light-nanosecond is famously almost exactly 1 foot.
That's why Lumen/Starcloud's designs all assume it'll be a space station with all containers connected to one central networking spine.