You don't even need neighborhood datacenters; you could just access your desktop machine in the basement with your lightweight portable mobile terminal in the living room. Just the other night I used mpv on one laptop to stream a video file from the other laptop over Wi-Fi with python2 -m SimpleHTTPServer, and XPra can do the same thing in the same way for remotely accessed applications. (Of course, ssh -X can kind of do that too, but it's a lot less efficient and more insecure.)

In theory we could have much-lower-power wireless communication systems: maybe using lasers, with MEMS corner reflectors on the mobile station to transmit, and a simple photodiode with a dichroic filter over it to receive. Or maybe using submillimeter waves from a phased-array antenna, like the Starlink terminal. Or just time-domain UWB pulse radio in conventional microwave bands, but optimized for low power usage instead of super high data rates or precise ranging.

But, right now, evidently even Bluetooth Low Energy from the leading ultra-low-power silicon vendor costs 10 milliwatts when you have it on. And it's not clear if the technologies I described above will materialize. So the amount of dumb that it makes sense to put into a wireless networked mobile terminal is only about 10 milliwatts of dumb. And 10 milliwatts is not that dumb. Even with a conventional low-power CMOS Cortex-M (300 pJ per cycle, 2 DMIPS/MHz) that's about 30 MIPS or 60 DMIPS of dumb. That's dumb like a SPARC 10 workstation from the mid-90s, not dumb like a VT100 or an analog TV tuner. With subthreshold logic it's more like 600 DMIPS of dumb, dumb like a 450 MHz Pentium II (introduced 01998, mainstream around 02000).