Yeah, I still think that's overkill, simply because the bulk of the computation is done on a remote server anyways. All you really need on the frontend is a TCP/IP stack to send telemetry and receive commands.
If the connection is lost, the exchange can just fallback to "naive" mode.
I guess using off-the-shelf mass market hardware combined with a software stack anyone can design, setup, and implement is way easier and cheaper than a customized solution.
Consider the overhead of maintaining two entirely separate software stacks, with different libraries and controllers then? And the ongoing costs of discovering your "minimal" hardware can't accomodate a future improvement, compared to just using general compute at a marginal upfront cost, and then having everything else be familiar?
If the connection is lost, the exchange can just fallback to "naive" mode.