My understanding is GPUs aren't general purpose. If you have to resort to setting a vast network of cellular atomata to do non-GPU workloads, those servers will get trounced by a raspberry pi that costs pennies in power.

Yes But CPU and GPU diverged since forever because they do not solve the same set of problems

Can I run k8s on a GPU ? Yes, why not. Will it be efficient ? No.

(replace k8s by "whatever random code you are mostly running")

So we outsource the serial calculations to the cloud but handle branching on our cpus?