To be fair, the authors of Kubernetes were pretty clear from very early on that it wasn't expected to scale to "one cluster to rule them all," at least in part since that's not how the system it was designed to replicate/replace worked. I don't think that's changed, but at the same time I think there was always an idea that the same concepts used to orchestrate e.g. pods and services, etc, could be extended to the clusters themselves. From what I can tell at a quick glance, that's exactly what karmada looks like it's intending to do.

One (most likely more) of the public cloud vendors actually does this (use k8s to orchestrate K8s).

It's actually the recommended approach from K8s upstream - https://github.com/kubernetes-sigs/cluster-api

This is how vmware tanzu works also.

But what will we use to orchestrate the Karmada cluster?

Some bash scripts

It's Turtles all the way down

write a systemd service ?

Admin panel