You got most of it correct:

Except I don't have to manually create a node when the notebook server is spun up.

It gets scaled up automatically using the cluster autoscaler.

About these instances being used in pipelines, I'm not quite sure, cause I haven't had a chance to play around with that.

Although, if pipeline pods are requesting for a GPU same way notebooks do, then your pipeline jobs should be able to use GPU nodes and automatically scale down to zero.

Hope this helps :)

Feel free to DM me on Twitter: @arjun921 (it's relatively easier to respond to texts there)