A more accurate point is not necessarily that Google will keep some more nodes up, but that it can fit small workloads like a Kubernetes master into existing no

A more accurate point is not necessarily that Google will keep some more nodes up, but that it can fit small workloads like a Kubernetes master into existing nodes that run other stuff and have bits of spare capacity. Borg's resource estimation is pretty good at squeezing things. I'm thinking specifically of masters for smaller clusters, where the fixed overhead costs of the control plane are relatively high compared to the node count. On larger clusters, the masters will be beefier, sure, but your customer will be paying for enough nodes to make up for their cost and the fixed overhead is more "diluted".