Assuming you're doing dedicated machines for services. My company runs on-prem with a big cluster scheduler & maintains headroom in it; small deployments of new services and modest scale-ups of existing services don't require explicit capacity requests. Only if you're going to provision a huge number of instances do you need to wait for infra to buy machines. Which also requires advance planning with the "elastic" cloud anyway.

As any cloud, you should always have spare capacity on-prem. Spinning up a VM on-prem doesn't have to be slower than in the Cloud. At this scale, you're just your own cloud provider.