Well I would be happy if we could turn off and on some lights but with our CSP (not a big name) we are at different levels of turning off lights:
VM level backups definitely slow down anything even more that wants to write against that one vmdk for everything and you can't have more controllers or something. Now that would be something you eventually could work around if your CSP was willing to set backup schedules for systems individually BUT there is only one backup window and that's for all customers. And the CSP wouldn't be able to tell us when exactly we'll be in line for VM backups (yes, we have SQL Server Backups) AND he might decide as done two weeks ago to change that window without notification so we literally are looking for a window to get our DWH Jobs done at all somewhen during the day - again.
But know what the best part about that CSP is? I can literally make my VM unresponsible because they are oversubscribing by more than double what VMware recommends for hosts running SQL Server VMs (instead of 4:1 we have rather something around 10:1) since my things are rather efficient and actually do push on the CPU during loads but as we know there is Nothing really to worry about when your taskmanager simply freezes a few times for a second or something, just that your stuff still isn't getting done any faster not due to provisioned hardware but the slices shared from it.
But when I had my own test environment around I would definitely have all unneeded hosts shut down via SCVMM. Still leaves a few machines running idle but these were lower power Desktop PC repurposed anyways so it would leave me with around 30 - 40 w during nights idle. Right now I've decided to go with a NVidia AGX Xavier SBC (30w) for one video stream processing project instead of something in a 1U formfactor with an Tesla GPU (150+w exkl. all other parts), I really do like to be efficient and if I should exceed processing capabilities of one unit, adding another one will still save huge amounts of power.