There's lots of reasons that might happen. Lots of businesses are tied to peaks at specific times of the day or seasonality. And not yet sophisticated enough to build in elasticity.

Even AWS has no reasonable automated way to be elastic in a vertical way, like auto changing instance size. Some apps can't scale well horizontally.

My database servers, for example, are built for "Easter Sunday Attendence" and underutilized the rest of the time. We do better with things like app and web servers, but there are inefficiencies.

The linked article mentions (on page 3) that, on servers, the gain was seen when not idling: "On this Tyan server, the idle power usage ended up being the same across these three most recent kernel branches. However, the power usage under load was found to have some nice improvements."

Many if not most servers are more idle than not in my experience. I haven’t seen many servers that utilize > 50% CPU on average.

Also, if you're designing for high availability, you're going to overprovision by definition, otherwise the loss of a server or datacenter is going to cause a cascading failure.

It's not just the power used while idling. It's the power usage in general.