That's an argument for using cloud for highly spiky traffic, or for not doing actual on prem primarily.
If you do in-colo hosting with leased servers you've already cut the big initial spend drastically (e.g. deploying to a new colo for a past employer involved buying a pair of switches and some cables), though you're still usually tied in to a lengthy contract. Doing managed servers usually comes with little to no upfront cost (sometimes a low setup fee) and often month-by-month contracts.
With respect to the on demand pricing benefit vs. monthly, my experience is that the price differential is so steep vs. the cheaper managed hosting providers that it's only usually viable for loads that run less than ~6 hours a day. Most sites do not have a variable enough load to justify the engineering cost to use cloud to handle their normal day/night cycle that way - usually the variations are too small. Some do.
But note that this only pays for itself if your base load is not on a cloud setup, as the base load tends to dominate and the base load cost of most cloud setups is high enough not to be outweighed by the increased flexibility vs. a pure managed/colo/on prem setup.
It gets worse for cloud: If looking for the most cost effective, they're not competing against a pure managed/colo/on prem setup, but against a hybrid setup which can auto-scale into the cloud but usually won't. If you e.g. deploy containers, all your need is to tie cloud instances automatically into an overlay network and feed data from your monitoring system into a tool that adjust min/max instances in for an autoscaling group in AWS when load exceeds a threshold, for example.
For a typical "pure" on prem setup you'd usually aim for the daily peaks to rarely exceed 50% of on prem capacity. Maybe a bit more if you have a very predictable business, or even less if your traffic is very unpredictable.
But if you add capability to that setup to spin up and tie in cloud instances to handle peaks, and you can push that into 90%+, or even above 100% - basically whichever number turns out most cost effective. Usually if you optimise that for cost, this means you'll end up with a setup which almost never spins up cloud instances, but which is now vastly cheaper relative to a pure cloud setup because it's gained the benefit of auto-scaling at a far higher load factor than would be safe in a pure on prem/managed setup.