In the public cloud, you can almost always just pick another AZ, another region, or another SKU. Because of how Spot priced compute works, a PayG customer will essentially always be able to provision something, somewhere.
Its actually one of the least bad ISP business models. Its fantastic for capacity planning. I worked for an ISP that got a lot of businesses signed up on 700 buck per month EOC services, 4-5 year terms, before the NBN rolled out and provided 20 times the bandwidth for 50 bucks.
Thing is, when the NBN passed those customers, they would reach out and try and move them sideways. Extend the contract for 2 years, drop 700 bucks down to 400 and bundle in voice and managed router services. Add in fixed wireless for backup. Offer them 100M business fibre for a transition up to 1k/ month.
An ISP that can do this can afford service techs, upgrades, more peering and transit. It can breathe. Trying to sell a 40 dollar TC-4 NBN service for 42 dollars is a losing proposition. The customer feels good when they get a good deal but they lose out when the ISP starts downsizing or crawling back costly transit. And if its month to month, the constant churn is crazy.
Never mind that most of that capacity will sit around unused for years. The real problem is that by the time they get around to filling the last half of the capacity, the exponential march of Moore's law -- or the equivalents for storage and networking -- will have discounted the cost of that capacity by 50% or more anyway.
Similarly, I saw a corporation "lock in" a sweet discount for the WAN network provider for 7 years where they got an upgrade from 1 Mbps to 2 Mbps and a discount of 40%. At the same time, residential broadband was already 25 Mbps minimum and business-grade fibre had three-digit Mbps speeds within months. The sales rep that got that deal signed probably laughed his ass off.
It's a fundamental misunderstanding to think that it's easier or better to do purchase things in big stair-step increments when following an exponential curve. The waste will be exponentially higher the longer the steps. Ideally, one would want to purchase hardware in the smallest possible time-steps, following product release cycles exactly. That's one of the benefits of the public cloud, you can switch CPUs (or whatever) with a button-press.
The example of this that's burned into my mind is that a vendor (Dell or HPE, I can't remember) convinced a government department to buy a 48-core AMD EPYC server for bioinformatics. This is one of those problems where you want the biggest possible single box because of the way the algorithms scale ("up", not "out"). They're stuck with that box for the next five or so years. Meanwhile the cloud public cloud is making available these monsters: https://techcommunity.microsoft.com/blog/azurehighperformanc...
Check out that beautiful exponential curve! If you "lock in" just 2 or 3 generations back, you're missing out on 90% of what you could be using.