One can add new servers in minutes, removing has a bit more latency to it, but I'd figure with the huge price difference between rented and cloud you'll come out on top with the former in most case. Also, just use a clustering or orchestration layer in between, they range from very simple to setup and use (e.g., Proxmox VE), to quite complex but also very capable (OpenShift, kubernetes, ...).
> 2. They have huge resource requirements and don't care to invest in their own infrastructure, but can negotiate lower rates with a cloud provider
Using hetzner or other providers is not investing in their own infra, that's using (= renting) the providers infra and ability (peering, fast uplinks, datacenter perks like utility redundancy and staff on site). The second sentence may be true but probably not for most use cases that aren't huge yet, like the post here.
> 3. When their resource usage is modest but profitability is high enough that cloud expenditure is a rounding error
IFF, yes, and often infra costs are relatively low compared to salary costs, so that's definitively some optimization problem one should go through when deciding such things. Chances are that for most projects the profitability can be good but not magic money printing and infra costs are a non-negligible part that eats on their revenue, and then it's definitively worthwhile to think about avoiding the high premium most of those cloud offerings ask for.
With Vercel I don't ever think about adding servers at all, huge win.
> infra costs are relatively low compared to salary costs
Enterprise SaaS here, this is it. Any second my team spends not caring about infra is well worth it.
1, when they need to adjust rapidly between different resource usage profiles, e.g. because they are growing rapidly and can't predict what the usage will be X days in advance
2. They have huge resource requirements and don't care to invest in their own infrastructure, but can negotiate lower rates with a cloud provider
3. When their resource usage is modest but profitability is high enough that cloud expenditure is a rounding error