This is simply not the case 51. When using virtual machines running on top of hypervisors, if a physical machine does fail, your virtual machine is dynamically migrated off the broken hardware onto a different functional one. You never need to know about a failed motherboard, power outage, or other event physical event. The most you will lose to your customer is maybe a ping. Nothing to repair, change, or do. The physical computer has been replaced by a file representation that is spread across a large number of hard disks that are all redundant in their design in such a way, that if they fail, you don't fail.
The cloud can have differing problems as you note but they have solutions you can choose to pay and design for. You move your worries up the chain, instead of starting at a power outlet.
You are right about costs and understanding them. As far as contracts you aren't large enough to need anything of the sort. You can pay as you go and size up as you needed. This is definitely an exercise in determining feasibility based on desired budgets.
As far as rolling your own, it's foolish to want to collect old obsolete hardware and incur that depreciating asset. It's also foolish to think you can build a better network topology than a big cloud provider. They are better than you and every hosting provider in every way. This is fact.
You are not abdicating control to Google, Amzn, or msft, for the most important part- The data. That should always be portable in a well designed infrastructure, and your exit strategy should always be intact, by picking that up and taking it elsewhere.
You will want to design your systems deployments in such a way that building them they are repeatable automated tasks and assume any server can die and you can rebuild it with no end user impact.
Now are they white glove? No, but you sure aren't getting it now, and if you are paying for it, then fire those people.
In the end it's far easier to add 8GB of memory to a machine by right clicking and adding it, than going to a datacenter and doing that physically or paying someone to do it for you. You pay for that memory per hour, but you don't have a care in the world about procuring it, getting it installed or it failing.