Just a week or so ago, the IMPORTHTML/XML and related functions in GSheets just...stopped working. There are Reddit and Google Helpdesk threads about it so it's clearly a global issue. As far as I know, no fix has landed yet.
I had to spend a weekend writing a bunch of scripts to reimplement importing functionality so an international competition hosting more than 600 people from 28 countries wouldn't collapse on the spot. Everyone on the team, including me, was initially glad we were using GSheets as it made our work much easier - until it suddenly made it much, much harder.
Google & co are doing their best to remove themselves from a rational person's critical path. They're too big, you're too small; they're full of AI and politics, you have no recourse.
Analogy doesn't work; "in-sourcing" of previously-IaaS resources as a cost-cutting measure, is used almost exclusively for "base load" (vs. elastic load, which is the comparative advantage of IaaSes); and so your in-house servers are usually running pinned at 100%. So the hardware components will wear out just as fast, if not faster, than the IaaS's servers will; and when they do, your ops team won't have as many spare parts on hand as the IaaS does; nor the ability to live-migrate the enforced VM to another exactly-equivalent substrate host in the fleet (which was already warm) to avoid downtime altogether.
Also, apart from that, there are economies of scale in reliability. Your in-house backups are probably just a ZFS pool or a RAID5+1 array or something. They're certainly not hosted within a 17+-copies Dynamo-ish system, the way most IaaS object storage (incl. archival storage) is, because the costs of doing that in-house are ridiculous.
Obviously, this puts them at the mercy of the okta if anything goes wrong with IdM. But then again, people are very bad at doing auth.
Are you implying that having more than 3 copies is somehow more reliable? Cause that sounds pretty insane.
Your wafer saw (or server) that is utilized 1% of the time is going to need a lot less maintenance than a shared wafer saw that is used 16 hours a day. You will actually experience a lot less downtime by investing in your own wafer saw (or your own servers) rather than outsourcing (to a foundry or a cloud provider).
I think the real draw of outsourcing is diffusion of responsibility: when it breaks, you can blame someone else.