I think you're either focusing on a very small part of the market - very large websites - or you're overestimating the requirements most websites have. Either way, it seems like you're taking about a very different scenario than I am.
Endless inventory and minimized administration but higher upfront and ongoing server investment is not a benefit when you need to write a website used by under 10k people a day. Nor is it a benefit for internal services moved to the cloud for nonsense reasons. Both of which I've seen done far too frequently. Simple tooling and small virtual machines or singular dedicated servers like are easily enough for these applications.
I can definitely understand the need for a small team working on a very large, high hit count application though. I'm not saying there are no uses for it, there totally are and you make valid points, just that the uses more limited than one might expect by the hype.
So...I run one service (gratis, as it's a friend) that gets about 8K uniques a day. It costs $26 a month in AWS (it was $35 but that's apparently gone down, cool!) and makes about $400/month. That price tag is despite me not being particularly cost-conscious when rolling it out. It does use free tier AWS resources, but EC2 is not in free tier right now; we're talking about keeping DynamoDB at free tier read levels and stuff like that.
And, unlike the still-really-weird-and-ad-hoc VPS world, stuff like DR is a solved problem when-not-if you need it. The vig for AWS is consistently between 20% and 25% and you are able to leverage implicitly all the incredibly useful tooling and systems around you. If 20% to 25% of $26 is going to materially damage your business, you do not have a business.
Endless inventory and minimized administration but higher upfront and ongoing server investment is not a benefit when you need to write a website used by under 10k people a day. Nor is it a benefit for internal services moved to the cloud for nonsense reasons. Both of which I've seen done far too frequently. Simple tooling and small virtual machines or singular dedicated servers like are easily enough for these applications.
I can definitely understand the need for a small team working on a very large, high hit count application though. I'm not saying there are no uses for it, there totally are and you make valid points, just that the uses more limited than one might expect by the hype.