Experienced dev with limited hands-on big tech infrastructure experience. Based on the results I get from LLMs in domains I understand how get they even get this stuff running using AI?

Monkeys and typewriters. Throw enough character input and "It's not working" into an LLM and it will eventually produce... something.

And since it tends to reach for the most web-represented solution, that means infinite redis caches doing the same thing, k8s, and/or Vercel.

Best mental model: imagine something that produces great tactical architecture, with zero strategic architecture, running in a loop.