For mixed workloads, I find that memory becomes the bottleneck first for running workloads on a given physical host. This changed after NVMe SSds became available – before that it was disk I/O.

But pricing per GB is also a factor I have to admit, currently we are outfitting with 1.5 TB memory per CPU.

I don't know, probably need to get finops to run a report to find out. More than 256gb but scaled out. Don't need a single virtual machine with that much for most things.

I imagine you rarely need a lot of RAM like that on one machine except for ML and some specialist stuff. Maybe a CDN node. Not sure.

The responses will self-select for people with trivial requirements. Nobody with real resources will divulge what those resources are.