The performance ceilings aren't that different, and not that impactful, at least not until you get to FANG scale, and I mean literally only FANG scale. We were running a billion dollar business with on 8 fairly small VMs for the API layer, which handled all of the ecommerce transaction handling. I remember at one point we encountered a memory leak of some sort in node, and the instances were falling over and dying about once an hour, but restarting and recovering- this was causing a few % error rates to our customers. I was insistent that we get all hands on deck to figure this out ASAP, and our head of Ops type person said "kevstev, we can throw hardware at this problem to meet SLOs until you get it under control. Your monthly server costs are less than my studio apartment cost me per month in Jersey City 15 years ago."
