Comments by "Maggie Jetson" (@maggiejetson7904) on "Serverless was a big mistake... says Amazon" video.

This depends on the nature of the load. A lot of simulation work are bursty in nature and nobody can build an infrastructure for the worst case scenario and idle most of the time cost effectively. In those cases you are probably better off renting a lot for a short period of time than buying a server and let your work waiting in line to be finished slowly.
1
NVMe guy here. This is like how SPDK made significant performance improvement just by putting the nvme stack in user plane to avoid task switching between kernel and user space by like 50% (50% of the latency is from these switching) and most SSD architecture went verticle slices than horizontal slices because of the overhead. I think this is pretty obvious that you will likely gain a lot of performance let alone cost reduction if you go from a lot of small overhead to a larger, less overhead architecture, and massive amount of small objects, software, services etc talking to each other is easy to start but hard to scale up.
1