Aurora Serverless packs a number of database instances onto a single physical machine (each isolated inside its own virtual machine using AWS’s Nitro Hypervisor). As these databases shrink and grow, resources like CPU and memory are reclaimed from shrinking workloads, pooled in the hypervisor, and given to growing workloads
Oh, wow, so the workload themselves are dynamically scaling up and down "vertically" as opposed to "horizontally" - I think this is a bit like dynamically changing the size of Docker containers that are running the databases while they're running