I see people responding to this thread with examples of high-volume workflows that use a small number of machines for the database layer.

I think those examples are missing the point. One of the examples uses a Redis cluster, which is a sign that the DB alone can't support the workflow.

Another example is using a database for certificate issuance. Which I suspect is using one table or multiple that can be shard and it doesn't suffer from lock contention or joints of multiple tables.

I wouldn't be brave enough to count the number of cases where horizontal scaling is needed but I would say it is definitely not zero, especially for read replicas when high availability is needed