Comment by jorams
This documentation page[1] seems pretty clear. One primary at a time, any number of read replicas that automatically proxy writes to the primary, when compute scales to zero the data is in object storage and a new primary can spin up elsewhere.