We have a read-heavy zpool with some data that's used as part of our build process, on which we see a roughly 8x savings with dedup - and because of this ZFS dedup makes it economically viable for us to store the pool on NVMe rather than spinning rust.