Essentially, inference is well-amortized across the many users.