The author links this in the fourth paragraph; their argument is that while compute is not a viable moat, training data may be.