Regressor and disturbance have moments of all orders, least squares estimator has none

We construct an example in which the least squares estimator has unbounded bias and no moments, even though the regressor and disturbance have moments of all orders and are independently distributed across observations.