Instances currently don't have a state like "allocating", "running", "down for maintenance", etc., so instances in a bad state show a misleading error when accessed.
They should have a top-level state and instances in a down state should provide clear messaging when accessed.