points
In cloud AI, sure. But in the smaller open weights model territory it feels like we're well into optimisation?