Industrial scale compute required for cutting edge AI may preclude efforts to get off of the cloud, or do you think that for 80% of useful queries we can use in-house models and servers and send the remaining 20% to the big brains in the cloud?