Nobody knows the answer to that, but if it’s efficiency they are aiming for, which results in lower inference costs, it’ll last for a while. The inference to training ratio will only get higher and the margins will only get lower.