As AI moves into production, the cost of AI inference decides which deployments survive. Lessons from a Fortune panel on running large models efficiently at scale.