The Price of Progress: Algorithmic Efficiency and the Falling Cost of AI Inference