Shifting AI Efficiency From Model-Centric to Data-Centric Compression