NVIDIA Vice President Ian Buck stated that in the era of AI factories, the most effective way to reduce token costs is to deliver significantly improved computing performance (10 to 30 times) with each generation, rather than lowering GPU prices. The AI hardware evaluation system has shifted from pursuing single-chip FLOPs to assessing the power conversion efficiency of data centers, specifically "total token throughput per megawatt (MW)."