NVIDIA Frames AI Infrastructure Performance as a Pareto Curve, Claims GB300 NVL72 Delivers Up to 25x Efficiency Over Hopper

NVIDIA is pushing back on the industry habit of judging AI infrastructure by a single benchmark, arguing that data-center performance is better understood as a Pareto curve balancing raw throughput against per-user responsiveness. In a technical explainer published this week, the company laid out why one headline number rarely captures how an AI system will actually behave under real workloads, and tied the argument to its latest rack-scale platform, the GB300 NVL72, which it says offers up to 25x better performance per watt than the previous Hopper generation.