Nvidia has achieved new performance milestones in MLPerf benchmarks using its latest Blackwell Ultra GB300 NVL72 rack-scale system, claiming a 45% increase in inference throughput for the DeepSeek R-1 model compared to its predecessor, the Blackwell-based GB200 platform. The results highlight the combined impact of advanced hardware and software optimizations, positioning the GB300 as a top performer across a wide range of AI workloads. The Blackwell Ultra architecture, which powers the GB300 system, represents a significant leap forward in AI compute efficiency and scalability. Nvidia emphasizes that this performance boost is not just about raw power—it reflects improvements in memory bandwidth, interconnect speed, and specialized AI accelerators, all enhanced by software stack innovations such as updated CUDA and TensorRT optimizations. These advancements are critical for organizations building large-scale AI infrastructure, often referred to as "AI factories," where efficiency and throughput directly influence operational costs and revenue potential. Nvidia argues that the GB300’s performance gains could enable faster model deployment, lower latency, and higher throughput for real-world AI applications, from generative AI to enterprise analytics. While the Blackwell architecture underpins Nvidia’s consumer-focused RTX 50-series graphics cards—offering top-tier performance for gaming—its real impact lies in the data center. The GB200 platform has already been deployed in data centers globally to support next-generation AI models. The GB300, as the enhanced version, is designed to meet the growing demands of AI training and inference at scale. Nvidia’s latest MLPerf results underscore its continued leadership in AI hardware, especially in benchmarks that measure real-world performance across diverse models and workloads. As competition intensifies from companies like AMD and custom silicon providers, these records serve as a key differentiator for enterprises investing in AI infrastructure.