Nvidia is repositioning AI data-center competition around token throughput per megawatt rather than raw GPU performance. At the AI Infrastructure Summit on the 15th in Santa Clara, California, the company said combining its next-generation Vera Rubin NVL72 platform with Groq 3 LPX can raise token throughput per megawatt by up to 35x. The strategy integrates GPUs, CPUs, networking, inference software, and power management to operate the data center as an AI factory, or token-producing system. Nvidia said Vera Rubin NVL72 delivered up to 30x higher throughput per megawatt than GB300 NVL72 on DeepSeek V4 Pro in SemiAnalysis's AgentX benchmark, while cost per million tokens was up to 45x lower. Groq 3 LPX can provide up to 35x higher throughput per megawatt than GB200 NVL72 for models exceeding 2 trillion parameters. Nvidia also introduced DSX MaxLPS, a power-management technology that reallocates surplus rack power to other workloads, potentially increasing GPU capacity by up to 40% and token throughput by up to 35% within the same power envelope. SemiAnalysis separately estimated that Vera Rubin NVL72 could produce about 39% more annual revenue per gigawatt and about 42% more modeled profit than the most powerful GB300 configuration, with up to 67x the throughput relative to total cost of ownership. Nvidia shares rose 0.7% in premarket trading as investors continued reassessing the pace and profitability of AI infrastructure spending. The company also introduced the RTX PRO 5500 Blackwell Workstation Edition, with 84GB of memory and MIG (Multi-Instance GPU, technology that partitions one GPU into separate environments), although its price and launch date remain undisclosed.