Nvidia Vera Rubin Achieves 30x Blackwell Throughput per Megawatt in DeepSeek-V4-PRO Test

Nvidia has evaluated its upcoming Vera Rubin NVL72 architecture using the SemiAnalysis AgentX benchmark on the DeepSeek-v4-PRO 1.6T model, showing a 30-fold increase in throughput per megawatt compared to the Blackwell GB300 NVL72. Additionally, Vera Rubin is projected to reduce the cost per million tokens to just 1/35th of Blackwell's cost. As AI models scale to trillions of parameters, energy efficiency and cost per token have become critical bottlenecks for data centers. This massive leap in efficiency suggests that next-generation hardware will allow operators to run complex agentic AI workloads at a fraction of current power and financial budgets. The benchmark utilized SemiAnalysis AgentX, which measures agentic reasoning workloads across metrics like end-to-end latency and Time to First Token (TTFT). For comparison, the Blackwell GB300 NVL72 itself delivers 15 times the throughput per megawatt of the Hopper-based H200 NVL8.