Deepseek-v4-pro AI model shows vera rubin outperforms blackwell by 30x in energy efficiency

Nvidia tested the DeepSeek-V4-PRO 1.6T model using the SemiAnalysis AgentX workload, revealing that the Vera Rubin system achieves 30 times higher tokens per megawatt (tokens/MW) compared to the GB300 NVL72. The tokens/MW metric is a key indicator of AI data center efficiency, measuring how many AI tokens can be generated per unit of power. The test highlighted four main performance metrics, including end-to-end normalized interactivity and first token latency (TTFT), which assess the user-visible output efficiency and the time taken for the first token to arrive after a request. In the DeepSeek-V4-PRO workload, the GB300 NVL72 system achieved 15 times higher tokens/MW than the H200 NVL8 Hopper setup. Nvidia noted that Blackwell’s cost per million tokens is about 10 times lower than its predecessor, allowing operators to support more agents with the same power and infrastructure budget or maintain the same capacity at a lower cost. Vera Rubin further enhances this advantage, with its NVL72 system achieving tokens/MW that is 30 times higher than the GB300 NVL72. The cost per million tokens for Vera Rubin NVL72 is 35 times lower than that of the GB300 NVL72, enabling continuous and large-scale operation of agents to meet diverse customer workloads.