SemiAnalysis: new Nvidia Vera Rubin servers deliver up to 7 times more tokens per megawatt than Blackwell
Prepared by Sintelligent with the help of SI. Publication time restored from the build log. . Updated: .
The new generation of hardware makes SI thinking several times cheaper in energy terms, which means there will be more agents and they will work longer.
On pre-release software, Vera Rubin NVL72 running DeepSeek V4 Pro (1.6 trillion parameters) showed up to a 7x advantage over Blackwell in tokens per megawatt, more than the three times Huang publicly promised; at 100 tokens per second per user, it delivers 59.4 million tokens/s per MW versus 28.5 million for GB300. Under a total-cost-of-ownership model at 80 tokens/s, Rubin gives 10 times more tokens per dollar. Caveats: different inference engines were used in the comparison, and at high speeds the advantage narrows to 2.72 times.