Nvidia's next-generation Vera Rubin NVL72 rack system is delivering roughly five times the performance per dollar and 5.4 times the performance per megawatt of today's Blackwell GB200 NVL72, according to early benchmark data reviewed by semiconductor research firm SemiAnalysis. The numbers come from engineering samples run on the AI model DeepSeek R1, and the gap widens further when compared against GB200's own rough early bringup period in 2025, since Rubin is still in that same early stage now and its results are expected to keep improving as the surrounding software matures, the same pattern SemiAnalysis says it saw with Blackwell. The early figures were gathered by cloud provider CoreWeave and have not been independently verified. Nvidia has committed to submitting checkable numbers to the industry benchmark group InferenceX by the third quarter of 2026, with Google expected to submit results for its TPUv7 chip and AMD for its MI455X chip around the same time, which would let investors compare all three on equal footing for the first time. One reason SemiAnalysis expects a smoother rollout for Rubin is software continuity, its chip can reuse Blackwell's low-level code, unlike Blackwell, which could not reuse the prior generation's, a technical detail that can determine how quickly a new chip generation actually reaches full performance in production.
Why it matters · If these early numbers hold up under independent verification, Rubin's efficiency lead would reinforce Nvidia's pricing power and add pressure on AMD and custom-silicon rivals trying to close the gap, while the smoother software transition lowers the risk that the new generation stalls in production the way earlier chip launches sometimes have.
Worth asking · How much should investors trust Nvidia-linked performance numbers before third-party verification lands, and will Rubin's lead hold up once InferenceX publishes checkable results in Q3 2026?