SemiAnalysis has published pre-release benchmarks of Nvidia's Rubin NVL72, and the headline figure is efficiency rather than raw speed. Testing on DeepSeek V4 Pro, the analysts measured 59.4 million tokens per second per megawatt at 100 tokens per second, against 28.5 million for the GB300 Blackwell system — a 2.1x gain. In the tested software configuration the advantage widened to 7.2x at 150 tokens per second. SemiAnalysis is clear about the caveats: the work was done with access to unreleased hardware and vendor assistance, so the numbers are directional, workload-specific, and not independent production results. Power, not silicon, is the binding constraint on most new data centers right now, which is why tokens per megawatt has quietly become the number operators care about most.