AI Signal 131
Vera Rubin NVL72 achieves up to 7x token throughput per MW over Blackwell on 1.6T DeepSeek model, surpassing Huang's 3x claim
↗
SemiAnalysis inference tests on Vera Rubin NVL72 show up to 7x better token throughput per MW versus Blackwell when running a 1.6T DeepSeek model, exceeding Jensen Huang's stated 3x improvement claim for 1T-3T parameter LLMs.
For teams planning data center capacity, the actual efficiency gain from Vera Rubin over Blackwell may be significantly higher than Nvidia's official positioning, which changes the cost calculus for inference infrastructure. The 7x throughput-per-megawatt figure suggests that large-model inference workloads could see substantially better economics than the publicly claimed 3x improvement.
Written by elseif from the cluster below · every claim links back to a sourceThe three things worth knowing
Vera Rubin NVL72 shows up to 7x better token throughput per MW vs. Blackwell on a 1.6T DeepSeek model in SemiAnalysis tests.
Jensen Huang had claimed a 3x improvement for 1T-3T parameter LLMs, which these results substantially exceed.
SemiAnalysis characterizes the gap as sandbagging and claims it translates to 2x more annual profit per gigawatt.
THE CLUSTER