NVIDIA Vera Rubin and Blackwell Set a New Standard for Agentic AI Performance per Watt
NVIDIA, Thursday, October 29th, 2026
NVIDIA cites SemiAnalysis AgentX results claiming Vera Rubin NVL72 gives up to 30x the throughput per megawatt of GB300 NVL72.
SemiAnalysis AgentX replays production-style agentic-coding sessions with interleaved reasoning and tool use.
On the DeepSeek V4-Pro workload at 160 tokens per second per user, Vera Rubin NVL72 reaches up to 30x the throughput per megawatt of GB300 NVL72.
GB300 NVL72 shows up to 15x higher throughput per megawatt than H200 NVL8 on DeepSeek V4 Pro and up to 80x on Kimi K3, and up to 10x lower cost per million tokens.
NVIDIA credits serving runtimes such as SGLang, TensorRT-LLM and vLLM, DeepGEMM kernels, Dynamo disaggregated serving and NVLink.