Up to 30x More Work Per Watt: NVIDIA Vera Rubin NVL72 Sets a New Efficiency Standard for AI Agents
nvidia ·
NVIDIA announced early performance results showing its Vera Rubin NVL72 systems deliver up to 30x higher throughput per megawatt and up to 35x lower token costs than NVIDIA GB300 NVL72 on agentic AI workloads. The measurements used the SemiAnalysis AgentX workload, which preserves real-world coding sessions with growing context, tool calls, and sub-agent spawning. The article explains that agentic AI consumes far more tokens than chat because agents perform multi-step reasoning and accumulate context across requests. Vera Rubin NVL72's efficiency gains stem from extreme codesign across hardware and software, including techniques like disaggregated serving, large-scale expert parallelism, distributed KV-caching, and CPU optimizations (though the current results do not yet reflect Vera CPU tool-calling performance). These advances matter for power-constrained AI factories, where throughput per megawatt drives revenue and cost per token affects profit margins. The results are currently pending SemiAnalysis review, and NVIDIA notes that Vera Rubin NVL72 is already in full production and scaling across the ecosystem.
nvidia