Nvidia Shifts AI Data Center Benchmark to Token Throughput Per Watt
Nvidia has shifted its competitive benchmark for AI data centers from 'GPU performance' to 'token throughput per watt'. The company announced that its Vera Rubin NVL72 platform, combined with Groq's 3 LPX chip technology, can boost token throughput per megawatt by up to 35x. This move comes as power constraints on data centers intensify amid the proliferation of AI agents.
Nvidia's AgentX benchmark measures context growth, tool calls, and sub-agent creation together, enabling efficiency comparisons that reflect real-world work environments. The company also unveiled its DSX MaxLPS power management technology, which dynamically allocates power based on per-rack demand in AI data centers, redirecting surplus power to other compute workloads.
SemiAnalysis estimated that Vera Rubin NVL72 can generate approximately 39% more annual revenue per gigawatt and approximately 42% more modeled profit compared to the most powerful GB300 configuration. Nvidia's stock rose 0.7% in premarket trading after the announcement, despite recent volatility due to investor reassessment of AI infrastructure spending.