← Back

Wire · technology

Vera Rubin NVL72 Agentic Inference: 67x better Performance per Dollar

Published

14 September 2026

Topic

technology

Sectors

AI InfrastructureSemiconductors

Geography

United States

Source

Read at newsletter.semianalysis.com

Verified

Fusion42 · 14 September 2026 · Fusion42 review

The Rubin NVL72 platform demonstrates up to 67x better performance per dollar on agentic inference workloads compared to prior platforms, delivering significantly higher token throughput per megawatt of power. Early software results indicate Rubin can generate over twice the profit per gigawatt than Nvidia's Blackwell, with further gains expected as software and developer optimizations mature.

This Wire brief sits within Fusion42's coverage of AI Infrastructure and Semiconductors.

◆ The Wire takeaway

You cannot still assume current AI chips are good enough. Rubin rewrites cost and energy efficiency rules for AI inference, opening a huge opportunity this week for AI startups to rethink deployment cost models and explore Rubin-optimized software stacks.

Coverage

1 source · 14 Sep 2026

Related on Wire

Topics

AI InfrastructureSemiconductorsagentic-inferenceperformance-efficiencygpu-platformnvidiaai-inferencebenchmark