Vera Rubin NVL72 hits up to 10X tokens per MW vs GB200 on DeepSeek R1, CoreWeave data shows

Beth_Kindig · x · 2026-09-03

CoreWeave offered the first look at Nvidia's Vera Rubin NVL72 performance on DeepSeek R1 in July, finding up to a 10X increase in tokens per MW compared to GB200.

Beth Kindig of I/O Fund's back-of-napkin math suggests revenue implications remain substantial: even assuming an 80% token price drop from a theoretical $5 to $1 per million tokens, a 10X jump in TPS per MW could still roughly double revenue.

The piece also covers Nvidia's earnings call: sovereign AI, regional AI, neoclouds and enterprise AI startups now represent about half of its business and are growing 100% a year. Nvidia broke its usual guidance cadence to project FY28 revenue of about $570B (70% growth), below analyst estimates of $691B, arguing the AI buildout is broadening beyond Big Tech even as its accelerator market share slides from 90% toward 70% by end of 2026.

Original post →

More from Venture

Venture channel →