Vera Rubin NVL72 3.8x RL output throughput vs GB200 for Cognition
Company: CoreWeave
The claim, verbatim
Vera Rubin NVL72 delivered 3.8x the output token throughput per GPU of GB200 NVL72 for Cognition's reinforcement learning workloads at matched interactivity.
Source (primary)
First Vera Rubin NVL72 Customer Sees 4.8x Throughput - CoreWeave (-, news_article)
View cached copy (2026-10-01)Live source ↗Archive.org ↗
Quote: “For reinforcement learning, NVIDIA Vera Rubin NVL72 delivers 3.8x the output token throughput per GPU of NVIDIA GB200 NVL72 at matched interactivity.”
How we checked this
Checked on October 2, 2026. The cited source supports every part of this claim.
The post's Cognition SWE-2 benchmark results report 3.8x output token throughput per GPU for RL versus GB200 NVL72 at matched interactivity.
Confirmed in the source:
- Vera Rubin NVL72 delivered 3.8x the output token throughput per GPU of GB200 NVL72
- The 3.8x result applies to reinforcement learning workloads
- The result is from Cognition's benchmarks
- The comparison was made at matched interactivity
What we did: Read our cached copy of the publisher (https://www.coreweave.com/blog/cognition-becomes-first-customer-for-nvidia-vera-rubin-nvl72-on-coreweave-cloud) in full (14,073 characters, retrieved October 1, 2026) and checked each assertion in the claim against it.
Additional evidence
confirms First Vera Rubin NVL72 Customer Sees 4.8x Throughput - CoreWeave
Quote: “For reinforcement learning, NVIDIA Vera Rubin NVL72 delivers 3.8x the output token throughput per GPU of NVIDIA GB200 NVL72 at matched interactivity.”