jain.com

Vera Rubin NVL72 4.8x token throughput (Cognition test)

Company: CoreWeave

Subject kind
product
Statement date
2026-10-02
Promised amount
4.8 x
Current status
stated

The claim, verbatim

Cognition's testing on CoreWeave showed Vera Rubin NVL72 delivered up to 4.8x higher total token throughput for SWE-2 inference versus GB200 NVL72, and 3.8x higher output-token throughput for RL workloads.

Source (primary)

CRWV's Vera Rubin Push: Can AI Infrastructure Fuel Its Growth Engine? - Yahoo Finance (-, news_article)
View cached copy (2026-10-04)Live source ↗Archive.org ↗

Quote: “the Vera Rubin NVL72 delivered up to 4.8x higher total token throughput for SWE-2 inference workloads versus a GB200 NVL72 baseline, 3.8x higher output-token throughput for reinforcement-learning workloads”

How we checked this

Checked on October 12, 2026. The cited source supports every part of this claim.

The article states that Cognition's testing on CoreWeave showed Vera Rubin NVL72 delivering up to 4.8x higher total token throughput for SWE-2 inference versus a GB200 NVL72 baseline and 3.8x higher output-token throughput for reinforcement-learning workloads. That matches every part of the claim.

Confirmed in the source:

What we did: Read our cached copy of the publisher (https://finance.yahoo.com/technology/ai/articles/crwvs-vera-rubin-push-ai-122000180.html) in full (12,270 characters, retrieved October 4, 2026) and checked each assertion in the claim against it.

Additional evidence

confirms CRWV's Vera Rubin Push: Can AI Infrastructure Fuel Its Growth Engine? - Yahoo Finance

Quote: “the Vera Rubin NVL72 delivered up to 4.8x higher total token throughput for SWE-2 inference workloads versus a GB200 NVL72 baseline, 3.8x higher output-token throughput for reinforcement-learning workloads”

View cached copy (2026-10-04)Live source ↗Archive.org ↗

confirms CoreWeave Forge Connects AI Training And Production On NVIDIA - Quantum Zeitgeist

Quote: “Cognition achieving 4.8x performance boost benchmarked against a GB200 NVL72 baseline using a real-world software engineering workload”

View cached copy (2026-10-08)Live source ↗Archive.org ↗