1,600 Gb/s per GPU on Vera Rubin NVL72
Company: CoreWeave
The claim, verbatim
CoreWeave's NVIDIA Vera Rubin NVL72 deployment runs ConnectX-9 SuperNICs at up to 1,600 Gb/s per GPU
Source (primary)
5 Lessons from Building a Multi-plane Network Fabric for Agentic AI - CoreWeave (-, news_article)
View cached copy (2026-09-01)Live source ↗
Quote: “CoreWeave pairs ConnectX-9 SuperNICs, running up to 1,600 Gb/s per GPU on our NVIDIA Vera Rubin NVL72 deployment”
How we checked this
Checked on September 15, 2026. The cited source supports every part of this claim.
The post states that CoreWeave pairs ConnectX-9 SuperNICs, running up to 1,600 Gb/s per GPU, on its Vera Rubin NVL72 deployment.
Confirmed in the source:
- CoreWeave has an NVIDIA Vera Rubin NVL72 deployment
- That deployment uses ConnectX-9 SuperNICs
- The SuperNICs run at up to 1,600 Gb/s per GPU
What we did: Read our cached copy of the publisher (https://coreweave.com/blog/5-lessons-from-building-a-multi-plane-network-fabric-for-agentic-ai) in full (19,954 characters, retrieved September 1, 2026) and checked each assertion in the claim against it.
Additional evidence
confirms 5 Lessons from Building a Multi-plane Network Fabric for Agentic AI - CoreWeave
Quote: “CoreWeave pairs ConnectX-9 SuperNICs, running up to 1,600 Gb/s per GPU on our NVIDIA Vera Rubin NVL72 deployment”