NVIDIA NVLink: The Scale-Up Network for AI Factories

Originally published at: NVIDIA NVLink: The Scale-Up Network for AI Factories | NVIDIA Technical Blog

The demand for AI continues to accelerate. Workloads are getting larger, models are becoming more complex, and there is mounting pressure to deploy AI compute infrastructure faster than ever. AI factories—data center-scale systems that continuously convert data and energy into intelligence—are being deployed to meet this insatiable demand. This AI factory approach to the data…

Hi @jwitsoe and the NVIDIA team,

Could you help clarify the NVLink 6 bandwidth figures in this article?

The “World-leading performance” section of the article lists 3.6 TB/s of bidirectional bandwidth per GPU and 260 TB/s of rack-level bandwidth for Vera Rubin NVL72.

However, as of September 28, 2026, NVIDIA’s product pages show different figures:

Official source Per-GPU bandwidth NVL72 bandwidth
US NVLink page 3,000 GB/s (3.0 TB/s) 216 TB/s
US Vera Rubin NVL72 product page 3.0 TB/s 216 TB/s
UK NVLink page 3,600 GB/s (3.6 TB/s) 260 TB/s

Both regional NVLink pages describe the interconnect bandwidth as bidirectional.

Which figures currently apply to Vera Rubin NVL72? Does the difference reflect a specification revision, different configurations or operating conditions, or different bandwidth definitions?

A link to the current authoritative specification would be appreciated. If some pages are outdated, could they be updated or annotated to clarify which figures apply?

I initially raised this through the Contact form, and the NVIDIA Developer Programs Team suggested asking on the forums.

Thank you!