NVIDIA: Spectrum-6 Ethernet switch delivers 102.4 Tb/s and double the throughput for gigascale AI factories

What is Spectrum-6 and why does it matter for gigascale AI?

NVIDIA has unveiled Spectrum-6, an Ethernet switch — a networking device that routes data traffic between servers — designed specifically for gigascale AI infrastructure: data centers with tens or hundreds of thousands of GPUs that jointly train and serve large AI models. The system delivers a capacity of 102.4 Tb/s per switch, crucial for moving enormous volumes of data between GPU clusters during training.

How much faster is Spectrum-6 than the previous generation?

Spectrum-6 achieves double the throughput of the previous generation of NVIDIA Spectrum switches, with up to 1.6x better AI networking performance compared to standard Ethernet. In deployments beyond 100,000 GPUs, the system reaches 95% network efficiency, and data center operators need 1.7x fewer switches than with the previous generation — a direct saving in space, energy, and cabling costs.

SuperNIC and the Vera Rubin platform

The Spectrum-6 chip is paired with the ConnectX-9 SuperNIC — a network card optimized for AI workloads that offloads networking tasks from the CPU — as part of the broader NVIDIA Vera Rubin platform. The system supports both pluggable and co-packaged optics as well as liquid cooling, giving operators flexibility when designing gigascale data centers according to their own energy and space constraints.

Who are Spectrum-6’s early partners?

Among the early partners integrating Spectrum-6 into their infrastructure, NVIDIA names CoreWeave, Microsoft, Nebius, SpaceX AI, and Tesla — companies already building or expanding gigascale AI data centers. Their early adoption suggests NVIDIA is positioning Spectrum-6 as the standard networking backbone for the next generation of AI factories, rather than a niche upgrade.