Built for Vera Rubin, NVIDIA Spectrum-6 Arrives for Gigascale AI Factories

As AI enters the gigascale era, the networking that connects hundreds of thousands of GPUs and CPUs has emerged as a key factor determining token generation speed. In line with this, NVIDIA announced the Spectrum-6 Ethernet switch system, a core component of its next-generation Vera Rubin platform.

Spectrum-6 delivers 102.4Tbps of bandwidth per second, doubling the capacity of the previous generation. It is the next-generation model of the NVIDIA Spectrum-X Ethernet platform, providing the bandwidth and intelligence needed to operate AI factories as a single end-to-end computing system.

Key features and adoption status are as follows:

• Major customers: CoreWeave, Microsoft, Nebius, SpaceXAI, Tesla, and others are participating as early adopters.
• Technical innovation: It combines with the ConnectX-9 SuperNIC to form an AI-dedicated scale-out fabric, intelligently distributing traffic and quickly rerouting around failures.
• Efficiency and cooling: It supports pluggable and co-packaged optics (CPO) formats, and supports Liquid Cooling to improve network power efficiency and help optimize cooling across the entire AI factory.

While existing Ethernet was optimized for enterprise traffic (North-South), Spectrum-X is designed to handle the intensive data exchange (East-West) patterns that occur in large-scale AI workloads. This ensures that every GPU receives an uninterrupted supply of data, maximizing GPU utilization during large-scale model training and inference and shortening the time to derive results.