NVIDIA Rolls Out 102.4-Tbps Spectrum-6 Switch to AI Factories, Doubling Capacity
Updated
Updated · NVIDIA Blog · Jul 21
NVIDIA Rolls Out 102.4-Tbps Spectrum-6 Switch to AI Factories, Doubling Capacity
3 articles · Updated · NVIDIA Blog · Jul 21
Summary
102.4-terabit-per-second Spectrum-6 is being deployed in gigascale AI factories worldwide, giving NVIDIA’s next Spectrum-X Ethernet platform 2x the capacity of the prior generation.
Gigascale training and inference now link more than 100,000 GPUs and CPUs, and NVIDIA says synchronized east-west traffic makes networking—not just chip speed—the key bottleneck.
CoreWeave, Microsoft, Nebius, SpaceXAI and Tesla are among the first adopters; CoreWeave, Microsoft and Nebius will offer Vera Rubin-based infrastructure with Spectrum-6 to customers.
Spectrum-X with Spectrum-6 and ConnectX-9 is pitched as a full AI fabric, delivering up to 1.6x higher networking performance than standard Ethernet and sustaining up to 95% efficiency beyond 100,000 GPUs.
NVIDIA says the design also cuts switch counts by 1.7x and supports liquid cooling, reinforcing its push to sell integrated AI factory platforms rather than standalone chips.
With networking bottlenecks now being solved, is sustainable power delivery the next insurmountable hurdle for building gigascale AI?
Can open standards prevent vendor lock-in as NVIDIA's integrated networking solutions come to dominate the AI hardware market?
Does the race for 'Sovereign AI' create a new digital dependency on a handful of powerful tech giants?
NVIDIA Spectrum-6 and Co-Packaged Optics: Powering the 102.4 Tbps AI Networking Revolution
Overview
NVIDIA's Spectrum-6 switch marks a major step forward in AI networking, ushering in the 102.4 Tbps era to meet the growing needs of hyperscale AI factories. By integrating silicon photonics directly into the switch, NVIDIA has transformed network efficiency, reducing reliance on traditional optical transceivers that often slow down high-speed data centers. This innovation boosts network efficiency by 3.5 times, which is crucial for handling the massive data flows required to train complex AI models. As a result, Spectrum-6 provides the high-bandwidth, low-latency connectivity essential for enabling the next generation of AI workloads.