AI has entered the gigascale era.
The world’s most advanced AI factories combine hundreds of thousands of GPUs and CPUs to train frontier models, power agent AI, and generate intelligence at unprecedented scale. At this level, networking becomes a crucial computing power multiplier that drives token generation.
NVIDIA Spectrum-6 – a 102.4 terabits per second Ethernet switch system that offers twice the capacity of previous generation systems and built as part of the NVIDIA Vera Rubin platform – marks a networking milestone and reaches the world’s gigascale AI factories.
Spectrum-6 anchors the next generation of the NVIDIA Spectrum-X Ethernet platformIt delivers the bandwidth, scalability and intelligence required to operate an AI factory as an end-to-end computing system.
Table of Contents
Leading AI developers go first
The world’s leading AI infrastructure builders – including CoreWeave, Microsoft, I won’t, SpaceXAI And Tesla – will be among the first to adopt Spectrum-6 to accelerate their AI factories.
For cloud providers, Spectrum-6 means more computing capacity can be run as a unified, high-performance resource, helping customers train models and deliver inference services faster.
“CoreWeave is designed for the most demanding AI workloads and networking is central to delivering this performance at scale,” he said Min Jun, product director for networks at CoreWeave. “Integrating NVIDIA Spectrum-6 and liquid-cooled Spectrum-X Ethernet infrastructure into our AI factories will help us provide the bandwidth, resiliency and efficiency customers need to train boundary models and deliver inference faster.”
“At gigascale, performance comes down to coordination: keeping each GPU in lockstep so that a slow connection doesn’t bring an entire job to a halt,” he said Laurelle Roseman, vice president of global partnerships at Nebius. “That’s what NVIDIA Spectrum-6 strives for and why we launched it early – a structure that remains fast and resilient even as our customers’ most demanding workloads scale.”
For AI pioneers building their own infrastructure, Spectrum-6 means more GPUs working in lockstep, higher utilization for demanding collective operations, and greater resilience for long-running tasks.
The result is faster delivery of results and better economic efficiency on an extraordinary scale in all use cases.
CoreWeave, Microsoft And I won’t will be among the first to deploy NVIDIA Vera Rubin-based infrastructure with Spectrum-6, expanding access to the platform to a broad community of developers, startups and enterprises.
AI performance is a network problem
Peak GPU performance alone can no longer predict the performance of an AI factory.
Large-scale training and inference workloads rely on thousands of accelerators continuously exchanging data. Collective communication is the fundamental process that synchronizes work between GPUs and creates intense east-west traffic, often with many systems transmitting simultaneously.
Ethernet was designed primarily for enterprise applications and north-south traffic between users, servers and storage. It was not built for the synchronized, collective-heavy communication patterns of gigascale AI.
Spectrum-X Ethernet changes that. Designed specifically for AI, it transforms Ethernet into a high-performance, scalable fabric designed to power any GPU.
The next generation of Spectrum-X Ethernet
The Spectrum 6 switch chip, together with the NVIDIA ConnectX-9 SuperNIC, forms the next generation of Spectrum-X Ethernet and was developed as part of NVIDIA Vera Rubin. bring together the NVIDIA Vera CPU, Rubin GPU, NVLink 6 Switch, ConnectX-9 SuperNIC, BlueField-4 DPU and Spectrum-6 Ethernet Switch.
Spectrum-6 supports both pluggable and packaged optical form factors. Additionally, the Spectrum 6 offerings support liquid cooling, enabling a comprehensive, end-to-end cooling approach for the entire AI factory while increasing network power efficiency.
Unlike standard Ethernet, the NVIDIA Spectrum-X Ethernet networking platform for AI factory scale-out combines intelligent switches, NVIDIA ConnectX-9 SuperNICs, and full-stack networking software, all designed together for AI. The platform’s advanced capabilities continually optimize the way traffic moves through the fabric.
Additionally, NVIDIA Spectrum-X technology intelligently balances traffic across available paths, quickly evades outages, and accurately recovers data when data traveling over a network does not reach its destination. Additionally, support for open network operating systems and a choice of RDMA transport models provides AI developers with flexibility without compromising performance.
Spectrum-X Ethernet is part of NVIDIA’s vertically integrated, horizontally open approach: co-designing silicon, systems and software across the entire computing platform while supporting standard Ethernet, open network operating systems, open protocols and a broad ecosystem of cloud providers, system manufacturers and infrastructure partners. Instead of assembling a collection of parts and then optimizing them, customers receive a complete AI factory platform designed to provide the fastest training time and lowest cost per token.
Demonstrating this approach, Spectrum-X Ethernet delivers up to 1.6x higher AI network performance than standard Ethernet and delivers up to 95% network efficiency in deployments with more than 100,000 GPUs.
Hardware-accelerated Spectrum-X multiplane topologies reduce the number of switches required for data centers by 1.7x to further accelerate network efficiency. In addition to the benefits achieved by Spectrum-X Ethernet Photonics, including 5x greater energy efficiency and 10x improved mean time between incidents, network fabric improvements continue to deliver workload accelerations beyond component-level optimization.
Find out more about the NVIDIA Spectrum-X Ethernet platform.
https://blogs.nvidia.com/blog/nvidia-spectrum-six-arrives-in-gigascale-ai-factories/
