NVIDIA has launched its Spectrum-6 Ethernet switch system for gigascale AI factories, with several large AI infrastructure operators among the first adopters.
Spectrum-6 is a 102.4-terabit-per-second Ethernet switch system built as part of NVIDIA's Vera Rubin platform. It offers twice the capacity of the previous generation and is the latest version of the Spectrum-X Ethernet platform for large AI deployments.
The rollout comes as AI developers assemble clusters with hundreds of thousands of GPUs and CPUs for model training and inference. In these systems, network performance has become critical because processors must exchange data continuously during large collective computing tasks.
Early users include CoreWeave, Microsoft, Nebius, SpaceXAI and Tesla. CoreWeave, Microsoft and Nebius are also among the first providers expected to deploy Vera Rubin-based infrastructure that includes Spectrum-6.
Early adopters
For cloud operators, a faster, more tightly managed network can allow more compute resources to function as a single pool. NVIDIA is targeting customers that need to train large models and run inference services across large clusters without bottlenecks from slower links or network failures.
"CoreWeave is built for the most demanding AI workloads, and networking is central to delivering that performance at scale," said Min Jun, Director of Product for Networking at CoreWeave. "Bringing NVIDIA Spectrum-6 and liquid-cooled Spectrum-X Ethernet infrastructure into our AI factories will help us deliver the bandwidth, resilience and efficiency customers need to train frontier models and deploy inference faster."
Nebius also described networking as a decisive factor once AI systems reach very large scale.
"At gigascale, performance comes down to coordination: keeping every GPU in lockstep so one slow link doesn't stall an entire job," said Laurelle Roseman, Vice President of Global Partnerships at Nebius. "That's what NVIDIA Spectrum-6 goes after, and why we brought it in early - a fabric that stays fast and resilient as our customers' most demanding workloads scale."
Network focus
NVIDIA argues that raw processor performance is no longer enough to judge the output of an AI factory. Large training and inference jobs rely on constant east-west traffic between accelerators, especially during collective operations that synchronise work across thousands of chips.
This creates a challenge for conventional Ethernet systems, which were largely designed for enterprise traffic moving between users, servers and storage. Spectrum-X is intended to adapt Ethernet for AI clusters that need dense, synchronised communication across many machines at once.
Spectrum-6 works with the ConnectX-9 SuperNIC in the new generation of Spectrum-X Ethernet. The wider Vera Rubin platform also includes the Vera CPU, Rubin GPU, NVLink 6 Switch and BlueField-4 DPU.
Spectrum-6 supports both pluggable and co-packaged optics, and the product range includes support for liquid cooling. That matters as AI data centres face growing power and thermal demands across compute and networking equipment.
Performance claims
The Spectrum-X platform combines switches, SuperNICs and networking software designed together for AI deployments. NVIDIA says the system manages traffic across available paths, bypasses failures and restores data flows when packets do not arrive as intended.
The platform also supports open network operating systems and different RDMA transport models, which may appeal to operators that want to retain some architectural flexibility while using NVIDIA hardware.
NVIDIA says Spectrum-X delivers up to 1.6x higher AI networking performance than standard Ethernet products and can sustain up to 95% network efficiency in deployments with more than 100,000 GPUs. It also says its hardware-accelerated multiplane topologies can reduce the number of switches required in data centres by 1.7x.
NVIDIA adds that its Spectrum-X Ethernet Photonics technology can improve power efficiency by up to 5x and increase mean time between incidents by 10x. The figures underscore how competition in AI infrastructure is expanding beyond processors into networking, cooling and systems design as operators seek to keep ever-larger clusters fully utilised.
CoreWeave, Microsoft and Nebius are expected to extend access to Vera Rubin-based systems with Spectrum-6 to developers, startups and enterprises, according to NVIDIA.