NVIDIA Spectrum-6 Ethernet Switch Enters Gigascale AI Factories

NVIDIA Spectrum-6 Ethernet Switch Enters Gigascale AI Factories

NVIDIA's latest networking hardware is now being deployed at the world's largest AI infrastructure sites, where hundreds of thousands of processors must work in concert to train frontier models and run inference workloads.

The Spectrum-6 Ethernet switch system delivers 102.4 terabits per second of bandwidth—double the capacity of prior-generation systems—and forms the core of NVIDIA's Spectrum-X Ethernet platform. According to NVIDIA, CoreWeave, Microsoft, Nebius, SpaceXAI and Tesla are among the first organizations bringing the technology into production.

The switch chip is engineered as part of NVIDIA's Vera Rubin platform, which integrates the company's Vera CPU, Rubin GPU, NVLink 6 Switch, ConnectX-9 SuperNIC, BlueField-4 DPU and Spectrum-6 Ethernet switch into a unified architecture. Spectrum-6 supports both pluggable and co-packaged optics and includes liquid cooling options.

Why it matters

At the scale where AI factories operate today—deployments exceeding 100,000 GPUs—network performance directly determines whether compute resources sit idle or generate tokens. Traditional Ethernet was designed for enterprise traffic patterns moving between users and servers, not the synchronized collective communications that training and inference workloads generate. A single slow network link can stall an entire training job involving thousands of accelerators exchanging data continuously. NVIDIA claims Spectrum-X Ethernet delivers up to 1.6x higher AI networking performance than standard Ethernet and sustains up to 95% network efficiency at extreme scale.

Addressing the coordination challenge

Min Jun, director of product for networking at CoreWeave, said bringing Spectrum-6 and liquid-cooled Spectrum-X infrastructure into the company's AI factories will help deliver the bandwidth and resilience customers need for frontier model training and faster inference deployment.

Laurelle Roseman, vice president of global partnerships at Nebius, framed the challenge directly: "At gigascale, performance comes down to coordination: keeping every GPU in lockstep so one slow link doesn't stall an entire job."

The Spectrum-X platform combines intelligent switches with ConnectX-9 SuperNICs and full-stack networking software engineered specifically for AI workloads. The system balances traffic across available paths, bypasses failures rapidly, and recovers when data fails to reach its destination. It supports open network operating systems and multiple RDMA transport models.

NVIDIA reports that hardware-accelerated multiplane topologies reduce the number of switches required for data centers by 1.7x. The company also cites 5x higher power efficiency and 10x improved mean time between incidents through its Spectrum-X Ethernet Photonics technology.

CoreWeave, Microsoft and Nebius will be among the first providers deploying Vera Rubin-based infrastructure with Spectrum-6, extending access to developers, startups and enterprises.

These details were first reported by NVIDIA.