Breaking

NVIDIA Begins Shipping Spectrum-6, a 102.4 Tb/s Ethernet Switch for Vera Rubin AI Factories

July 21, 2026 at 11:12 EDT

Infra & Chips

NVIDIA has begun rolling out Spectrum-6, a single-chip Ethernet switch capable of 102.4 terabits per second of switching capacity, positioning it as the networking backbone for the company's forthcoming Vera Rubin AI supercomputing platform. CoreWeave, Microsoft, Nebius, xAI and Tesla are among the first operators bringing the switch into gigascale AI clusters, according to NVIDIA.

NVIDIA Networking · Spectrum-6

A Single-Chip Switch Hits 102.4 Terabits Per Second

NVIDIA has begun rolling out Spectrum-6, the Ethernet backbone built for its Vera Rubin AI platform. It doubles both switching capacity and per-lane bandwidth over the previous generation — with CoreWeave, Microsoft, Nebius, xAI and Tesla among the first adopters.

Switching capacity, per single ASIC

Spectrum-4 (SN5000) vs Spectrum-6 (SN6000) — same unit, Tb/s

51.2 Tb/s

Spectrum-4

102.4 Tb/s

Spectrum-6

2× the capacity — and 2× the per-lane bandwidth.

128

ports of 800 Gb/s on the flagship SN6810-LD (2U, liquid-cooled)

409.6 Tb/s

SN6800-LD: four ASICs, 2,048 ports of 200 Gb/s in 5U

1.6×

AI networking performance vs conventional Ethernet

95%+

sustained effective bandwidth utilization across GPU meshes

Built for the Vera Rubin platform

Rubin GPUs

Vera CPUs

NVLink 6

Spectrum-6

Rack-scale systems designed to interconnect clusters scaling toward millions of GPUs. Many flagship models use co-packaged optics — integrating the optical engine directly with the switch silicon to cut power and improve reliability.

What excites operators

Co-packaged optics reduce power draw at scale
Better uptime and fault tolerance
Gains in training efficiency and GPU scaling
Widening ecosystem — Cisco building on the ASIC

Practical hurdles

Liquid-cooled flagships need 48–60 VDC busbar power
Requires liquid-cooling-ready data centers
Migrating from pluggable optics adds complexity
Pricing undisclosed; field evaluation still thin

Timeline: Early units are reaching lead customers now, with broader partner shipments expected to ramp through the second half of 2026 as Vera Rubin pod deployments come online.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.