NVIDIA Spectrum-6 Targets Gigascale AI Networks at 102.4 Tb/s

Spectrum-6 combines NVIDIA's switch silicon with ConnectX-9 network adapters and its networking software. NVIDIA says the platform is designed for the heavy east-west traffic created when many GPUs repeatedly exchange data during distributed training and inference. It supports pluggable and co-packaged optics, as well as liquid-cooled configurations.

The company named CoreWeave, Microsoft, Nebius, SpaceXAI, and Tesla as early adopters. It said CoreWeave, Microsoft, and Nebius would be among the first providers to deploy Vera Rubin infrastructure using Spectrum-6, extending access beyond companies that build their own clusters. These are vendor-reported adoption plans, not a claim that every named deployment is already generally available.

CoreWeave separately said it has deployed the liquid-cooled Spectrum-X SN6600-LD for Vera Rubin NVL72 systems. The provider reports 102.4 Tb/s across 64 ports running at 1.6 Tb/s, using the Spectrum-6 ASIC in a 2U switch. CoreWeave also says a rack containing 16 of the switches can reach 1.64 petabits per second of aggregate switching capacity.

CoreWeave's comparison with its prior air-cooled Spectrum-4 setup claims twice the switching performance per rack, a 62.5% smaller network footprint for a 16,000-GPU topology, and 30% better switch-layer power efficiency. Those figures come from the deployer and should be read as configuration-specific results rather than universal benchmarks. Independent reporting from SiliconANGLE corroborates the Spectrum-6 rollout and the list of early infrastructure operators.

For data and ML teams, this is primarily an infrastructure signal. At very large scale, accelerator utilization depends on how reliably the network keeps GPUs synchronized, so denser switching can affect training throughput and serving economics even when the accelerators themselves do not change. Most practitioners will encounter Spectrum-6 through cloud capacity rather than operating the switches directly; the practical test will be whether providers translate the claimed network gains into measurable workload performance and pricing.