Nvidia is rolling out Spectrum-6, a 102.4-terabit-per-second Ethernet switch system that doubles the capacity of its previous generation and is built as part of the Vera Rubin platform. CoreWeave, Microsoft, Nebius, SpaceXAI and Tesla are among the first operators wiring Spectrum-6 into production AI factories that connect hundreds of thousands of GPUs. The pitch, in one line: at gigascale, the network is now the bottleneck, and Nvidia wants to sell the whole fabric — silicon, NICs, DPUs and software — as a single integrated system.
The bandwidth math is the headline. Spectrum-6 pairs with the ConnectX-9 SuperNIC to form the next generation of the Spectrum-X Ethernet platform, and it slots alongside the Vera CPU, Rubin GPU, NVLink 6 Switch and BlueField-4 DPU in the Vera Rubin stack. Nvidia claims Spectrum-X delivers up to 1.6x higher AI networking performance than off-the-shelf Ethernet and holds 95% network efficiency across deployments of more than 100,000 GPUs — a threshold that until recently was reserved for InfiniBand.
The efficiency gains cascade at the data-center level. Multiplane topologies hardware-accelerated by Spectrum-X reduce the number of switches a facility needs by 1.7x, cutting cabling, power draw and floor space. Spectrum-X Ethernet Photonics adds 5x higher power efficiency and a 10x improvement in mean time between incidents — a resilience number that matters when a single stalled link can idle thousands of GPUs mid-training run.
Key facts
- 01Nvidia Spectrum-6 is a 102.4 Tb/s Ethernet switch system, twice the capacity of the previous generation, engineered as part of the Vera Rubin platform.
- 02Spectrum-X Ethernet delivers up to 1.6x higher AI networking performance than off-the-shelf Ethernet and sustains 95% efficiency across deployments exceeding 100,000 GPUs.
- 03CoreWeave, Microsoft, Nebius, SpaceXAI and Tesla are among the first operators bringing Spectrum-6 into production AI factories.
- 04Multiplane topologies cut the number of switches needed by 1.7x, while Spectrum-X Ethernet Photonics claims 5x higher power efficiency and 10x improved mean time between incidents.
CoreWeave is one of the launch partners moving Spectrum-6 into liquid-cooled deployments.
“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.”— Min Jun, Director of Product for Networking at CoreWeave
Nebius is another. Both providers, alongside Microsoft, will be among the first to run Vera Rubin infrastructure with Spectrum-6 on top, extending the platform to developers and enterprises through cloud access rather than direct hardware purchase.
The underlying argument Nvidia is making is that peak GPU flops no longer predict AI factory performance. Large-scale training and inference depend on synchronized collective operations across thousands of accelerators, which generate intense east-west traffic patterns that standard Ethernet was never designed to handle. Ethernet was built for enterprise north-south flows between users, servers and storage. Spectrum-X reworks it into a scale-out fabric with adaptive traffic balancing, rapid failure bypass and precise packet-loss recovery.
Spectrum-6 supports both pluggable and co-packaged optics, and the platform is offered with liquid cooling for end-to-end thermal design across the AI factory. Nvidia is also keeping the platform open at the OS layer, supporting standard Ethernet, open network operating systems and multiple RDMA transport models — a hedge against customers who want Nvidia's performance without Nvidia's full software lock-in.
The competitive frame matters. InfiniBand has been Nvidia's traditional high-performance networking product, and it still dominates the highest-end training clusters. Spectrum-X is aimed at operators who want Ethernet's operational familiarity and ecosystem breadth but need something closer to InfiniBand's efficiency at 100,000-GPU scale. Broadcom, Arista and Cisco all sell into this same buyer, and merchant-silicon Ethernet has been closing the gap on AI workloads. A 95% efficiency claim on Nvidia's own benchmarks is a defensible position, but it will be tested by hyperscaler internal comparisons.
Caveats are worth flagging. The 1.6x and 95% numbers are Nvidia's own measurements against unspecified off-the-shelf Ethernet baselines, and real-world results depend heavily on workload mix, topology and tuning. The 5x power-efficiency and 10x reliability figures for the photonics variant are similarly vendor-supplied. Independent benchmarks from CoreWeave, Microsoft or Nebius — if any are published — will do more to settle the question than launch-day marketing.
Networking is quietly becoming the second front in the AI infrastructure war, and Nvidia is trying to close it before merchant silicon catches up. Selling a 102.4 Tb/s switch is one thing; selling it as an inseparable piece of Vera Rubin, alongside Rubin GPUs, ConnectX-9 NICs and BlueField-4 DPUs, is a business model bet — that hyperscalers will pay a premium for an integrated fabric that saves them the systems-integration work. If the 1.7x switch reduction holds up in production, the total cost of ownership case gets sharper by the quarter, and the pressure on competing Ethernet vendors to match the full stack, not just the switch chip, gets harder to answer.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
