Adding more GPUs no longer translates directly into more AI throughput. NVIDIA has announced that Spectrum-6, an Ethernet switch designed as part of the Vera Rubin platform, is arriving across the world's gigascale AI data centers[1]. Each system delivers 102.4 Tb/s, twice the capacity of the previous generation. CoreWeave and Microsoft are among the companies named as early adopters[1].
The Bottleneck Has Moved From the GPU to the Network
In gigascale AI factories, where hundreds of thousands of GPUs and CPUs operate as a single machine, assembling the fastest GPUs no longer guarantees results. Large-scale training and inference depend on thousands of accelerators exchanging data continuously[1].
The difficulty is that this traffic looks nothing like ordinary enterprise workloads. In collective communications, where GPUs synchronize their results, large numbers of nodes begin transmitting at the same moment. Ethernet was designed primarily for north-south traffic moving between users, servers and storage, not for the synchronized, collective-heavy patterns of gigascale AI[1]. When a single link stalls, the delay holds up the entire job.
That gap is why NVIDIA layered Spectrum-X, a purpose-built framework, on top of standard Ethernet. Spectrum-6 anchors the next generation of that platform[1].
How 102.4 Tb/s Is Built
Spectrum-6 doubles per-switch-chip bandwidth to 102.4 Tb/s using 200G PAM4 SerDes, bundling 512 ports of 200 Gb/s each[2]. The higher port density makes it easier to build large fabrics tuned to AI traffic patterns.
Optics are the other defining feature. Spectrum-6 supports both pluggable transceivers and co-packaged optics (CPO), which place the optical engines inside the same package as the switch chip[1]. The CPO variant carries 32 silicon photonics optical engines at 3.2 Tb/s each, combined with micro-ring modulators and detachable fiber connectors[2]. Liquid cooling is also supported, allowing the entire AI factory, networking included, to be handled by a single cooling approach[1].
The chip is not meant to stand alone. Paired with the ConnectX-9 SuperNIC, it forms the next generation of Spectrum-X Ethernet. Within Vera Rubin, Spectrum-6 sits alongside the Vera CPU, Rubin GPU, NVLink 6 Switch, ConnectX-9 SuperNIC and BlueField-4 DPU as parts of the same generation[1].
What Removing the Optical Modules Changes
Eliminating pluggable transceivers and DSP retimers brings concrete returns. According to NVIDIA, Spectrum-X Ethernet Photonics improves network power efficiency roughly 5x and cuts optical loss from about 22 dB to about 4 dB, raising signal integrity by up to 64x[2]. Fewer components and fewer failure points matter quietly but consistently for training jobs that run for weeks.
At the fabric level, NVIDIA says Spectrum-X Ethernet delivers up to 1.6x higher AI networking performance than off-the-shelf Ethernet and sustains up to 95 percent network efficiency across deployments exceeding 100,000 GPUs[1]. Hardware-accelerated multiplane topologies, the company adds, cut the number of switches a data center needs by a factor of 1.7[1].
All of these figures come from NVIDIA itself. The configuration of the off-the-shelf Ethernet used as a baseline, and the measurement conditions, have not been published, so there is no guarantee the numbers reproduce in production. Reports from the early adopters will be the more useful evidence.
Early Adopters Span Cloud Providers and In-House Builders
NVIDIA names CoreWeave, Microsoft, Nebius, SpaceXAI and Tesla among the first to bring in Spectrum-6[1]. Of those, CoreWeave, Microsoft and Nebius are described as among the first providers to deploy Vera Rubin-based infrastructure with Spectrum-6[1].
Min Jun, director of product for networking at CoreWeave, said that bringing Spectrum-6 and liquid-cooled Spectrum-X Ethernet infrastructure into the company's AI factories will help deliver the bandwidth, resilience and efficiency customers need to train frontier models and deploy inference faster[1]. Laurelle Roseman, vice president of global partnerships at Nebius, said performance at gigascale comes down to coordination, keeping every GPU in lockstep so one slow link does not stall an entire job[1].
For cloud providers, the appeal is that more compute capacity can operate as a unified, high-performance resource. For those building their own infrastructure, the payoff is more GPUs working in lockstep, higher utilization during demanding collective operations and greater resilience for long-running jobs[1].
Summary
Spectrum-6 is a 102.4 Tb/s Ethernet switch engineered as part of the Vera Rubin platform, with twice the capacity of the previous generation. It bundles 512 ports of 200 Gb/s using 200G PAM4 SerDes and supports co-packaged optics and liquid cooling. NVIDIA claims up to 1.6x the performance of off-the-shelf Ethernet and up to 95 percent efficiency at more than 100,000 GPUs, though all figures are vendor-supplied and await independent verification. CoreWeave, Microsoft, Nebius, SpaceXAI and Tesla are listed among the early adopters.
Source: https://blogs.nvidia.com/blog/nvidia-spectrum-six-arrives-in-gigascale-ai-factories/
