Nvidia Launches Spectrum-6 Switch, Doubling Bandwidth to Serve Gigawatt-Scale AI Factories
Alina Collins
Nvidia unveiled Spectrum-6, a next-generation Ethernet switch that doubles bandwidth over its predecessor, purpose-built for gigawatt-scale AI factories; CoreWeave, Microsoft, Tesla and two other major customers have already signed on, repricing the network bottleneck in large-scale AI infrastructure.
Why do AI factories need a new switch?
AI training runs hundreds of thousands of GPUs in parallel; data flows laterally between GPUs (east-west traffic). Legacy Ethernet was designed for up-down enterprise traffic (north-south) — the architecture simply doesn't match.
This means → the network is not a supporting act but the binding constraint on how much compute actually gets used. No matter how fast the GPUs, a slow network leaves the whole system waiting.
Spectrum-6 is Nvidia's answer: 102.4 T capacity, double the prior generation, built to solve lateral communication at massive AI-cluster scale.
What do the performance numbers really mean?
Compared with generic Ethernet, the Spectrum-X platform delivers up to 1.6× AI networking performance and sustains 95% network efficiency at deployments exceeding 100,000 GPUs.
Hardware-accelerated multi-plane topology — routing data across multiple parallel paths — cuts switch count by 40%. Silicon photonics — transmitting data with light instead of electricity, faster and more efficient — delivers a 5× energy-efficiency gain and a 10× improvement in mean time between failures.
In plain terms = fewer devices, lower power, bigger AI clusters — directly compressing the cost per token.
Why are five major customers betting early?
CoreWeave, Microsoft, Nebius, SpaceX AI and Tesla have adopted Spectrum-6. CoreWeave, Microsoft and Nebius will be the first cloud providers to deploy Vera Rubin infrastructure on top of it.
Nebius VP of Global Partnerships Laurelle Roseman: "At gigawatt scale, coordination determines overall performance — keeping every GPU in sync and preventing one slow link from dragging down the entire job."
This means → for cloud providers, Spectrum-6 unifies more GPUs into a schedulable, high-performance resource pool, shortening training cycles and accelerating inference response. For enterprises building their own infrastructure, it raises system utilization on communication-heavy collective operations.
Where does Spectrum-6 sit in Nvidia's roadmap?
It is not a standalone product but the networking layer of the Vera Rubin compute platform. That platform integrates Vera CPU, Rubin GPU, NVLink 6 switches, ConnectX-9 SuperNICs, BlueField-4 DPUs and Spectrum-6 into an end-to-end co-designed system.
The switch ships in pluggable and co-packaged-optics form factors, both compatible with liquid cooling for full end-to-end liquid-cooled AI factory deployments.
Nvidia describes the design philosophy as "vertically integrated, horizontally open" — chips, systems and software co-designed inside the platform, while remaining compatible with standard Ethernet and open protocols on the outside.
What should we watch next?
Five marquee customers adopting in unison signals that Spectrum-6 has passed validation at commercial-deployment scale.
This reflects Nvidia's broader shift from "selling GPUs" to "selling an entire AI infrastructure platform" — the networking layer is an indispensable piece of that play.
The real observation window: the pace at which gigawatt-scale AI compute infrastructure reaches commercial operation — that will test whether the Vera Rubin platform ecosystem can deliver on schedule.
Content is for reference only, not financial advice.