Nvidia unveiled the Spectrum-6 Ethernet switching system on July 21, a foundational piece of infrastructure designed to connect tens of thousands of GPUs and CPUs inside massive AI factories. The system delivers 102.4 terabits per second—double the previous generation—enabling the kind of scale required to train the next wave of foundation models.

The Spectrum-6 comes in two variants. The SN6810 features 128 ports at 800 Gb/s, while the SN6800 offers 409.6 Tb/s of bandwidth across 512 ports at 800 Gb/s. Both models are liquid-cooled, a requirement when dealing with this much data throughput and the heat it generates. Early adopters include Microsoft, Tesla, CoreWeave, Nebius, and SpaceXAI.

Spectrum-X and Cross-Datacenter AI Supercomputers

The Spectrum-6 forms the core of Nvidia’s Spectrum-X platform, which also introduces Spectrum-XGS technology for linking geographically distributed data centers into a single unified AI supercomputer. That’s the future: organizations won’t build AI infrastructure in one place anymore. They’ll link multiple data centers across regions into a cohesive computing fabric that operates as a single entity.

That capability matters because it lets companies operate AI factories that span multiple facilities while maintaining unified control. If one region experiences power issues or natural disaster, the workload shifts seamlessly. Data center geography becomes a risk management decision, not a constraint on AI scale.

The Networking Bottleneck Nobody Talks About

Most AI industry coverage focuses on GPU compute and model training. Few discuss the network fabric that ties it all together. Networking can become the bottleneck. If your switches can’t move data fast enough between GPUs, your expensive chips sit idle. Nvidia’s investment in networking—with switches that cost millions—is Nvidia recognizing that the constraint will shift. Compute alone isn’t enough.

Every major AI company now faces the same question: buy Nvidia switching or build custom networking. Most will buy. That’s another revenue stream Nvidia has locked up.