NVIDIA Spectrum-6 switch arrives to power gigascale AI factories

NVIDIA unveiled Spectrum-6, the next generation of its Spectrum-X Ethernet networking platform, on July 21, 2026, as the interconnect layer for its upcoming Vera Rubin AI platform. The 102.4-terabit-per-second switch system is built to keep pace with GPU clusters that now scale past 100,000 accelerators.

What's new

Spectrum-6 doubles switch capacity over NVIDIA's previous-generation networking system, delivering 102.4 Tbps of throughput per switch. It pairs the new Spectrum-6 switch chip with NVIDIA's ConnectX-9 SuperNIC and supports both pluggable and co-packaged optics, with liquid cooling built in for the higher power density gigascale data centers now require.

NVIDIA says the platform delivers 1.6x higher AI networking performance than off-the-shelf Ethernet, while hardware-accelerated multiplane topologies cut the number of switches a data center needs by 1.7x. Across deployments exceeding 100,000 GPUs, NVIDIA is claiming up to 95% network efficiency, along with a 5x gain in power efficiency and a 10x improvement in mean time between incidents compared to prior-generation networking.

Early adopters named in the announcement include CoreWeave, Microsoft, Nebius, SpaceX AI, and Tesla. Nebius's Laurelle Roseman captured the design goal: "At gigascale, performance comes down to coordination: keeping every GPU in lockstep so one slow link doesn't stall an entire job."

Context

Spectrum-6 is engineered specifically as part of the Vera Rubin platform, NVIDIA's successor to Grace Blackwell, which the company has separately said delivers up to 10x more tokens per megawatt than the current generation. As GPU clusters grow into the hundreds of thousands of chips, networking has increasingly become the bottleneck limiting how much of that raw compute a data center can actually use — a single slow link can stall an entire distributed training or inference job. NVIDIA's Spectrum-X line, launched to compete directly with InfiniBand and standard Ethernet fabrics, has steadily added generations roughly in step with each new GPU architecture.

Why it matters

As frontier labs and cloud providers push toward gigawatt-scale AI factories, the switch fabric connecting GPUs matters as much as the chips themselves — wasted network efficiency translates directly into wasted power and capital at that scale. With hyperscalers and neoclouds like CoreWeave, Microsoft, Nebius, and SpaceX AI already lined up as early adopters, Spectrum-6 positions NVIDIA to capture not just GPU revenue but a growing share of the networking budget for next-generation AI data centers, deepening customers' dependence on NVIDIA's full-stack platform rather than mixing in third-party interconnects.