AIBRIEF
Nvidia's new Spectrum 6 switch doubles the Ethernet capacity of the prior generation, aimed at linking hundreds of thousands of chips inside AI data centers built for Vera Rubin, the company's next AI chip platform.
Picture an AI data center stacked with hundreds of thousands of expensive chips, where one slow link between them makes all the others wait. Nvidia's Spectrum-6 is built to make that link stop being the choke point, doubling the capacity of the company's previous-generation Ethernet switch and aimed at sites designed for the upcoming Vera Rubin platform.
The launch lands with a customer list that is itself the story. Nvidia named Microsoft, CoreWeave, and Nebius as cloud operators taking Spectrum-6, and, more unusually, Tesla and SpaceX as in-house AI builders. Tesla cited the network for self-driving and robotics training; SpaceX cited it for xAI's Grok workloads. The pattern: the companies already running their own AI factories, not the ones renting them, are the early adopters.
What remains open is whether Spectrum-6's capacity claim holds up against the obvious alternatives: Broadcom's Tomahawk, Marvell's Teralynx, Cisco's Silicon One. Independent benchmarking is still pending, and it is not yet clear whether the named in-house customers are running production silicon or pilot deployments.