Nvidia used Hot Chips 2026 to lay out how it thinks the networks inside AI data centers should be built, presenting its Spectrum-X Ethernet "multiplane" network architecture.
According to ServeTheHome, which covered the session, Nvidia showed how Spectrum-X and a multiplane network form the company's vision for what it calls "AI Factory" networks — the large clusters of accelerators that train and serve modern AI models.
The presentation venue matters as much as the product. Hot Chips is a long-running technical conference where chip companies walk engineers through architectural details rather than pitch customers, so a Spectrum-X talk there signals Nvidia making its case to the people who actually design and buy this infrastructure.
The framing is also notable. Nvidia is best known for GPUs, but in a large AI cluster the chips are only useful if they can talk to each other quickly enough. Networking is increasingly the bottleneck, and by pushing an Ethernet-based architecture Nvidia is planting its flag in territory traditionally held by dedicated networking vendors.
The details beyond that come from the Hot Chips session itself. ServeTheHome's writeup is the source for what was shown; Google News carried the same story via ServeTheHome's syndication, so the reporting here traces back to a single on-the-ground account rather than multiple independent confirmations. Nvidia has not been quoted directly in the available coverage on pricing, availability, or customer deployments.
Why it matters: as AI systems scale past what any single machine can handle, the network connecting thousands of chips becomes as decisive as the chips themselves — and Nvidia is moving to own that layer too.