Nvidia and Cisco push the enterprise AI factory into the rack-scale era

Cisco and Nvidia are pushing their joint Secure AI Factory deeper into the enterprise stack, and the big shift is physical as much as software. The latest version is built for rack-scale deployment, with rack-to-fabric liquid cooling that supports systems above 200 kilowatts per rack. That’s aimed at the heavy stuff: reasoning systems, agentic AI and trillion-parameter training.

The pitch is simple enough. Companies can buy GPUs. The hard part is turning those GPUs into something that can actually run in production without turning the whole place into a troubleshooting exercise. Nvidia’s Marc Hamilton said the point is to close that gap, and Cisco is leaning hard on reference architectures to do it.

On the networking side, the two companies are mixing Nvidia Spectrum-X switch silicon with Cisco software, specifically NX-OS or SONiC. Cisco is also pairing its Silicon One architecture with the same software stack, giving customers a way to build around Nvidia Cloud Partner-compliant reference designs or Cisco Silicon One-based ones. Will Eatherton said that setup lets Cisco use Spectrum silicon on the scale-out GPU side and Silicon One for front-end connections, with an option for high-performance storage networking too.

Cisco is also folding in tools it had already built. Nexus One serves as a common management plane, while Cisco Cloud Control is where it brings in AgenticOps across the full system. Eatherton said Cisco is wrapping software, sales and support around the cluster as well. That sounds mundane, but in enterprise AI, mundane is the point. The prize is not a flashy demo; it’s an architecture teams can actually operate when something slows down and nobody wants to guess where the problem lives.