Cisco and Nvidia take AI factories from rack to runtime
SiliconANGLE Victoria Gayton ● Covered by 2 sources
Cisco and Nvidia are widening their AI factory push to rack-scale compute. The real battle now is getting these systems running fast enough to matter, not just bought.
Based on reporting by SiliconANGLE, Victoria Gayton — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Cisco and Nvidia are trying to turn AI infrastructure from a promise into something customers can actually switch on. The new push extends Cisco’s Secure AI Factory beyond networking and into full rack-scale compute, with liquid-cooled systems starting on Blackwell and moving to Vera Rubin. Cisco says it has partnered with Supermicro on that rack-scale layer, while wrapping the whole thing in its own sales, support and software stack.
The timing matters because the buyers are coming from different directions but hitting the same wall. Neoclouds want capacity for waiting customers. Enterprises want inference closer to home so they can move away from API-heavy setups. Sovereign AI programs are also being built now. Will Eatherton of Cisco said the pressure is about speed: getting systems up so companies can start shifting workloads into local inference.
That speed problem is exactly where the Cisco-Nvidia pitch gets interesting. Marc Hamilton said old-school enterprise buying breaks down in AI because server teams and networking teams used to meet late in the process, but AI factories need everything to line up from the start. Cisco and Nvidia are leaning on Cisco Validated Designs that match Nvidia’s Cloud Partner reference architecture, with the goal of avoiding the messy integration failures that happen when customers mix and match too late.
But the first token is not the finish line. Cisco says it is also focusing on monitoring, availability, software upgrades and lifecycle management once the cluster is live. That is the dull, unglamorous part of AI infrastructure, and it is probably the part that decides whether these systems become business plumbing or very expensive trophies.
Nvidia’s Gilad Shainer also stressed that the networking layer is meant to stay flexible. Spectrum-X lets customers run their own technologies on top of the platform, while Spectrum-X licensing allows Cisco Silicon One switches to connect to the access network. The pitch is simple enough: keep the performance promise, reduce the integration pain, and don’t force customers to throw away the tools they already know.
My take — AI-written commentary, not fact-checked reporting
This is where AI gets less mystical and more like actual infrastructure, which is a relief. The industry loves talking about tokens and model magic; the bill arrives later in racks, cooling, and the horror of day-two operations. The winners will be the vendors who make the boring parts work without turning the customer into a systems integration intern.
Read more about this at: SiliconANGLE
