Cisco Adds Supermicro Racks to Its Secure AI Factory
2 min readCisco is bringing a third party into its AI data center stack. The company is expanding its Secure AI Factory with NVIDIA to the rack-scale era through a partnership with Supermicro, adding high-density liquid and air-cooled compute to an architecture Cisco previously built almost entirely from its own parts.
What a Secure AI Factory is
Cisco launched Secure AI Factory with NVIDIA as a prevalidated blueprint: networking, compute, storage and security assembled and tested as one unit so a customer does not have to integrate the pieces themselves. The pitch is time. Standing up an AI cluster from components can take quarters. A reference architecture compresses that to weeks.
The gap in that blueprint was density. Training a trillion-parameter model needs racks drawing far more power than traditional enterprise servers, which in practice means liquid cooling. Supermicro has built that hardware at volume for years.
What Cisco announced
Under the agreement described by Cisco, Supermicro liquid and air-cooled systems will be validated and sold as part of the broader Cisco AI infrastructure portfolio. The combined stack is an NVIDIA Cloud Partner compliant reference architecture, and it supports NVIDIA Vera Rubin NVL72 and HGX Rubin NVL8 platforms for trillion-parameter training and high-throughput inference.
Cisco Silicon One switches handle the front end, NVIDIA Spectrum-X switches handle the back end, and Nexus One provides a single management layer across both. Supermicro compute becomes orderable through Cisco starting in October 2026.
Why it matters
The target customers are the tell. Cisco is aiming this squarely at neoclouds and sovereign clouds: GPU specialists and government-backed operators building national AI capacity. Those buyers want a vetted full stack with a single support contract, and they are spending heavily right now.
It is also an admission that no vendor wins AI infrastructure alone. Cisco is strongest in networking and security, NVIDIA owns the accelerators, and Supermicro knows dense cooling. Bundling all three is a more realistic path than trying to build every layer in house. Watch whether Cisco converts these reference architectures into disclosed AI order numbers, which is the figure investors have been pressing for.
The AI buildout is shifting from parts to prevalidated systems, and Cisco wants to be the one selling the whole rack.