Cisco and Nvidia Move Secure AI Factory to Rack Scale, Targeting September Orders

Cisco and Nvidia are taking their Secure AI Factory collaboration to rack-scale systems for enterprises, neoclouds and sovereign AI programs. The full solution is planned to be orderable through Cisco in September 2026, bundling the components needed to move from hardware delivery to a running AI environment.

The move addresses a systems problem rather than a simple compute purchase. A production AI cluster brings together GPUs, network interface cards, switches, cables, storage, models and software libraries; each must work in concert for the whole system to perform. Cisco and Nvidia are packaging those interdependent layers as a designed rack-scale deployment rather than leaving customers to assemble them separately.

That integration is aimed at a real operational pressure. Cisco networking engineering head Will Eatherton said neocloud customers can be lined up before the GPUs arrive, making delays between a purchase order, software deployment and handover a direct constraint on bringing capacity online. The companies are also targeting enterprises and sovereign AI programs, where performance and data-control requirements can shape infrastructure choices.

A designed rack, not a collection of parts

The expanded offering combines liquid-cooled compute, AI-optimized networking, validated designs and unified operations. Cisco is partnering with Supermicro on high-density liquid-cooled systems, including support for Nvidia NVL72 architectures. It also supports Nvidia HGX and MGX form factors, with a path toward the Vera Rubin platform.

The networking design combines Nvidia Spectrum-X Ethernet with Cisco technology, including Cisco Silicon One and NX-OS or SONiC. Spectrum-X is intended to handle adaptive routing, congestion control, remote direct memory access and lossless networking for distributed AI computing, while Cisco Silicon One covers front-end, storage and data-center interconnect needs.

Validation moves integration work upstream

Cisco announced support for Nvidia Certified Partner Reference Architecture compliance across both front-end and back-end network fabrics. That reference architecture establishes how a system should be constructed, tested and operated; Cisco Validated Designs and Cisco Validated Infrastructure Services adapt those requirements to Cisco networking and management technology.

The commercial logic is to make a complex build more repeatable. Nvidia vice president Marc Hamilton said some large infrastructure lenders have indicated they would offer preferred financing rates to customers following the Nvidia Cloud Partner reference architecture. That is a company executive’s account, not a disclosed financing program, but it shows why validated deployment can matter beyond technical setup.

What Cisco says it will handle before and after deployment

• Pre-validating full-stack performance with a 1,000-GPU internal engineering cluster before customer hardware reaches the data center.
• Using an automated toolkit and service framework that Cisco says is intended to cut deployment time from months to weeks.
• Managing compute sleds, switches and security appliances through Cisco Cloud Control’s telemetry, observability and day-two operations capabilities.

The sale extends past initial cluster startup

Cisco is positioning Nexus One and Cisco Cloud Control as a shared management layer across routing, front-end networking, storage networks and the Spectrum-X backend. It is also adding AgenticOps for AI-assisted infrastructure monitoring and lifecycle management, extending the package beyond initial installation.

That focus reflects the fact that an AI factory changes after it is deployed: models, inference software and workloads evolve, and the system needs upgrades and optimization without sacrificing availability. Hamilton said Nvidia reduced Blackwell-generation inference costs by factors ranging from 10x to 30x through software optimization over the product’s lifetime. The figure is Nvidia’s own account, but it explains why the companies are emphasizing ongoing operations rather than time to first token alone.

The target market includes enterprise data centers, neocloud providers and sovereign-cloud operators. Cisco’s next defined milestone is September 2026, when it says the full rack-scale solution will be orderable through Cisco. Customer deployment results, including whether pre-validation consistently shortens installation time, remain the consequential test of the package.