Cisco Adds Supermicro Rack-Scale Compute to Secure AI Factory with NVIDIA

Cisco extends the stack to rack-scale compute

Cisco has expanded its Secure AI Factory with NVIDIA through a partnership with Supermicro, adding high-density air- and liquid-cooled compute systems to an architecture previously centered on networking, security, observability and validated reference designs. Cisco announced the move on August 25, 2026.

The company says it will offer the Supermicro systems as part of its broader AI infrastructure portfolio, including dense GPU servers and rack-scale configurations. Cisco’s supporting FAQ identifies NVIDIA NVL72 rack-scale systems as well as NVIDIA HGX- and MGX-based dense GPU servers among the Supermicro options planned for the architecture.

This is an expansion of Cisco’s route to market and integration scope, not an announcement that Cisco is replacing its own UCS portfolio with Supermicro hardware. The practical change is that a customer or channel partner can source more of a large AI cluster’s compute, networking and supporting services within a Cisco-led validated design.

October 2026 is the stated ordering window

The central timing claim is supported by Cisco’s announcement and rack-scale FAQ: Cisco says Supermicro compute solutions will begin to be offered as part of Secure AI Factory with NVIDIA in October 2026. The FAQ further specifies that organizations will be able to order the rack-scale architecture from Cisco’s authorized channel partner ecosystem.

Cisco has not published a specific October order date, full regional availability schedule, detailed bill of materials or customer pricing in the materials reviewed. Procurement teams should therefore treat October as the announced availability month and confirm configuration eligibility, lead times, services coverage and local channel availability during quotation.

Cisco also positions its new Cisco Validated Infrastructure Services program as a design-validation service aligned with NVIDIA Infrastructure Services methodology. That could matter on projects where acceptance criteria must cover not only server delivery but the installed, integrated cluster.

A mixed networking design for AI fabrics

Cisco describes the expanded architecture as using Silicon One-based Cisco switches on the front end and Cisco switches based on NVIDIA Spectrum-X silicon on the back end, unified under the Cisco Nexus One architecture. Cisco’s FAQ says customers can use either Cisco NX-OS or SONiC in the design.

The front-end/back-end distinction matters for build planning. AI deployments commonly require separate or logically segmented networks for client and data access, storage, cluster management and the high-bandwidth east-west traffic that connects accelerated compute. Cisco’s published architecture materials also place storage, management, security and observability within the solution scope.

NVIDIA’s Cloud Partner reference architecture program provides a structured hardware and software blueprint for high-performance AI cloud infrastructure. Cisco says the new rack-scale offer is compliant with that program for neocloud and sovereign-cloud use cases, while its enterprise reference architectures are being updated for newer NVIDIA infrastructure generations.

Cooling becomes part of the fabric conversation

The most consequential physical-infrastructure element is Cisco’s stated support for rack-to-fabric liquid cooling. In this model, liquid cooling is not limited to the GPU server row: Cisco says liquid-cooled Supermicro servers can be deployed alongside liquid-cooled Cisco AI networking systems.

That approach raises the importance of coordinated design reviews before equipment is ordered. Integrators should establish the facility water interface and capacity, coolant distribution unit arrangement, supply and return routing, leak-detection approach, service clearances, rack loading and emergency response procedures. They should also account for the heat load of switches, transceivers and power equipment rather than sizing cooling solely around the GPU servers.

The announcement does not publish rack power figures, water temperatures, connector standards or component-level cooling requirements. Those details will remain configuration-specific and should be verified against final Supermicro, Cisco and facility documentation.

What installers and procurement teams should do now

For channel partners and data-center integrators, the announcement is a signal that large AI projects are increasingly being packaged as integrated rack-scale systems. Network switching, NICs, optics, server racks, liquid cooling, storage connectivity, power distribution, security controls and telemetry must be designed together because changes in one layer can affect commissioning and support responsibility in another.

A useful pre-quote checklist should identify the selected NVIDIA compute platform, GPU density, network topology, port and optic counts, fiber and copper pathways, out-of-band management, storage network requirements, rack power budget, cooling method, commissioning tests and support boundaries. It should also identify which parts are supplied through Cisco, Supermicro, NVIDIA software channels and any storage or Kubernetes ecosystem partner.

Cisco’s October availability target gives project teams time to complete these dependencies before ordering opens. It does not eliminate the need for site-specific engineering, especially where liquid cooling, high rack densities or new electrical distribution are involved.