Cisco will expand its Secure AI Factory with NVIDIA offering with high-density servers from Supermicro, liquid cooling, and new rack-scale configurations built for NVIDIA’s latest platforms. The company will begin offering these solutions in October 2026, a move that reflects how AI infrastructure is shifting from buying individual GPU servers to deploying complete systems where compute, networking, power, and cooling need to be designed together.
Cisco’s new AI factory in 30 seconds
- Cisco will add air- and liquid-cooled Supermicro systems to Secure AI Factory with NVIDIA starting in October 2026.
- The architecture includes platforms such as NVIDIA Vera Rubin NVL72 and HGX Rubin NVL8.
- Cisco is combining its own Silicon One networking technology with switches based on NVIDIA Spectrum-X.
- The offering targets enterprises, neoclouds, and sovereign cloud providers.
- The goal is to deploy high-density AI clusters using previously validated architectures.
The announcement comes as building AI infrastructure starts to raise issues that go well beyond simply getting hold of GPUs. The new racks pack in more compute capacity and drive up power, thermal, and connectivity requirements. For operators, properly integrating all of those components is becoming as important a part of the project as picking the accelerator.
Cisco is trying to cover exactly that layer. Its offering brings together servers, front-end and back-end networking, cooling, observability, security, and validation services under a single, common architecture.
Supermicro joins Cisco’s architecture to boost density
The main news is the addition of Supermicro systems, cooled by both air and liquid, to Cisco’s AI infrastructure catalog.
The equipment will be validated and sold as part of Secure AI Factory with NVIDIA. This will let Cisco offer configurations considerably denser than those aimed at conventional enterprise deployments.
Liquid cooling takes center stage here. Cisco is planning installations where both Supermicro’s servers and part of the networking infrastructure can use liquid cooling, extending the thermal design from the equipment all the way to the communications fabric.
The goal is to support workloads that require huge numbers of accelerators working together, from training models with hundreds of billions or trillions of parameters to high-performance inference.
Among the platforms under consideration is NVIDIA Vera Rubin NVL72, NVIDIA’s new generation of rack-scale systems, along with NVIDIA HGX Rubin NVL8.
Rubin’s presence here matters because it places the announcement beyond current generations of GPU infrastructure. Cisco is preparing its architecture for a stage in which AI systems are increasingly designed as complete rack units rather than simply servers later connected over a network.
That shift also affects data centers. Rising densities require power supply, cooling, networking, storage, and the physical layout of equipment to be coordinated from the start.
Cisco wants the network to be part of the AI factory
The other important piece is communications.
Large AI clusters need at least two networks that serve different functions. On one hand there’s front-end connectivity, linking the infrastructure to users, applications, storage, and external services. On the other, there’s the back-end used to exchange enormous amounts of data between accelerators during training and distributed inference.
Cisco will use Silicon One-based switches for the front-end and Cisco N9100 devices based on NVIDIA Spectrum-X for the back-end. Management is integrated through Nexus One.
According to Cisco, it’s the only NVIDIA technology partner that uses its own switches and network operating system within a solution that meets NVIDIA Cloud Partner (NCP) requirements.
This certification is of particular interest to so-called neoclouds, providers specializing in GPU capacity, as well as to operators building sovereign clouds who need to deploy AI infrastructure following pre-defined configurations.
The decision also shows how competition around NVIDIA is evolving.
The economic opportunity created by GPUs no longer ends with the accelerators themselves. Around every cluster there’s a considerable amount of infrastructure made up of switches, network interfaces, optics, servers, storage, power, and cooling.
As model sizes and GPU counts grow, effective performance increasingly depends on that surrounding infrastructure.
An accelerator waiting on data or on communications with another GPU is installed capacity that isn’t producing useful work. That’s why metrics like utilization, performance per watt, and cost per token are gaining more weight compared with a comparison based purely on each chip’s theoretical power.
From installing GPU servers to certifying the entire infrastructure
Cisco will also introduce Cisco Validated Infrastructure Services (CVIS), validation services aligned with NVIDIA Infrastructure Services.
The purpose is to verify that a physical installation matches the architecture it was designed and validated against beforehand.
It might sound like a minor detail, but it matters once deployments reach hundreds or thousands of accelerators. A problem with cabling, optics, cooling, firmware, or network configuration can reduce the performance of a cluster whose cost is measured in millions of euros.
Cisco is also setting up a dedicated lab for large-scale AI infrastructure to develop tools and test software related to these services.
The company is also proposing a common observability layer. With NVIDIA AI Enterprise and the AgenticOps capabilities in Cisco Cloud Control, operators will be able to correlate the state of AI jobs with information coming from compute, network cards, optics, and the fabric’s own performance.
This correlation addresses another change brought on by distributed AI: tracking down the source of a performance drop can be considerably more complex than identifying a failed server.
The problem could be in a GPU, but it could just as easily be in a network interface, an optical link, localized congestion, or any other element that keeps the accelerators from being fed fast enough.
Neoclouds and sovereign cloud enter the target market
Cisco identifies three main markets for its architecture: enterprises, neocloud providers, and sovereign cloud operators.
Not all of them need the same scale. An enterprise might deploy private infrastructure for inference, model fine-tuning, or internal applications, while a neocloud needs to operate GPU as a service and maintain high utilization levels.
Sovereign cloud adds further requirements around the location and control of data and infrastructure.
Cisco aims to cover these scenarios with an architecture that can scale from enterprise installations up to high-density clusters.
The addition of Supermicro also fills a gap Cisco needed to close in order to compete for projects where the GPU server and rack-scale cooling carry as much weight as networking.
The move also shows where the AI infrastructure market is headed. The next battle isn’t just about selling the fastest accelerator. The ability to integrate thousands of accelerators, keep them fed with data, cool them, and make them work as a single system is becoming a central part of an AI data center’s real-world performance.
Frequently Asked Questions
What is Cisco Secure AI Factory with NVIDIA?
It’s an infrastructure architecture for running artificial intelligence workloads that integrates compute, networking, security, software, and management following designs developed jointly around the NVIDIA ecosystem.
What does Supermicro bring to the Cisco partnership?
Supermicro will provide high-density server systems cooled by air and liquid. Cisco will integrate and sell them as part of its Secure AI Factory with NVIDIA architecture.
Will it be compatible with NVIDIA Vera Rubin?
Yes. Cisco has said the expansion covers next-generation platforms such as NVIDIA Vera Rubin NVL72 and NVIDIA HGX Rubin NVL8.
When will the new systems be available?
Cisco expects to start offering Supermicro’s compute solutions within Secure AI Factory with NVIDIA in October 2026.
via: newsroom.cisco
