The short answer
On August 25, 2026, Cisco announced it is expanding the Secure AI Factory with NVIDIA through a new partnership with Supermicro. The addition brings high-density, liquid-cooled rack-scale GPU systems into Cisco’s existing full-stack architecture—validated to run NVIDIA Vera Rubin NVL72 and HGX Rubin NVL8 clusters capable of trillion-parameter training. Liquid cooling now supports densities beyond 200 kW per rack. New Cisco Validated Infrastructure Services (CVIS), aligned with NVIDIA’s NVIS program, will certify deployments against reference architectures. Solutions go on sale in October 2026.
For enterprise teams evaluating cloud and AI infrastructure, this matters because it signals that rack-scale AI compute is now entering the enterprise buying cycle as a pre-validated, end-to-end product—not a custom integration project.
What does the expansion add?
Cisco’s original Secure AI Factory with NVIDIA provided enterprise networking (Cisco Silicon One–based front-end switches and NVIDIA Spectrum-X back-end switches, unified by Cisco Nexus One) combined with standard-density GPU server options. The August 2026 update adds a third layer: Supermicro rack-scale and high-density server systems in both air-cooled and liquid-cooled variants.
The practical upgrade is the power budget. Standard enterprise AI systems typically operate in the 20–50 kW per rack range. The new liquid-cooled configurations support over 200 kW per rack, unlocking platforms like the NVIDIA Vera Rubin NVL72 (72 linked GPUs in a single NVL domain) that simply cannot run in conventional air-cooled environments. This shifts the hardware boundary from “inference and fine-tuning” to “pre-training and trillion-parameter inference”—a category that was previously limited to hyperscaler data centers.
The announcement also introduces rack-to-fabric liquid cooling: Cisco’s liquid-cooled networking systems now work alongside Supermicro’s liquid-cooled servers, letting operators run the entire rack—networking included—in a direct liquid cooling loop without air gaps that introduce thermal management complexity.
Full-stack architecture and NCP compliance
The expanded Secure AI Factory offers NVIDIA Cloud Partner (NCP) compliant solutions—a program that certifies entire infrastructure stacks (not just individual components) against NVIDIA’s reference architecture for training and inference workloads. NCP compliance matters for procurement because it removes the validation step that normally delays enterprise AI cluster deployments by weeks or months.
Cisco notes it is the only NVIDIA technology partner to supply its own networking switches and operating system within an NCP-compliant solution. This is technically significant: most NCP deployments use NVIDIA InfiniBand or third-party Ethernet for the high-speed GPU interconnect fabric, and Cisco is the first vendor to offer a fully Cisco-managed back-end using NVIDIA Spectrum-X rather than a mix of vendors in the same rack. For enterprise buyers, that means a single throat to choke: one support contract, one set of firmware update cycles, and unified observability through AgenticOps on Cisco Cloud Control.
The new Cisco Validated Infrastructure Services (CVIS) program, aligned with NVIDIA Infrastructure Services (NVIS), will certify that each deployment is built exactly to the reference architecture. Cisco is building a dedicated large-scale AI Lab specifically to develop CVIS tooling and test software before it reaches customers.
Why sovereign cloud is the hidden story
The headline coverage has focused on rack-scale compute numbers, but the more consequential element for US and EU enterprises may be the sovereign cloud focus. Cisco explicitly names sovereign cloud operators as a target market alongside enterprises and neoclouds—and the architecture directly addresses why sovereign AI has been slow to get started: validated reference designs for the latest NVIDIA GPU platforms have historically been available only through hyperscalers, not on-premises or in regionally isolated infrastructure.
EU enterprises in regulated industries—financial services (MiFID II, DORA), healthcare (GDPR, MDR), and public sector (NIS2)—have faced a genuine constraint: the most capable AI training infrastructure required data to pass through hyperscaler regions, creating data residency and regulatory risk. NCP-compliant Vera Rubin NVL72 deployments available to sovereign cloud operators change that calculus directly. An EU neocloud or national cloud operator can now offer hyperscaler-class AI compute within a border without custom engineering.
For teams building AI products for EU-regulated clients, this opens a path to on-premises or EU-sovereign inference that was not commercially practical before this announcement.
What does this mean for US and EU dev teams?
For teams evaluating AI infrastructure: The “build vs. buy” question for AI compute is shifting. Rack-scale GPU infrastructure has historically required custom integration: designing the network fabric, sourcing compatible servers, validating cooling, and negotiating separate support agreements for each component. Cisco’s expansion—with CVIS certification and a single vendor support model—turns rack-scale AI from a systems-integration project into a procurement decision. Teams planning training infrastructure in the 12–18 month horizon should evaluate this against colocation or hyperscaler alternatives now, before October availability creates lead-time constraints.
For teams relying on public cloud for AI workloads: This announcement does not immediately change the economics of running AI workloads on AWS, Azure, or GCP. Hyperscaler on-demand compute remains lower-friction for variable workloads. But for teams with steady-state inference at scale, fine-tuning pipelines, or pre-training runs where token costs compound, the long-run economics of owned or leased rack-scale infrastructure are becoming more favorable as the procurement process simplifies. The Cisco/NVIDIA path is designed for exactly that segment.
For EU sovereign cloud considerations: If your roadmap includes large-scale AI in a GDPR-constrained environment or you are building for public sector or financial services clients with data residency requirements, the expanded Secure AI Factory provides the first commercially available path to NCP-validated Vera Rubin–class compute outside a hyperscaler. Follow the CVIS certification program closely—it is the mechanism that makes sovereign deployments practically feasible, not just theoretically possible.
For teams not yet at training scale: Inference at current LLM sizes (7B–70B parameters) still runs well on HGX H100 or H200 class hardware available from multiple cloud providers. The Vera Rubin NVL72 is a next-generation platform primarily relevant to organizations training or running frontier-class models (>200B parameters). If your AI workload fits in a single H100 node today, the rack-scale compute story is still 12–18 months away from directly affecting your infrastructure decisions.
Frequently asked questions
What is Cisco Secure AI Factory with NVIDIA?
Cisco Secure AI Factory with NVIDIA is a validated, full-stack enterprise AI infrastructure architecture combining Cisco networking (Silicon One and NVIDIA Spectrum-X switches unified by Cisco Nexus One) with NVIDIA GPU compute. The August 2026 expansion adds rack-scale Supermicro server systems, enabling trillion-parameter training and inference at densities exceeding 200 kW per rack. It targets enterprises, neoclouds, and sovereign cloud operators who need AI infrastructure that is pre-validated, supply-chain-managed, and secure from silicon to application.
What NVIDIA hardware platforms does the expanded Secure AI Factory support?
The expanded architecture supports NVIDIA Vera Rubin NVL72 and NVIDIA HGX Rubin NVL8. Both are next-generation GPU cluster platforms designed for large-scale AI training and inference. Supermicro provides server systems in liquid-cooled and air-cooled variants; Cisco supplies the end-to-end networking. The combination is NVIDIA Cloud Partner (NCP) compliant, meaning the reference architecture has been validated against NVIDIA’s requirements for inference and training workloads.
When will the Cisco and Supermicro AI Factory solutions be available?
Cisco will begin offering Supermicro compute solutions as part of the Secure AI Factory with NVIDIA starting October 2026. Cisco Validated Infrastructure Services (CVIS)—aligned with NVIDIA Infrastructure Services (NVIS)—will certify that infrastructure is built to reference architectures, reducing deployment risk and shortening time to production.
What does the expansion mean for EU sovereign cloud customers?
Sovereign cloud customers in the EU face two constraints: regulatory requirements (GDPR, AI Act, data residency) and the need for capable AI compute that doesn’t route data through hyperscaler regions. The expanded Secure AI Factory directly addresses both: NCP-compliant solutions mean EU operators can deploy validated Vera Rubin-class infrastructure without custom integration, and Cisco’s networking layer provides the observability and security controls regulators expect. EU enterprises in FinTech, HealthTech, and regulated verticals now have a pre-certified pathway to trillion-parameter-class compute within EU borders.
Sources
Cisco Newsroom — Cisco Expands Secure AI Factory with NVIDIA for the Rack-Scale Era, August 25, 2026 (primary source)
SiliconANGLE — Cisco expands rack-scale secure AI factory infrastructure for neocloud and sovereign clouds, August 25, 2026
Cisco Investor Relations — Official announcement filing, August 25, 2026