The AI infrastructure that provides the lowest cost per token is the AI infrastructure waiting to be used.
We’ve spent a lot of time over the last two years in rooms where the same thing happens. A team shows a genuinely impressive AI pilot. Everyone nods. Then someone asks what it takes to run this for real, at scale, under the security and compliance rules the business actually lives with. The room goes quiet.
That gap between a working AI model and a production environment is where time, money and momentum disappear. Because the hard part of AI isn’t just training a model or buying the compute. It’s standing up compute, networking, storage, software, power, cooling, security, observability and operations management as one system that a real team can actually run.
As AI infrastructure gets larger and denser, organizations can’t afford months of integration and validation before those investments start producing value. And at rack scale, operationalizing that infrastructure becomes even more critical.
That’s the idea behind Cisco Secure AI Factory with NVIDIA: give customers a pre-validated path to production AI instead of leaving every organization to figure it out themselves.
And now, we’re partnering with Supermicro to deliver NVIDIA Cloud Partner Reference Architecture (NCP RA)-compliant rack-scale AI infrastructure, including liquid and air-cooled systems for high-density training, inference, and agentic workflows.
The result is that we’re giving customers a more predictable path from design to deployment to validation, taking uncertainty out at each step, from the edge to the enterprise, to the neocloud and sovereign cloud organizations that serve the enterprise.
Design it. Deploy it. Prove it.
Every AI build starts with the same deceptively simple question: What should we build?
NCP RA compliance answers a big part of that question before a rack ever ships. It gives customers a known architectural foundation for how rack-scale compute, frontend and backend AI fabrics, power, cooling and management should fit together.
Cisco is the only NVIDIA technology partner to utilize its own networking switches and network operating system in an NVIDIA Cloud Partner (NCP) compliant solution. #
But a certification is the starting line, not the outcome. You still have to translate it into a specific customer environment, deploy it correctly and prove that what got built actually performs the way it was designed to.
That’s where Cisco Validated Infrastructure Services, or CVIS, comes in. CVIS carries the architecture into the customer environment, from detailed design and deployment through post installation compliance verification, and performance validation of the completed cluster using Cisco tooling. Every CVIS cluster is handed over with a complete evidence package, an end-of-test report documenting the as-built configuration, test results, and conformance to the reference architecture, so the cluster is not just deployed, but provably compliant and support-ready from day one.
In other words, NCP RA helps define what to build; CVIS helps turn that blueprint into a deployed, validated system.
The prize isn’t a certified bill of materials, but a more predictable path from design to first token, and from first token to business value.
Operationalizing AI at scale
Of course, the blueprint and deployment process only matter if you’ve got the right technology underneath them. Cisco Secure AI Factory with NVIDIA brings accelerated Cisco compute together with Cisco networking, security and observability as one architecture.
Beginning in October, Cisco will expand this to rack-scale, offering Supermicro liquid-cooled and air-cooled systems on the Cisco Global Price List, giving customers access to a broader range of dense infrastructure directly from Cisco.
That includes NVIDIA HGX and NVIDIA MGX-based platforms and NVIDIA G300 NVL72 systems, with Vera Rubin NVL72 planned to follow. This brings the rack-scale engineering, cooling expertise, and manufacturing scale needed to extend Cisco’s portfolio into the most demanding AI environments.
But this is about more than adding rack-scale compute. It’s about turning that compute into infrastructure customers can actually operate in production.
Cisco Nexus One provides a high-performance AI networking fabric with a choice of NX-OS or SONiC, built on Cisco Silicon One and NVIDIA Spectrum-X Ethernet switch silicon. Cisco AI Defense, Hybrid Mesh Firewall, Live Protect and Isovalent Runtime Security help build security into the architecture from the start rather than adding it later.
And the system has to remain manageable after deployment. Cisco Cloud Control with AgenticOps brings signals across GPUs, NICs and the network together so teams can see what’s happening across the infrastructure and identify problems before they become stalled jobs. Cisco engineering, support and lifecycle services extend that operating model into Day 2 and beyond.
That’s what operationalizing AI at rack scale means. The point is to bring the pieces production AI depends on together as a system, instead of leaving customers to integrate and operate them after the fact.
As Sharon AI co-founder and CEO James Manning put it, “with Cisco Secure AI Factory with NVIDIA, we no longer have to choose between performance, reliability or ease of management. NCP RA validation gives us the confidence that our infrastructure is optimized from day one, while rack-scale capabilities provide a seamless path to scale our AI operations as our business grows.”
One architecture, different AI needs #
Not every AI workload needs the same infrastructure. What customers do need is an architecture that can adapt as those requirements change.
At distributed sites, Cisco Unified Edge brings compute, networking, security and cloud management together to run AI closer to where data is created. In the data center, Cisco UCS, available standalone or in full-stack solutions like Cisco AI PODs continue to support enterprise AI and traditional workloads.
And for the highest-density AI environments — including neocloud and sovereign AI deployments — rack-scale systems add the performance, density and cooling required to operate at much greater scale, helping these providers deliver production AI infrastructure to the enterprise customers they serve.
The infrastructure can change with the workload. The operating model doesn’t have to. Ultimately, the value should be measured by how quickly customers can put it to work.
For more on the announcement and what Cisco is bringing to market, read the #
[full press release]. Want more? Check out the[FAQ.]