Cisco Puts Supermicro Inside: Secure AI Factory Grows Rack-Scale Teeth for the Rubin Era

When Cisco first launched the Secure AI Factory with NVIDIA, the pitch was elegantly self-interested: enterprises were suddenly building GPU clusters, and nobody sells the network around those clusters quite like Cisco. But the architecture had an obvious gap. Cisco could hand you the fabric, the firewalls, the observability and the management layer — yet the actual accelerated compute, the roaring heart of an AI factory, always came from someone else’s price list.

On August 25, Cisco moved to close that gap. The company announced it is expanding the Secure AI Factory with NVIDIA through a new partnership with Supermicro, folding the server vendor’s high-density, liquid- and air-cooled GPU systems directly into Cisco’s AI infrastructure portfolio. Starting in October 2026, customers will be able to buy Supermicro rack-scale compute as part of an integrated, pre-validated Cisco stack — aimed squarely at enterprises, neocloud operators and sovereign clouds that want frontier-class AI capacity without becoming systems integrators themselves.

What’s actually in the box

The Secure AI Factory was never a single product; it’s a reference architecture that bundles Cisco UCS servers with NVIDIA GPUs, Cisco Silicon One or NVIDIA Spectrum-X switching, Cisco AI Defense, the Hybrid Mesh Firewall, Isovalent Runtime Security, Splunk-powered monitoring, and NVIDIA AI Enterprise software. The Supermicro expansion adds the missing layer: dense, rack-scale accelerated compute.

The target platforms tell the story of where Cisco thinks this market is going. Systems based on NVIDIA Vera Rubin NVL72 — the rack-scale platform following Blackwell’s GB200 NVL72 — and NVIDIA HGX Rubin NVL8 are the named anchors, positioning the partnership for the October-era NVIDIA platform refresh. In other words, Cisco isn’t backfilling last generation’s hardware; it’s lining up for the next one.

The cooling story is the technically interesting part. Modern rack-scale systems like the NVL72 can draw more than 200 kW per rack, a density at which, as Cisco’s own FAQ puts it, liquid cooling “becomes a system-level requirement rather than an option.” The partnership pairs Supermicro’s liquid-cooled servers with Cisco’s 100% liquid-cooled AI networking systems — specifically liquid-cooled N9000 Series switches — to deliver what Cisco calls a “rack-to-fabric” liquid-cooled AI factory. Cooling the switches as well as the servers matters more than it sounds: at these densities, the network is no longer a passive bystander to thermal design but part of the same physical engineering problem.

Why Cisco is doing this

Cisco’s president and chief product officer Jeetu Patel framed the moment bluntly: “We are at the beginning of one of the largest datacenter buildouts in history. Every organization is racing to scale AI — but speed only counts if it comes with control of data, managed token costs, and real ROI.”

That’s the enterprise-anxiety checklist in one sentence, and the expansion speaks to each item. The stack ships with security controls baked in (AI Defense, the Hybrid Mesh Firewall), validated end-to-end under Cisco’s and NVIDIA’s architecture qualification programs, and — new with this announcement — supported by Cisco Validated Infrastructure Services, aligned with NVIDIA Infrastructure Services, which certify that a deployed system actually matches the reference architecture it was sold as. Cisco is also building a dedicated large-scale AI lab to develop and test the tooling behind that validation service.

For day-two operations, NVIDIA AI Enterprise is being combined with Cisco’s AgenticOps through Cloud Control, with observability that drills down across AI job health, compute, NICs, optics and network performance. Anyone who has run a large training job knows that the failure modes live precisely in that stack of tiny components; observability that spans from the job to the optic is a genuine operational differentiator, not brochure-ware.

The strategic read

The move fills the most conspicuous hole in Cisco’s AI infrastructure strategy. The company already had a strong position in Ethernet switching, Silicon One silicon, optics, security and management, and its NVIDIA partnership gave it access to the Spectrum-X ecosystem and NVIDIA’s software stack. What it lacked was the compute layer — which meant every Secure AI Factory deal still required a third party’s servers and a customer willing to stitch the pieces together.

There’s also a competitive subtext. The networking side of the AI buildout is a three-way fight between Cisco, Arista (which just posted its first $3B quarter on AI networking demand), and NVIDIA itself, whose Spectrum-X pulls the fabric toward its own silicon. By packaging Supermicro compute with its own switches into a validated rack-to-fabric system, Cisco is arguing that the integrated-system sale beats the best-of-breed parts sale — the same logic that made HCI a multi-billion-dollar business in the storage world a decade ago.

Power density is forcing the issue. As racks blow past 200 kW, servers, switches, optics, cooling, management and software increasingly have to be designed and validated as a system rather than procured separately. The Cisco-Supermicro pairing is an acknowledgment that in the Rubin era, the rack is the new unit of deployment — and whoever assembles and validates the whole rack owns the customer relationship.

Timing-wise, the announcement lands amid a run of Rubin-related news — from India’s AM Intelligence ordering 9,000 Vera Rubin NVL72 systems for an $8B buildout, to AWS linking its Trainium strategy to NVIDIA’s NVLink and Vera platforms. Cisco’s bet is that not every Rubin-class cluster will be built by hyperscalers with infinite engineering staff; the neoclouds and sovereign operators popping up worldwide need someone to hand them a working factory.

Whether enterprises actually buy full-stack AI infrastructure from a networking company remains the open question. But with October availability timed to NVIDIA’s next-generation platform refresh, Cisco has at minimum guaranteed itself a seat at the table where the next wave of AI factories gets specced — and it arrives with a whole rack in tow, liquid-cooled and ready.