Cisco Secure AI Factory, implemented in a home lab.

"What would Cisco Secure AI Factory look like at our scale?"

Every Cisco DC SE gets some version of that question every week — from a customer, from a peer, from someone at a conference. A slide deck doesn't answer it. A slide deck describes what a customer could build. A running lab shows what is running now, in the same product line-up Cisco and NVIDIA co-market to enterprise DC customers.

This site is that running lab. Every artifact here has a live URL, a public repository, or a published guide. The AI Defense demo is reachable now. The SAIF stack shows implementation status honestly, layer by layer — partial, adjacent, or planned — no overclaiming. The roadmap lays out T1–T4 rollout tiers side-by-side with the enterprise version.

The full enterprise SAIF is serious infrastructure — Cisco AI PODs built on UCS C885A M8 dense GPU servers packing 8× NVIDIA HGX H200 SXM per node (or 8× AMD MI300X / MI350X OAM) — NVIDIA's 2-8-9-400 pattern: 2 CPUs, 8 GPUs, 9 NICs, 400 Gbps per GPU to the backend fabric. Networking rides a Nexus 9364E-SG2 fabric on Cisco Silicon One G200 (51.2 Tbps, 64× 800GbE), with Cisco Hybrid Mesh Firewall offloaded to NVIDIA BlueField-3 DPUs. Scale units are 32, 64, or 128 GPUs (4 / 8 / 16 nodes), with incremental scale to 256, 512, or higher GPU clusters. Not something an SE spins up over a weekend.

But the pattern is reproducible on $0/mo substrates: OCI Free Tier compute (2× AMD E2.1.Micro + 1× ARM Ampere A1 with 2 OCPU / 12 GB†), NVIDIA NIM Free Tier hosted inference (up to 5,000 credits total: 1,000 on signup + 4,000 with business-email upgrade to a 90-day NVIDIA AI Enterprise trial — not monthly-replenishing; 40 RPM default per model), Cloudflare DNS, Caddy for auto-TLS via Let's Encrypt, GitHub Pages for static sites. Every layer that can be implemented on free-tier substrates is implemented. Where the reference requires enterprise hardware or paid services, the layer is honestly labeled — the reference architecture page spells it out.

† Oracle halved the Always Free Ampere A1 allowance from 4 OCPU / 24 GB to 2 OCPU / 12 GB on 2026-06-15; historical references to 24 GB in guides / screenshots are pre-cut. Current allowance per docs.oracle.com Always Free Resources.

Every artifact on this site is open source under Apache-2.0. The infrastructure setup is a chapter-by-chapter guide — thirteen chapters, ~10K words, from Oracle account signup through the last Caddy hostname. The AI Defense demo runs from docker-compose up. The ConsolePi build is a shopping list and 30 minutes of work.

No secret sauce. No "call me for the config." Any SE on the team who wants to reproduce this can — start-to-finish, over a weekend, at zero incremental cost. That's the whole point. When a peer asks how to build one, the answer is a URL, not a screen-share.

Cisco Cloud Control is Cisco's AI-native, agentic operating platform spanning Networking, Security, Compute, Observability, and Collaboration — the foundation of Cisco's AgenticOps model. It entered Controlled Availability in the US on 2026-06-02 at Cisco Live, with Global Availability to follow. When Cloud Control reaches customers at scale, it's where they'll run every product on this site. This lab isn't a substitute — it's the hands-on foundation SEs need to be fluent in the portfolio Cloud Control will orchestrate.

Build the stack here now. Be ready to operate it there when the cockpit ships. Full positioning on the roadmap page.

Billy Garcia, Cisco DC Systems Engineer, based in Upper Nyack, NY. This site is built on my own time, on my own OCI tenancy. It's a personal project — not a Cisco-official artifact, not affiliated with, endorsed by, sponsored by, or representing Cisco Systems, Inc. Product references are to publicly available Cisco offerings; the assembly and the voice are mine. Full statement on the disclaimer page.