Reference architecture/SAIF stack · L1-L5 · GTC 2026 product line-up 3 partial·1 adj·3 planned
§ Reference architecture Cisco Secure AI Factory with NVIDIA narrative view: /roadmap ↗

The Cisco Secure AI Factory with NVIDIA reference architecture, layer by layer. Each row shows the reference products, deployment-tier coverage (via chips), and the home-lab implementation status. Product names taken from the GTC 2026 Cisco+NVIDIA co-announcement. A green ● marker means the product actually runs in the AI Defense demo.

Cisco Secure AI Factory with NVIDIA home-lab: 3 partial · 1 adj · 3 planned
Observability ThousandEyes · Talos · Splunk (future) · XDR (future) CLEEESP planned
L1
Access & Apps
Duo · Umbrella · TLS ingress · AI Defense on API edge
CLEEESP Planned
L2
AI Runtime
NVIDIA AI Enterprise · NIM · Nemotron · OpenShell · Agent Toolkit
CLEEESP Partial · 2/5
L3
Data
NetBox · vector store · knowledge base sources
CLEEESP Partial · 1/3
L4
Compute
Cisco UCS + NVIDIA RTX PRO 4500 Blackwell · Cisco Unified Edge (edge tier)
→ OCI Free Tier Lab — home-lab compute substrate (2× AMD Micro + 1× Ampere A1)
CLEEESP Adjacent
L5
Network Fabric
Cisco N9100 (102.4T · Spectrum-6) · Silicon One G300 / P200 · NVIDIA Spectrum-X · Cisco Nexus One (mgmt + Hyperfabric) · NX-OS / SONiC
CLEEESP Planned
Security Cisco AI Defense · Cisco Hybrid Mesh Firewall on NVIDIA BlueField DPUs · NVIDIA NeMo Guardrails · Tetragon eBPF · Cisco XDR (future) CLEEESP Partial · 2 / 5
Partial — at least one product from the layer's reference list runs in the demo Adjacent — home-lab substitute for the reference product Planned — reference-only ● = product live in demo compose stack
Deployment tiers C Central DC LE Local Edge EE Enterprise Edge SP Service Provider Edge

Deployment tiers. SAIF is positioned across four tiers per the GTC 2026 co-announcement — Central DC · Local Edge · Enterprise Edge · SP Edge — reflected as chips on each layer above. Lab coverage is Central DC only; Local Edge (real-time inference), Enterprise Edge (mission-critical on-prem), and SP Edge (managed multi-tenant) are reference-only. Side-by-side write-up with rollout tiers: /roadmap ↗

§ Agent security — emerging cross-cutting pattern

As LLM systems become agentic — invoking tools, calling other agents, taking actions — a distinct security surface emerges that doesn't cleanly belong to one SAIF layer. It runs at L2 · AI Runtime (via NVIDIA OpenShell + Agent Toolkit) and at the Security wrap (via NVIDIA NeMo Guardrails + Cisco AI Defense). All four products are already listed in their home layers above; the pattern is the framing — GTC 2026 calls this out as a first-class integration.

Lab coverage: Cisco AI Defense is live (three-point guard in the AI Defense demo). NeMo Guardrails · OpenShell · Agent Toolkit remain planned.