From the Cisco edition of November 6, 2026
Cisco IT says its multi-tenant enterprise AI factory tripled GPU utilization to ~50% and hosts MyAgent plus Cisco IQ at 99.99%+ SLA
Cisco · Vendor IT (Cisco-on-Cisco) AI infrastructure, security and observability; transferable to SLED estates with caveats · Global (Cisco enterprise data centers); U.S. SLED transferability limited
- Publisher
- Cisco Blogs
- Original publication
- November 6, 2026
- Source retrieved
- 2026-11-07
What happened
Cisco IT describes moving from project-specific GPU silos to a pooled multi-tenant AI factory after MyAgent rollout to 90,000 employees and Cisco IQ production hosting. Self-reported results: GPU utilization from ~15% to ~50% (peaks ~80%); active GPUs from ~250 to 1,568 without new real estate via rear-door heat exchangers; preference for ~80% steady-state inference on-prem; 99.99%+ SLA for MyAgent and Cisco IQ. Stack cited: Silicon One fabric, UCS-X with NVIDIA/AMD GPUs, Cisco AI Defense runtime guardrails, planned Hypershield segmentation, Live Protect on Nexus, Splunk Observability/Cloud/ITSI with Agent Observability and Tokenomics still expanding.
Why it matters
Relevant as an architecture pattern for SLED agencies and research campuses building on-prem/hybrid AI factories on Cisco networking and Splunk observability, not as a measured SLED deployment. Transferability requires matching power/cooling, multi-tenancy, FedRAMP/state privacy constraints and independent verification of utilization claims.
Evidence and measured results
Vendor-reported metrics and architecture narrative from Cisco IT authors; no independent audit, baseline methodology detail or third-party measurement of GPU utilization or SLA.
Limitations and uncertainty
Self-reported Cisco IT results; power, cooling and GPU supply assumptions may not hold in typical state data centers; Agent Observability/Tokenomics not finished.
Put this evidence to work
Lighthouse Advisory interpretation, grounded in this source. Enriched 2026-11-07; this does not change the original publication date. Each role follows the same checklist so takeaways compare across sources.
Sales
Role takeaway
- Customer problem
- Agencies and campuses need to scale agents without stranded GPUs, runaway cloud token spend or bolted-on security.
- Stakeholders
- CIO
- CTO/infrastructure
- CISO
- FinOps/budget
- Research computing lead
- Discovery questions
- What is current GPU utilization and how siloed are clusters?
- Which agent workloads must stay on-prem for sovereignty?
- Do you already own Cisco networking, AI Defense and Splunk?
- Value hypothesis
- Pooling fabric and denser cooling may raise GPU use and keep inference on-prem; Cisco’s 15%→50% figure is self-reported, not a SLED guarantee.
- Potential engagement
- A 6–8 week AI landing-zone assessment measuring baseline utilization, cooling headroom and security control gaps on one production-like workload.
- Unsupported claims
- Guaranteed tripling of GPU utilization
- 99.99% SLA in customer environments
- Universal fit of rear-door heat exchangers
Pre-sales engineering
Role takeaway
- Fit
- Conceptual fit for Cisco-centric estates building multi-tenant AI factories; not a drop-in for non-Cisco fabrics.
- Architecture and integration
- Pooled Silicon One/Nexus fabric, UCS-X NVIDIA/AMD GPUs, AI Defense runtime, planned Hypershield, Live Protect on Nexus, Splunk Observability/Cloud/ITSI.
- Prerequisites
- Sufficient power/cooling for denser racks
- Shared scheduling/tenancy model
- Identity and network path for AI Defense
- Splunk metrics/logs entitlements
- Constraints
- Vendor-reported metrics only
- Tokenomics/Agent Observability still expanding
- Thermal plant may block density gains
- No FedRAMP mapping in this post
- Security
- AI Defense prompt/response guardrails plus planned microsegmentation and Live Protect between patches; validate against agency ATO boundary.
- Proof of value
- Measure GPU utilization before/after pooling on one workload
- Confirm zero GPU packet-drop stalls under load
- Verify AI Defense blocks on a test prompt set
- Track on-prem vs cloud inference cost share
Delivery
Role takeaway
- Work
- Design and migrate from project GPU silos to a shared AI factory with security and observability wired in.
- Dependencies
- Facilities cooling upgrades
- GPU procurement
- Network fabric change windows
- SOC playbooks for AI Defense alerts
- Ownership
- Joint platform ownership: infrastructure ops owns fabric/compute; security owns AI Defense/Hypershield; FinOps owns token attribution.
- Skills and adoption
- Need fabric engineers, GPU/MLOps operators and security engineers fluent in agent guardrails; train app teams on shared tenancy rules.
- Governance checkpoints
- Architecture review of multi-tenant isolation
- Security ATO delta for AI Defense path
- FinOps gate before broad MyAgent-like rollout
- Acceptance criteria
- Documented utilization lift on named workload
- Agreed SLA and on-call model
- Token/cost dashboard for pilot agents
- Written rollback to dedicated clusters
- Risks
- Self-reported Cisco metrics may not reproduce
- Cooling project slips block density
- Multi-tenant noisy-neighbor incidents
- Unfinished Tokenomics leaves cost blind spots
Implementation considerations
Lighthouse Advisory interpretation across the operating dimensions a public-sector buyer must settle before this evidence becomes a design. Each note answers the question under its heading for this specific source.
Architecture and integration
What must connect, and where does the AI sit in the workflow?
Pooled high-bandwidth fabric, heterogeneous UCS GPUs, rack-level liquid assist and embedded AI Defense/Live Protect are the control points Cisco IT credits for utilization and security; SLED designs should validate Silicon One/Nexus and Splunk telemetry paths against local standards.
Governance
Who approves, reviews and stays accountable for outcomes?
Token visibility and closed-loop routing are described as in progress, so cost governance is incomplete even in this Cisco-on-Cisco story; agencies should require token/cost attribution before scaling agents.
Security and privacy
What data, permissions and controls need testing?
Cisco embeds AI Defense for prompt/response inspection and plans Hypershield segmentation plus Live Protect virtual patching; claims are vendor architecture, not validated against a SLED threat model.
Accessibility and workforce
Who is affected, and what skills or accommodations follow?
MyAgent at workforce scale implies change management and accessibility of agent interfaces for 90k employees; no ADA or public-sector workforce evidence is provided.
Procurement
What should contracts, pricing and exit terms secure?
Stack spans compute, networking, AI Defense, Live Protect and Splunk entitlements; SLED buyers should map to existing Cisco/Splunk contracts rather than assume a single SKU AI factory.
Operating model
Which teams own the service once it runs?
Shifts from project-owned GPU islands to shared multi-tenant ops with enterprise SLA; requires FinOps for tokens and platform ownership across network, security and AI teams.
Publication history
- 2026-11-06Cisco · Issue 095 resources
Stable resource ID: cisco-enterprise-ai-factory-gpu-utilization-onprem-security-20261106