Technology

As AI needs evolved from isolated projects to mission-critical enterprise-wide platforms, Cisco IT needed to address physical, economic, and operational barriers, including:

• Stranded capacity & low GPU utilization: Fragmented, project-specific clusters trapped compute in isolated silos, stranding GPUs at low utilization.
• Unpredictable cloud costs & workload placement: Relying solely on closed-source cloud models became cost-prohibitive as AI scaled enterprise-wide. Without token-level visibility, forecasting expenses and proving business ROI was nearly impossible.
• Expanding security attack surfaces: Scaling multi-tenant autonomous agents introduced new threats, from prompt injection and intellectual property loss to lateral network threats moving across data centers undetected.
• Thermal & power density constraints: Traditional air-cooled facilities hit physical heat limits long before they could energize high-density GPU clusters — and waiting years to build costly new data center space would stall AI innovation.
  
  
  

Cisco IT modernized its infrastructure into a full-stack, multi-tenant AI factory that embeds compute, networking, security, and observability across global data centers using the following model:

• High-bandwidth pooled network fabric: Programmable Cisco Silicon One and Cisco switches eliminate data bottlenecks and GPU packet drops, pooling network capacity to dynamically route multi-tenant workloads and future-proof the fabric.
• Heterogeneous compute architecture: Cisco UCS X-Series servers with a mix of NVIDIA and AMD GPUs provide flexibility to optimize token economics and adapt to evolving hardware supply. 
• High-density rack cooling: Rear-door heat exchangers retrofitted onto existing cabinets absorb heat at the source, doubling compute density to two full AI nodes per rack within the existing data center footprint.
• Natively embedded AI security: Cisco AI Defense inspects runtime prompts and responses, while Live Protect delivers real-time vulnerability shielding — securing the network switches and the data traversing them. Cisco Hypershield is architected for future fabric-wide segmentation.
• Real-time observability and tokenomics: Splunk Observability Cloud and Splunk Cloud Platform unify infrastructure metrics, traces, and logs to accelerate Mean Time to Resolution and enable closed-loop operations, with expansion to Agent Observability and Tokenomics in progress.