Cisco IT AI factory triples GPU utilization to ~50%

Cisco's account traces a shift from dedicated, project-specific AI clusters to a shared, multi-tenant setup as AI moved from prototypes to mission-critical operations. The stated catalyst was the global deployment of MyAgent across its 90,000-person workforce and scaled capacity for Cisco IQ, which hosts live production workloads for thousands of enterprise accounts. That surge in agentic workflows exposed four barriers Cisco lists: stranded capacity with GPU utilization at 10%-15%, unpredictable token costs, expanding security attack surfaces, and thermal and power density ceilings.

The company's response combined a high-bandwidth pooled network fabric, heterogeneous compute using Cisco UCS-X servers with a mix of NVIDIA and AMD GPUs, rear-door heat exchangers, and natively embedded security through Cisco AI Defense and Hypershield. Splunk Observability Cloud, Splunk Cloud Platform, and Splunk IT Service Intelligence form the observability layer. Cisco says it is expanding that layer to Agent Observability and Tokenomics.

Looking ahead, Cisco IT plans to shift toward inference as demand outpaces training by 60%, with dynamic resource pooling, automated GPU reclaims, and adaptive networking. It also cites agent-generated traffic projected to be 450% higher per task than human traffic as a reason to move to an AgenticOps model. The results are a single company's internal deployment, so the figures may not translate directly to other enterprises.