This paper presents a validated reference architecture combining NVIDIA Dynamo for distributed KV cache management, IBM Storage Scale Erasure Coding Edition (ECE) as a high-performance shared storage tier, Supermicro Petascale servers, and NVIDIA Spectrum-X Ethernet to address the critical infrastructure challenge of key-value (KV) cache management in large-scale generative AI and agentic AI inference deployments. The paper provides deployment sizing options ranging from small proof-of-concept environments to large enterprise AI factories, demonstrating that shared storage-based KV cache infrastructure provides a scalable and cost-efficient path to large-scale inference deployment.