When Photonics Becomes a Network Architecture Constraint in AI Data Centers

Abstract AI training, inference and agentic workloads are placing new stresses on data center network architectures. Large accelerator clusters generate sustained, highly synchronized east–west traffic that challenges traditional assumptions around bandwidth density, electrical reach, optics power, and fault isolation. As these environments scale, limitations emerge that cannot be addressed through incremental upgrades to existing designs. This talk examines how photonics is now becoming a binding architectural constraint in AI data center networks rather than a transparent transport layer. It outlines where current architectures encounter scaling limits, including faceplate density, thermal and power budgets for pluggable optics, and the operational complexity of rapidly expanding fiber plants. The discussion focuses on architectural impacts rather than specific products. The presentation then reviews several optical architectural approaches operators are actively evaluating, including co-packaged optics, external laser source models, and optical circuit switching as a complement to packet-switched fabrics. For each, it highlights the problems these approaches aim to address, the new constraints they introduce, and where they may—or may not—fit within real-world operational environments. The goal is to provide network operators with a practical framework for evaluating how AI-driven traffic growth affects data center network design in current and near-term deployments. Attendees will gain clarity on where existing assumptions hold, where they break down, and which architectural decisions warrant early consideration. Christian Urricariet: Christian Urricariet is Senior Director of Product and Strategic Marketing at Lumentum, with over 25 years of experience in photonics and optical networking, primarily in product management roles. His work focuses on the development and deployment of optical interconnect technologies for AI and cloud hyperscale data centers and large-scale network infrastructures. Prior to Lumentum, he spent 21 years at Finisar and 3.5 years at Intel, contributing to multiple generations of optical connectivity solutions for cloud and data center networks. His experience spans both component- and system-level perspectives, with a focus on how these technologies integrate into real-world network architectures. He holds a degree in Electrical Engineering with a specialization in optoelectronics and has presented at several NANOG meetings. https://nanog.org/events/nanog-97/con...