Energy-Efficient Optical Interconnect Architectures for AI Data Centers

Abstract

The rapid expansion of artificial intelligence data centers has transformed interconnect design from a supporting engineering concern into a central determinant of energy efficiency, scalability, and service fairness. As large models scale in parameters, training parallelism, and inference concurrency, electrical packet-switched networks face growing pressure from bandwidth-distance limitations, thermal density, and power delivery constraints. Optical interconnect architectures offer a promising alternative through wavelength parallelism, low propagation loss, and reduced switching energy per delivered bit, but their system-level benefits depend on topology, switching granularity, co-packaging strategies, control-plane maturity, and deployment governance. This paper examines energy-efficient optical interconnect architectures for AI data centers from an interdisciplinary systems perspective. It analyzes silicon photonics, wavelength-switched fabrics, optical circuit switching, hollow core fiber, and co-packaged optics as elements of a broader socio-technical infrastructure. The discussion emphasizes structural trade-offs among bandwidth, latency, radix, reliability, reconfiguration overhead, and sustainability. Particular attention is given to governance mechanisms that shape energy accounting, multi-tenant fairness, carbon-aware scheduling, and infrastructure retrofitting. The paper argues that optical interconnects should be evaluated not merely by device-level efficiency metrics but by lifecycle energy, operational resilience, equity of access, and policy accountability. It concludes that AI data centers require co-designed optical architectures, open interfaces, and transparent energy governance to realize sustainable performance at scale.