OFC 2026 Series

OFC 2026 Deep Dive: The Optical Interconnect Revolution in the AI Era

701 papers. 6 deep-dive articles. From optical switching to CPO packaging, materials war to AI-driven operations — OFC 2026 marks the inflection point where optical networking becomes AI infrastructure.

6 Articles 701 Papers AI Datacenter Optical Interconnect
OFC 2026 AI 光互连
701
Papers
50.3%
AllReduce reduction
600km
Cross-DC training
1Tbps/λ
BTO per-λ rate
0.11
dB/km HCF loss
0.894
AI ops F1 score

AI Cluster Optical Interconnect Layers

SVG · scale-up / scale-out / scale-across
AI 集群光互连三层结构图

Three layers: intra-chassis (Scale-Up), inter-rack (Scale-Out), cross-DC (Scale-Across) — each with distinct latency, bandwidth, and power requirements.

01

AI Clusters Are Rewriting the Foundation of Optical Networking

Series overview. The optical networking narrative shifts from "bandwidth carrier" to "AI system architecture." Three mainlines: optical switching, CPO packaging, materials war.

OCSCPOTFLNAllReduce
02

Optical Circuit Switches Are Becoming the Nervous System of GPU Clusters

OCS reconfiguration drops below 2μs, enabling AI training phase coupling. O/E hybrid cuts AllReduce by 50.3%. 600km cross-DC LLaMA2-70B training demonstrated.

OCSAllReduceLLM 训练W4H.4
03

CPO Enters Engineering Deep Waters: The "Last Centimeter" Battle

CPO's real challenges: external laser power, glass substrate coupling yield, and thermal management. Photonic Chiplets targeting 204.8 Tb/s.

CPO玻璃基板Photonic Chiplets散热
04

400G/lane Materials War: Silicon, TFLN, or BTO — Who Wins?

TFLN driver-less operation (~1V drives 1.2Tbps). BTO monolithically integrated for 1.6T net. Scene-based coexistence, not winner-takes-all.

TFLNBTO硅 MZMDriver-less
05

Hollow-Core Fiber: Faster Than You Think (By More Than a Little)

HCF propagates light in air, cutting latency 31-47% vs SMF. Loss record refreshed to 0.11 dB/km. Real demand from AI cross-cluster training and HFT.

HCF空芯光纤低时延0.11 dB/km
06

When AI Manages Optical Networks: Optics GPT and Self-Healing Infrastructure

Optics GPT: first vertically pre-trained model for optical networking. Baidu production data validates AI ops at F1=0.894. Networks shift from rule-based to model-based.

Optics GPTAI 运维数字孪生自愈网络