1. Introduction

A two-site data center interconnect carrying a handful of 400G or 800G flows and a tier-1 carrier metro grooming Ethernet, mobile backhaul, and wholesale wavelengths look like the same optical problem from a distance. They are not. The first wants the shortest possible layer stack and a coherent pluggable in a router port; the second needs a packet-optical or Optical Transport Network (OTN) layer to aggregate, monitor, and isolate services before light hits the fiber. Choosing the wrong one wastes capital in one direction and strands operability in the other. This article is about making that choice on physics and economics rather than habit.

The pressure forcing the question is machine traffic. Optical transport was engineered for decades around predictable, peak-hour residential broadband and north-south enterprise cloud flows. Trillion-parameter model training and distributed high-performance computing instead generate dense east-west collectives between synchronized graphics processing unit (GPU) clusters, and the racks driving them draw between 100 kW and 600 kW each. When a single building can no longer power a training job, operators split it across buildings or metros, and the optical layer has to merge data center fabric behavior with wide-area reach. A new operator class — the "neocloud," a pure-play accelerated-compute utility — has grown up entirely inside this constraint.

Data center interconnect (DCI) is engineered for a different objective than the telecom access-to-core hierarchy. Where a service provider aggregates millions of small connections and funnels them toward the internet, DCI moves colossal datasets at high speed between a small number of large endpoints. That shifts the design center toward ultra-high capacity per fiber, deterministic latency, and reliability budgets that treat optics as a first-order failure source rather than an afterthought.

What follows builds from the layer model out: the distance classes that define DCI, the four networks inside an AI cluster, the physical-layer architectures that carry them, the link-budget math that decides which architecture closes, and the design, implementation, monitoring, and troubleshooting practice that turns a feasible link into a deployed one. The numbers are attributed to their source class throughout — measured, standardized, or vendor-claimed — because in this domain the difference between those three is the difference between engineering and marketing.

2. Foundational concepts

An AI cluster is four distinct networks stacked on top of each other, and only the outermost is DCI. Inside a rack, the scale-up fabric connects GPUs over copper — an NVLink-class domain such as a 72-GPU rack reaching roughly 130 TB/s aggregate bisection over thousands of copper cables, where copper saves on the order of 20 kW per rack versus optics at that distance. The scale-out fabric is the optical back-end: one transceiver per GPU at 400G or 800G over InfiniBand or RoCEv2, optical because 800G PAM4 over copper degrades beyond about one meter. A front-end Ethernet network handles data loading and checkpointing. The fourth network, scale-across, is coherent Dense Wavelength Division Multiplexing (DWDM) DCI between sites, stitching geographically separate GPU fabrics into one logical machine.

Scale-across is where the optical engineer lives, and its binding constraint is not optical reach. Coherent pluggables reach roughly 2,000 km at reduced rates, but distributed training generates synchronized all-reduce and all-gather collectives that stall on the slowest replica, so every additional 100 km of fiber adds about 0.5 ms of one-way propagation. Published demonstrations have kept distributed training within roughly 1,000 km even though the optics could go further. The light is not the limit; the speed of light is. This single fact reorders the whole design: latency is the ceiling, and reach margin beyond the latency budget is wasted.