From Telecom Consumable to AI Compute Vascular Network

Introduction#

For roughly two decades, optical fiber behaved like a commodity consumable. Its demand curve tracked telecom carrier spending and general-purpose cloud build-outs, prices were set by preform capacity and Chinese vendor competition, and the buyer's question was almost always the same: how many strands per dollar. Fiber was treated as passive plumbing, purchased late in a project, and rarely discussed at the board level.

That framing has broken down. In an AI training cluster, the interconnect is not plumbing around the computer, it is part of the computer. A model's training step cannot complete until every accelerator finishes its share of work, and that synchronization depends on fibers carrying tens of terabytes per second across the fabric with microsecond-scale latency. Remove the fabric and the GPUs are simply expensive space heaters.

This guide explains the underlying logic of that transition, quantifies how much fiber a single large GPU cluster actually consumes, and provides a practical method for building a bottom-up fiber demand model that can be validated against public capital expenditure data.

Why AI Clusters Break the Old Fiber Math#

Traditional enterprise and cloud data centers are dominated by north-south traffic: requests arrive from users, servers respond, and the network is engineered around oversubscription ratios that often sit between 3:1 and 10:1. AI workloads invert this. Training and inference on large models are dominated by east-west traffic, where every accelerator talks to every other accelerator through collective operations such as all-reduce and all-to-all. That traffic pattern requires a fabric with low oversubscription, frequently approaching 1:1, because a congested link stalls the entire job rather than a single user session.

The second break is the ratio of optics to compute. A conventional server may host two to four processors and expose two 25G or 100G network ports. An AI node typically pairs every accelerator with its own high-speed network interface, so a node with eight GPUs may present eight 400G or 800G backend ports. The number of optical endpoints per unit of compute has therefore multiplied by a factor of four to eight before any consideration of cluster size.

The third break is physical. AI clusters are dense and compact, usually spanning a hall or a campus rather than a metro region, with link reaches of 50 to 500 meters. This favors parallel multimode optics in some tiers and short-reach single-mode in others, and it concentrates an enormous number of connector pairs into a small physical space. The constraint shifts from fiber kilometers to fiber strands, connector density, and cable management inside the rack.

Sizing Fiber Consumption for a 10,000-GPU Cluster#

Consider a reference cluster of 10,000 GPUs arranged in 1,250 nodes of eight accelerators each, using a rail-optimized backend fabric where each GPU's network interface connects to a different leaf switch rail. With 400G interfaces, the backend alone exposes 10,000 ports and roughly 4 petabits per second of aggregate interface bandwidth. Using 400G-DR4 optics, each duplex link