AI’s next infrastructure bottleneck may be hiding inside the fibre network

Artificial intelligence has made the data centre a technology story in its own right. Much of the discussion has understandably centred on processors, electricity consumption and cooling. Yet there is another physical constraint becoming increasingly important: moving enormous quantities of data between all of that computing hardware. Modern AI infrastructure depends on optical fibre, and the requirements become more demanding as networks grow in capacity and density.

The scale of this transition is apparent in the move towards faster Ethernet. IEEE standards now cover 400G and 800G Ethernet, but increasing the headline transmission speed does not solve the physical engineering problem. Reach, wavelength, fibre count, connector type, error correction and interface specifications all have to work together. This means the AI infrastructure race extends considerably further down the supply chain than the GPUs sitting inside a server rack.

The AI fibre boom

One indicator comes from TrendForce. According to information provided to Digital Journal by STL, citing TrendForce, global optical-fibre demand is projected to reach 577 million fibre-kilometres in 2026, with AI data-centre expansion and investment in high-capacity networks among the factors driving demand.  That forecast matters because AI computing creates a networking problem as well as a computing problem.

Large clusters must exchange data between accelerators, storage and other parts of the computing environment. Consequently, operators are confronting not simply the question of how much fibre they need, but how much connectivity can practically be fitted into a finite physical space.

This is leading to renewed scrutiny of the entire optical chain: glass manufacturing, fibre production, cable construction, ‘connectorisation’ and testing. A defect or inappropriate design choice early in this chain does not simply disappear when the cable enters a data centre. Physical characteristics influence how optical connectivity can be routed, installed and maintained.

AI infrastructure presents a simple but awkward engineering equation: more computing capacity requires more connectivity, yet data centres do not have unlimited room for cabling. At extreme fibre counts, pathways themselves become a constraint. Information supplied by STL, for example, describes pre-terminated trunks ranging from 48 fibres to 6,912 fibres. Its indoor assemblies cover 48 to 576 fibres, while indoor/outdoor variants extend from 144 to 6,912.

The figures illustrate how different contemporary fibre density can be from the relatively simple patching arrangements associated with traditional enterprise networks. This makes physical cable dimensions consequential. The same STL portfolio specifies a 48-fibre indoor trunk with a diameter of 5.8 mm and a 576-fibre version measuring 12.2 mm. At the other end of its range, the 6,912-fibre indoor/outdoor cable has a stated outer diameter of 32 mm.

The underlying engineering objective is to increase connectivity without allowing trays, ducts and conduits to become unmanageable. There is also the question of bending. Routing thousands of fibre connections through densely populated infrastructure inevitably requires cables to negotiate restricted spaces. The choice of fibre therefore becomes part of the density calculation, rather than an isolated purchasing decision. STL, for instance, specifies ITU-T G.657.A2 bend-insensitive single-mode fibre as one option for space-constrained routing in its pre-terminated assemblies.

Increasing Ethernet speeds further complicates matters. Moving a network from 100G to 400G or 800G cannot safely be regarded as a straightforward switch replacement. Optical specifications at both ends of a connection must match. Even familiar terminology may disguise important differences. Something described loosely as “LR4”, for example, does not identify a universally interchangeable interface across Ethernet generations.

For data-centre operators, this places increasing value on considering fibre, cable and connector as parts of one engineered system. High-density infrastructure also has to accommodate different connector architectures. STL says its latest assemblies can be factory-terminated using MMC, MTP/MPO and LC formats, with Base-8, Base-12, Base-16 and Base-24 configurations.

The broader implication is that “fibre” should not be treated as a single commodity. Fibre type, cable geometry, connector format and the optical interface supported by the equipment all affect whether an installation meets its intended purpose.

Moving work out of the data hall

Another development is the increasing use of pre-terminated fibre assemblies. Traditionally, some optical installation work can be completed at the deployment site. At AI-data-centre scale, however, multiplying field operations across very large numbers of connections increases the amount of work that must occur during deployment.

The alternative is to move termination and testing into the controlled factory environment.

STL says the assemblies in its recently expanded Neuralis portfolio are factory terminated and 100 per cent tested before deployment. The company positions this as a means of reducing field-installation work while supporting higher-density deployment. [PRESS

This should be understood as a manufacturer’s claim rather than evidence that pre-termination eliminates deployment risk. Installation still requires correct planning, routing, handling and compatibility between the components at each end. Nevertheless, changing where work occurs is important. As data-centre builds become larger and more repetitive, converting site activity into factory-controlled assembly and testing offers operators a different approach to managing installation complexity.

Rahul Puri, CEO of the Optical Networking Business at STL, characterises the issue as one of integrating optical connectivity “from glass to connection”. The proposition is that decisions made across fibre manufacture, cabling, connectorisation and testing increasingly need to be viewed together when infrastructure is built for AI scale.

There is a sound wider context for looking beyond computing hardware. AI data centres are already placing pressure on electricity infrastructure, and data-centre expansion is prompting significant debate over power and other physical resources. Recent analysis reported by Digital Journal demonstrates the extraordinary scale of the energy infrastructure now being contemplated alongside AI expansion. Networking presents a different constraint, but the lesson is similar: AI ultimately depends on physical systems. Processors may attract the headlines, but GPUs that cannot exchange information efficiently cannot deliver their intended performance as part of a cluster. As data-centre networks progress through 400G and 800G architectures, optical infrastructure therefore moves from a background utility towards a central engineering consideration.