Building a fully scalable fiber infrastructure

Immagine

For most of the last decade the data center was, in essence, a cloud machine. The facilities built through 2023 and 2024 were designed to run the applications we now take for granted, from email and collaboration tools to streaming and enterprise software. Then, almost overnight, artificial intelligence changed the brief. Once large language models proved how powerful they could be, investment poured into a new class of facility built specifically for AI, and these sites bear little resemblance to what came before.

The difference is not incremental. An AI data center consumes in the region of ten times the fiber of a comparable cloud facility. Where a cloud campus might be measured in tens of megawatts, the AI sites now being planned are measured in hundreds of megawatts. A national AI data center build program will require tens of millions of fiber kilometres. Multiply that across a national build program and the figures quickly become difficult to picture.

That trajectory raises an awkward question for anyone specifying infrastructure now. If demand is climbing this steeply, is there any point installing fiber today that will simply be redundant by the time the next generation of hardware arrives?

Why the fiber is rarely the bottleneck

Immagine

The reassuring answer is that the fiber itself is seldom the limiting factor. Optical fiber is an exceptional medium with enormous inherent capacity, and most networks draw on only a fraction of it. The constraint sits in the combination of the fiber and the electronics at either end.

Most systems today run on a single wavelength, so there is considerable headroom to open up further wavelengths and lift the data carried across a single strand, all without disturbing the cable already in the ground.

This is central to any honest conversation about future-proofing. Glass does not wear out. Well-designed cable is proven for at least twenty-five years of service, and some designs have been tested to demonstrate a working life of fifty. Performance does not drift over time. When networks do fail, the cause is almost always a connector that has been contaminated or physically damaged, or external damage to the cable, rather than any deterioration in the fiber. In practice, fiber that is installed well can stay in place for decades while its capacity is upgraded again and again through the equipment around it.

Designing for the way AI scales

Immagine

The harder design challenge is density. AI workloads run on tightly interconnected processors that all need to communicate with one another, and that generates a volume of connections the industry has not previously had to manage. Inside a single AI rack there can be more than five thousand connections, while the cabling running between racks is a small fraction of that. As the data moving between processors grows, connections traditionally made in copper are shifting to fiber, and once fiber enters the rack the quantity required rises sharply.

It helps to picture the problem in three layers. Scaling up means connecting processors within a rack, the domain now moving from copper towards fiber and, increasingly, towards optics built directly into the chip.

Scaling out connects racks and rows into a single compute cluster. Scaling across links buildings, campuses and geographically separated sites. Each layer pulls in a different direction

Over very short distances it is more cost-effective to use more fibers and carry less data on each, whereas over long distances the economics reverse and the priority becomes pushing more data down fewer strands.

This is why the right answer is not a single product but a portfolio matched to the application. High fiber count cables, now reaching several thousand fibers in one cable, address the density problem between racks and across campuses. Reduced-diameter fibers, down to 160 microns, pack more capacity into the same physical space. Multi-core fiber, which places several light-carrying cores inside one strand, offers a further route to density. And hollow core fiber, which guides light through air rather than glass, cuts latency by roughly a third and allows operators to site facilities further apart, closer to cheaper power and land, without paying a performance penalty.

Building for what comes next

None of this requires operators to predict the future precisely. The sensible approach is to install high-quality fiber with capacity to spare, route and protect it carefully, and adopt a pay-as-you-grow model that adds fibers and upgrades electronics as demand materializes. Chosen and installed properly, the fiber laid today should still be earning its place long after this generation of AI hardware has been retired. The task for the sector is less about guessing where AI goes next, and more about building infrastructure with the flexibility to follow it.