AI accelerators (e.g., GPUs, ASICs) evolve over relatively short technology cycles, while grey-area infrastructure (e.g. power, cooling, fiber cabling, pathways, and connectivity) is expected to remain in service for much longer.

In practice, this means that what works for one generation of AI infrastructure may look very different by the next, making flexibility at the physical layer increasingly important.

I see this as a central consideration for organizations planning for the next phase of AI capacity. The broader point is that there is no single path to scaling AI infrastructure.

Designing fiber infrastructure around current AI workload demands can quickly create constraints as hardware refresh cycles bring new, higher-performance requirements. For me, this makes adaptability a fundamental part of AI infrastructure planning.

Different paths to AI capacity

Neoclouds are one example of how AI capacity is being deployed differently. These providers are focused on scaling specialized accelerated computing to meet growing AI demand. That means the infrastructure supporting these environments needs to look beyond the immediate workload.

Power and cooling are obvious considerations, but fiber capacity, routing, connectivity, and physical cable management also need to be considered as part of the same long-term plan.

Brownfield deployments bring a different set of challenges. Existing data centers can provide a faster route to AI capacity because operators can leverage facilities, power, cooling, and connectivity already in place.

But those facilities may have been designed around much lower fiber densities, and the differences at ground level for planners can be substantial.

For example, a conventional cloud rack might support 15–30 fibers, while a modern, high-performance rack capable of managing complex AI workloads can support in excess of 1,000 fibers, considerably changing physical infrastructure requirements across pathways, patching capacity, routing, and cable management.

DCI adds another layer to the challenge. AI infrastructure can span buildings, campuses, and geographic locations, giving operators more options for distributing capacity while addressing power availability, resilience, and other facility constraints.

However, DCI also raises questions about the availability and capacity of the underlying civil infrastructure. Existing ducts, pathways, and other assets may already be heavily utilized, meaning any increase in interconnection requirements could create demand for additional capacity or new routes.

Latency, optical performance, network architecture, fiber connectivity, and pathway capacity all need to work together.

Then there is the evolution of the optical interface itself. CPO is expected to bring optical connectivity closer to compute and networking components before 2030, changing how fiber is routed and managed around high-density equipment.

VSFF connectivity (which supports dramatically higher fiber counts in the footprint traditionally occupied by larger connector formats) is similarly addressing the need to accommodate significantly more optical connections within limited rack and panel space.

These technologies are important, but they are only part of the bigger picture, as higher connector density does not, by itself, solve the physical challenges created by higher fiber counts. The surrounding infrastructure needs to provide the capacity, accessibility, routing, and manageability required to make those technologies practical at AI scale.

Designing beyond the current hardware cycle

One potential approach is modular data center design, where infrastructure is deployed in repeatable building blocks that can be scaled or adapted as AI requirements evolve.

Modular approaches add capacity in stages and reduce the need to redesign an entire facility as hardware generations change. However, the benefits depend on ensuring that the underlying infrastructure (i.e., power, cooling, fiber, pathways, and connectivity) can support that modular approach.

For the fiber, layer this means defining repeatable pathways, distribution, patching and connectivity blocks with sufficient headroom to allow additional capacity to be introduced without redesigning the surrounding infrastructure.

The physical layer also deserves more attention in AI infrastructure planning. The objective is not to predict exactly what the next AI architecture will look like. That would be unrealistic given the pace of change.

Instead, the objective should be to create enough capacity and flexibility that the physical infrastructure does not become the limiting factor when technology changes.

That means thinking beyond the immediate deployment. Power, cooling, fiber pathways, connectivity systems, patching environments, and cable management all need to accommodate increasing density while retaining the flexibility to support future configurations.

A forward-looking approach can also reduce the disruption associated with future upgrades. If infrastructure is designed with sufficient capacity from the outset, fiber infrastructure planners have more options when new generations of AI equipment arrive, enabling network expansion or deployment reconfiguration without having to fundamentally redesign the physical layer each time the technology changes.

For data center operators, I see long-term infrastructure planning translating into flexibility without unnecessary disruption.

The case for adaptable infrastructure

AI growth will continue to take different forms. Some capacity will be deployed through Neoclouds, some through brownfield facilities, and some through new greenfield builds. Distributed architectures will become increasingly important, while technologies such as CPO and VSFF will influence how optical connectivity is delivered at higher densities.

The common requirement across all of these approaches is a physical infrastructure foundation capable of adapting alongside them. This is why fiber infrastructure needs to be considered a long-term strategic asset, rather than simply an installation requirement for a particular generation of AI hardware. The fiber infrastructure decisions made today will remain in place long after the current AI hardware cycle has been replaced. The goal isn’t to predict the future of AI workload demand perfectly. The goal is to create infrastructure foundations flexible enough to accommodate it.

For more information, read: What Does Sustained AI Growth Mean for Data Center Fiber Infrastructure?