There was a time when building a data center felt a bit like buying a suit off the rack. You picked a standard size, made a few minor adjustments, and trusted that what fit most people would fit you well enough. It wasn’t perfect, but it was predictable and efficient.

Today’s AI era, however, has untethered the data center from that one-size-fits-all world, and pushed it into a new design regime, one in which infrastructure must be shaped much more deliberately around workload and business need – exposing just how poorly legacy assumptions now fit.

As AI deployment models diversify, infrastructure can no longer be standardized around a single archetype – workloads are too large, too varied, and too strategically important.

These shifting operational priorities are driving fundamentally different infrastructure choices, pushing toward more tailored network architectures and fiber plant designs, making it clear that evolving market demands are now tied directly to physical-layer design.

When the pattern no longer matches the fabric

Traditional network and fiber plant designs for legacy data center workloads have come to form a standardized pattern – made up of general-purpose fabrics, pre-defined pathways, centralized patching, and conservative fiber counts.

Even as link speeds evolved incrementally – from 10G to 40G to 100G – most facilities did not have to accommodate the combination of extreme bandwidth, density, low latency, and rapid scaling that AI clusters now demand.

Having spent more than 40 years specializing in computing and networking, Dr. Alan Keizer, special technical advisor at AFL, has seen just how much that model has shifted. Hyperscale environments are now built around racks drawing more than 100kW, and fabric links run at 400G and 800G.

The pace of change has been astounding. As Keizer puts it:

“What’s happened in the last five years exceeds the rate of change that I’ve seen in any decade before that – it’s an amazingly dynamic time period. The advent of AI has pushed people and organizations to dive in, push harder, do new things, and combine technologies in new ways. It’s an exciting place to be.”

The risk is that teams continue to apply familiar patterns long after the assumptions behind them have broken down. In AI infrastructure, what worked last year may already be suboptimal. That can lead to underperforming systems, inefficient use of space and power, slower deployment, and constrained upgrade options. As Keizer notes:

“What they did last year is no longer the smart thing to do. They have learned a lot more, and different tools are available. So we see the big hyperscalers producing maybe several large AI data centers in a given model, and then the next generation is different.”

Part of the challenge is that ‘AI’ isn’t one thing. It’s a catch-all term that covers everything from training to fine-tuning to inference – each demanding fundamentally different design priorities.

There is no single template for building an AI data center, and the network fabric cannot simply follow a legacy standardized pattern. Trying to force one is like asking everyone to wear the same suit, whether they’re running a marathon or heading to a black-tie dinner.

“AI is not one workload, and the fiber plant can no longer be designed as though it were,” Keizer affirms.

Training environments, for instance, are centralized, resource-intensive facilities that pack large numbers of accelerators into tightly interconnected clusters. Latency and bandwidth directly shape how fast models can be developed, and the fiber plant must support extreme density and near-perfect reliability – any small disruption can strand very expensive compute resources.

Inference environments can look very different. Some are highly distributed and latency-sensitive, especially where user response time is critical. Others are large centralized clusters optimized for throughput, utilization, and cost per token. In both cases, the design priorities differ from training, and the physical infrastructure must reflect that.

These fundamental differences introduce a tradeoff in how these environments are built. Larger, more centralized clusters may reduce cost per token through scale, but at the expense of added latency as requests travel further; smaller, more distributed clusters improve responsiveness but at higher operational costs due to duplication and lower utilization.

Either way, neither scenario can be served well by a one-size-fits-all approach.

Managing fiber count and shrinkage

In the short span between the rise of popularized AI applications like ChatGPT and today, much of the industry’s focus has been on building facilities to train and scale these models. The challenge is now translating those shifting AI requirements into practical physical-layer decisions – cabling, connectivity, and overall infrastructure design.

This is a substantial undertaking, considering it has required rethinking the fundamental building blocks of traditional computing to accommodate vast numbers of parallel processing accelerators. These accelerators, arranged in tightly coupled racks and pods, depend on extensive interconnectivity to operate in unison, with high-speed communication flowing across complex fabric topologies.

The physical consequences of these stark changes include surging fiber counts, shrinking available space, and rising thermal pressure – all of which put pressure on cable routing, connector density, service access, and error avoidance.

In that context, today’s small form factor and very small form factor connectors are part of a broader response to the need for much higher connectivity density within the same or smaller physical envelope.

What’s needed, according to Keizer, is a shift toward tailored infrastructure – designs that start with the specific requirements of the workload and build outward from there. This doesn’t mean abandoning standardization altogether; there is still clear value in repeatability and modularity, but these need to be applied within a design framework that begins with the workload, the topology, and the operator’s growth path.

“The first thing we focus on is meeting demand, but going beyond that, it’s providing more and more capability in terms of cable density and connectivity at every layer in the network,” he says.

At the physical layer, that means treating connectivity as a design system rather than a fixed layout. AI fabrics create many-to-many connectivity requirements that are far more demanding than traditional point-to-point builds. Pre-terminated high-fiber-count assemblies can accelerate deployment and reduce installation error, but only if the underlying topology, routing, polarity, and service strategy have been thought through in advance.

Cooling also becomes part of the tailoring process. The adoption of liquid cooling and advanced air management strategies unlocks higher rack densities and allows cabling and connectivity to be engineered more efficiently within tighter physical constraints.

Even ergonomics – which is often overlooked – becomes a design advantage. Where thousands of fiber connections coexist in a single rack, well-designed patch panels and cable management systems improve accessibility, simplify maintenance, and reduce the risk of operational error.

This is where specialist engagement matters. The question is no longer just how to connect point A to point B. It is how to build a fiber plant that supports scale, serviceability, installation speed, and future migration without creating operational drag.

Tailoring for purpose, not convenience

One of the most challenging aspects of AI infrastructure is that, while it must be designed for current workloads, it must also work for future ones – workloads that are inherently difficult to predict.

Today’s focus on large language models may well give way to new paradigms tomorrow – agentic AI, real-time multimodal processing, and Edge-based intelligence embedded directly into physical environments. Each will bring its own set of requirements, reshaping the infrastructure landscape once again.

This only reinforces the need for adaptability, where systems can accommodate new technologies and shifting priorities without requiring complete redesigns. As Keizer frames it:

“AI infrastructure will be a rich palette with many colors and textures, and the designer not only can, but really has an obligation to assemble it in a way that reflects the specific needs of the operator.

What are the workloads? What are the business and operational objectives? What is the outlook over time? We are building a very rich portfolio for the connectivity infrastructure designer to draw from, and at the same time, we’re evolving new standards of practice.”

Because hyperscalers and neocloud workloads are so nuanced, and the models being trained so individualized, it is unlikely that a standard plug-and-play architecture will meet their holistic needs. It’s precisely that specificity of operators’ requirements that makes an adaptable and practical approach so crucial.

“If you’re a big, well-established hyperscaler or AI lab with significant resources, you almost certainly want to optimize the solution for your specific problem.”

Instead of a single dominant architecture, we are seeing a proliferation of specialized approaches, each optimized for different use cases – and different situations call for different fits. In this context, connectivity has to be considered as part of the site’s ongoing evolution.

Against this backdrop, network topology can be analyzed and optimized in line with the specific workload requirements of the operator, with specialist providers such as AFL working closely with operators early in the design phase, bringing expertise in structuring fiber connectivity at every level.

A wardrobe, not a uniform

Looking ahead, Dr. Keizer envisions AI infrastructure as a broad design space, where operators should thoughtfully select from a diverse array of technologies and architectures to meet evolving requirements. This is the emerging reality of AI infrastructure.

Building AI infrastructure is becoming a craft in its own right. It requires an understanding of both the materials and the forms they must take in different contexts. In that sense, a well-tailored system does not just function; it performs in a way that aligns with its purpose, its workload, and its environment.

The key point is simple: in AI, the physical layer is no longer passive plumbing. It is part of the performance envelope.

To learn how AFL helps data center operators design, deploy, and scale high-density optical connectivity for AI and hyperscale environments, visit aflhyperscale.com.