Become a TechRadar Insider
Artificial Intelligence (AI) and very large (hyperscale) datacenters are not new technologies. What is new is how they are both now commingled and supporting each other in an unprecedented growth spurt.
The datacenter (DC) industry is currently going through a period of fundamental structural change, driven mainly by the rapid transition from traditional cloud computing to AI computing.
Foundational to this transformation and a key enabler is fiber and the extreme fiber densification that is occurring within AI DCs.
Chief Business Officer, Hyperscaler Data Centers, STL.
This shift is fundamentally reimagining the "Physical Layer" from a passive utility into a strategic asset. While traditional IT infrastructure relied on hierarchical, north-south traffic flows, AI "factories" demand a massive, east-west parallel architecture where thousands of GPUs synchronize in real-time.
This architectural pivot creates a literal space-crunch; we are attempting to squeeze terabits of intelligence through physical pathways—conduits, trays, and ducts - that were never designed for such volume.
Consequently, the industry is moving past the "standard" 250-micron fiber toward ultra-thin, high-density 160-micron solutions. It is no longer just about the speed of light, but about the density of glass. To scale intelligence, we must first master the physics of the fiber that carries it.
The Density Dilemma
The difference in how AI workloads are processed compared to how traditional cloud workloads are processed is the basic driver of this increase in fiber volume. In a traditional cloud environment, tasks like web hosting, database management, and file storage are typically processed by Central Processing Units (CPUs) in a hierarchical fashion via access, aggregation and core layers.
Here, data mostly moves between the end-user, who may be located anywhere outside the DC, and the server within the DC. This type of traffic flow is described as ‘North-South’ and requires robust external gateways but relatively modest internal interconnects within the DC.
However, these traditional network architectures are ill-suited for the ‘all-to-all’ communication required by Graphics Processing Unit (GPU) or Accelerator clusters in AI DCs because they introduce latency and bandwidth bottlenecks at oversubscribed aggregation points; this cannot be tolerated by AI compute stacks.
AI factories use iterative, compute-heavy processes where thousands of GPUs must work in tandem as a single logical entity. This creates a huge amount of internal traffic because of the enormous parallel processing needs across the interconnected GPUs.
In such environments, the network must support the constant synchronization of the AI model parameters and the timely exchange of vast amounts of mathematical modelling data, creating a huge volume of internal traffic. This type of traffic flow is referred to as ‘East-West’ traffic and it now accounts for most of the movement of data within AI DCs.
The architecture to support large East-West traffic is very fiber dense because the number of East-West connections scales quadratically with the number of nodes (GPUs/Accelerators).
Why ‘Fiber-First’ is the New Standard
The scale of East-West traffic is orders of magnitude larger than the traffic moving North-South. Any link error or packet loss in this network is catastrophic; for instance, it can force the entire training batch to be restarted, leading to significant financial losses and delays in model development. Only fiber can meet the bandwidth and speed requirements needed for AI processing.
AI DC operators are also deploying network fabrics such as InfiniBand or high-performance Ethernet and these require lots more high-performance fiber than traditional CPU-based networks to meet AI performance metrics.
To build a data center whose infrastructure will not be obsoleted at the next step-increase in speed (say from 800G to 1.6T), we have to adopt a fiber-first approach at the network design stage. Scalability is now a basic DC necessity.
It’s not ‘Speed of light’, It’s now ‘Speed to Light’
The pressure to deploy AI-ready capacity quickly is unending and high-density, pre-terminated fiber solutions are really the only deployment option available when building AI DCs. Instead of spending weeks on-site splicing multi-fiber elements, pre-terminated fiber solutions offer advantages such as:
Plug and Play: Significantly reducing deployment timelines (from weeks to days).
Reduction in Human Error: Pre-terminated solutions are factory assembled and tested.
Simplification: Make it easier to upgrade or swap components as technology evolves and cable management is more controlled.
Pre-terminated solutions are a key enabler of the speedy deployment of fiber networks, allowing for quicker commercial activation of AI DCs.
The Shift from Compute-Led to an Infrastructure-Led Design
Optimised fiber connectivity in an AI DC is no longer an afterthought; it’s now an important expectation within best-practice AI DC design. We are moving away from a world where we buy servers or AI compute pods and then figure out how to connect them.
Instead, we are now designing DC environments where fiber connectivity is key to determining the efficiency, speed of deployment, scalability, resilience, operational agility, and long-term readiness of the DC facility.
This shift is especially important as AI workloads demand denser interconnects, faster data movement, and lower latency across the compute environment. A well-planned fiber architecture helps reduce future rework, supports smoother upgrades, and ensures the DC can scale with evolving AI compute requirements.
We feature list the best Linux server distros.
This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.
The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit