For decades, cabling infrastructure remained on the margins of data centre planning, categorised as passive and treated as a late-stage procurement decision. That positioning is being fundamentally revised. As artificial intelligence (AI) workloads push rack densities towards 100 kW and beyond, east-west traffic within facilities now accounts for roughly 80 per cent of total data movement and the cabling layer has become critical to the performance of data centre infrastructure. In a panel discussion on “Scaling Connectivity and Cabling Infrastructure for Modern Data Centres”, organised by tele.net in association with Belden, Inc., Ankit Agarwal, General Manager – Network Services BU, Yotta Data Services; Nalin Agrawal, Director – Data Centre Advisory, JLL; Saurabh Gupta, Solution Sales Director – Smart Infrastructure Solutions and Data Centres, India and SAARC, Belden, Inc.; Santosh Mahapatro, Enterprise Architect, HPE; Amit Mishra, Head – Credit Strategy and Portfolio Intelligence, Axis Bank; and Jagat Ram, Head – Data Centre Operations, Larsen & Toubro – Vyoma, examined how the industry is rethinking the design, specifications and future-proofing of connectivity infrastructure for modern data centres. Key takeaways from the discussion…

The data centre industry is witnessing one of its most consequential structural shifts. AI, cloud migration and the exponential growth of high-density computing are not simply adding capacity to existing infrastructure frameworks, they are changing the architecture of data centres from the ground up. At the centre of this transformation lies an element that has historically been assigned peripheral status – cabling infrastructure. Long categorised as passive and treated as a later-stage procurement decision, cabling infrastructure is being repositioned as a foundational layer that determines the performance, scalability and longevity of an entire facility.

This repositioning carries real consequences. Infrastructure decisions made at the design stage are extremely difficult and costly to reverse once a facility is operational. Across the industry, from hyperscale operators and colocation providers to enterprises and regulated financial institutions, there is growing consensus that cabling must receive the same level of strategic attention as power supply, cooling infrastructure and compute architecture.

From passive infrastructure to critical
infrastructure

The terminology used to describe cabling within data centre builds is also shifting. For decades, the phrase passive infrastructure grouped cabling alongside elements considered static, secondary and largely invisible. That perception is being actively contested by practitioners who design, operate and invest in data centres. In a facility running graphics processing unit (GPU) clusters connected by InfiniBand networks at 400 gigabits per second, cabling is not passive in any meaningful sense. It is the medium through which the entire value of active infrastructure is realised, and if it fails to function at the required specification, the entire investment is compromised.

The financial picture reflects this changing status. The share of cabling in total data centre build costs, which industry experience placed at approximately 3 per cent through much of the previous decade, is now trending towards 8-10 per cent in AI-focused deployments. Going forward, as GPU density and interconnect requirements continue to rise, this allocation is expected to reach 15-20 per cent. This is not an anomaly driven by a single procurement decision, but it reflects a structural change in the role that connectivity infrastructure plays in overall data centre design and in the capital expenditure that governs it.

This reclassification extends beyond budget share. The development of 400-gigabit InfiniBand interconnects, which now underpin AI cluster design, was made possible only through the active participation of cabling manufacturers in the research and development process. Cabling suppliers are no longer simply fulfilling specifications set by compute and networking vendors. They are co-developing the standards that make next-generation infrastructure possible.

Three forces reshaping data centre decisions

When enterprises and institutions evaluate data centre options today, three forces consistently define their decisions. The first is power density. Deployments that once operated in the range of 5-8 kW per rack are giving way to configurations reaching 30 to 40 kW, and in AI-specific cases, towards 100 kW per rack. The most demanding GPU architectures now require rack designs of 250 kW and beyond, requirements that cascade directly into how the network and cabling infrastructure within those facilities must be specified and managed.

The second force is latency. Physical proximity to network hubs and subsea cable landing stations is increasingly non-negotiable for operators serving latency-sensitive applications. The ability to reach key network interconnects within two milliseconds is becoming a baseline expectation, shifting site selection criteria away from a purely city-centric approach towards a connectivity-first model.

The third force is sustainability. Carbon reduction commitments, with many large organisations targeting carbon neutrality by 2030, are influencing infrastructure design and site selection. Power usage effectiveness (PUE) targets of 1.3 are becoming standard expectations, and the sourcing of renewable power is a firm requirement in most major mandates. Cabling choices are not neutral in this equation. Poorly routed or densely packed cabling obstructs airflow paths, raises PUE figures and directly undermines the thermal efficiency targets that operators are contractually obligated to achieve. The transition from copper-based cabling to single-mode and multimode fibre is not a future consideration in this context. It is already the present baseline.

The primacy of east-west traffic

One of the most consequential structural changes driving cabling requirements is the inversion of traffic patterns within data centres. AI and machine learning workloads generate substantial volumes of intra-cluster traffic, that is, the movement of data between servers and GPU nodes within the same facility. This now accounts for approximately 80 per cent of total data centre traffic, with the remaining 20 per cent being north-south traffic that flows outward to users and external networks. Most cabling infrastructure installed in previous build generations was sized for a very different traffic distribution and is not adequate for the current workload profile.

The performance of the cabling layer directly determines how effectively east-west traffic is handled and, by extension, how efficiently AI workloads execute. GPU clusters function as a form of distributed computing, with multiple nodes operating collectively as a single virtual processing system. The network and cabling connecting those nodes are not support infrastructure but part of the compute architecture itself. InfiniBand has become the de facto standard for this interconnect layer and its integration into GPU cluster design is now treated as an embedded design requirement.

The practical implication is that 400 gigabit is the entry-level specification for east-west connectivity in AI environments. The industry is even preparing for 800 gigabit as the next standard, with 1.6 terabits of capacity expected to achieve widespread deployment within four to five years. Operators designing GPU clusters today are specifying their requirement for fibre infrastructure capable of supporting these future throughput levels, even where current deployments do not require them.

Challenges of scaling up

Despite growing awareness around the strategic importance of cabling, significant gaps between intent and execution persist. Several recurring challenges are emerging as operators and enterprises attempt to scale their connectivity infrastructure in line with rising workload demands.

Physical planning failures are among the most common. A frequently encountered scenario involves a data centre deploying racks progressively, only to find that cable trays installed during construction are insufficient for the density of cabling required. In documented cases, operators running at just 25 per cent of intended rack capacity have already exhausted available cable tray space. Cables obstructing airflow paths disrupt cooling system performance, raise PUE figures and undermine both operational efficiency and sustainability commitments.

Redundancy architecture presents a related challenge. Data centres that position themselves as carrier-neutral facilities with multiple connectivity providers may have those providers physically routed through a single path. A civil works event in the vicinity, such as road excavation, can sever what is presented as a fully redundant connectivity architecture. The physical diversity of fibre routes is as important as the commercial diversity of providers and the two do not always align.

Technology refresh timelines create a third pressure point. The cycle that once extended over a decade has compressed to three years and in some cases to a single year. Customers want shorter, more flexible arrangements, but high-specification cabling is a long-term capital investment by nature. Bridging that mismatch requires cabling architectures capable of supporting multiple generations of active equipment without replacement.

A skills gap is also emerging as a significant constraint. Engineers with practical experience in the installation, certification and management of InfiniBand networks and high-density fibre infrastructure are in short supply. This reflects a structural gap in workforce capability that will require deliberate and sustained attention from the industry and training institutions.

For regulated industries such as banking and financial services, data localisation requirements and regulatory mandates around data sovereignty add further complexity.

As these institutions move towards AI at scale, their connectivity requirements will shift from data-centric to compute-centric architectures, generating a substantial increase in east-west traffic intensity that the infrastructure built for data storage was not designed to handle.

Best practices for the modern data centre build

The industry is converging on a set of design and procurement principles gaining traction as standard practice in new builds. Among the most significant is the adoption of base-8 and base-12 fibre topologies.

These structured cabling architectures support the transition from current 400 gigabit deployments to future 800 gigabit and even 1.6 terabit configurations without requiring a wholesale replacement of the physical layer. Operators building on these topologies today can upgrade active equipment at subsequent technology cycles without disrupting the underlying cabling infrastructure, significantly reducing capital cost and operational disruption at each future transition.

A second principle is the separation of cabling management from the active rack environment. Rather than consolidating all cabling within operational racks, leading operators are deploying interconnect modules and optical distribution frame systems in dedicated spaces outside the primary operational area.

This approach improves airflow, supports better cooling performance and makes ongoing maintenance substantially less disruptive. Optical distribution frame systems offer modular, scalable architectures suitable across hyperscale, co-location and enterprise environments. Keeping the rack environment clean of excess cabling has a direct and measurable impact on cooling efficiency.

A third principle is discipline in fibre supply chain management. In high-stakes AI infrastructure projects, some operators are deliberately restricting procurement to a small number of vetted suppliers whose quality, delivery and performance standards are well established. In a market where supply chain disruptions can delay critical infrastructure milestones, this approach trades procurement breadth for operational certainty.

The adoption of splice-on connectors and ultra-low-loss components is the fourth area of growing focus. As data rates increase, the tolerance for signal degradation decreases sharply.

Traditional connector interfaces introduce potential signal loss that compounds over time through dust ingress and temperature variation. Splice-on connectors eliminate physical contact points, reducing attenuation risk and improving the reliability of high speed links across the operational life of the facility. Demand for these components has risen sharply in line with the growth of high-density AI deployments and is expected to continue on the same trajectory.

Outlook for future-proofing

The outlook for cabling infrastructure in data centres is one of sustained and accelerating relevance. The industry is converging on a set of forward-planning principles. The first is to design for peak workload rather than for the current or day-one deployment state. Changing active infrastructure is relatively straightforward; changing passive infrastructure is not.

Building cabling to a specification that anticipates the next two or three technology generations is a more durable and cost-effective investment than deploying for immediate requirements alone. Data centre agreements should carry provisions that allow for infrastructure upgrades as technology evolves. Invisible infrastructure must be the first thing planned and the last thing compromised.

The second principle is modularity. Rather than committing the full capital expenditure for peak-capacity infrastructure from the outset, operators are adopting phased approaches that allow the physical layer to scale as demand grows. Leaf-and-spine network architectures designed for incremental expansion, where additional capacity can be added floor by floor without disrupting existing operations, are the preferred model for large AI cluster deployments. This preserves capital while maintaining the upgrade pathways that future technology transitions will require.

On the technology front, silicon photonics promises to reshape the economics and performance profile of high speed data centre connectivity.

Network management is simultaneously moving towards software-defined architectures with AI-driven analytics, enabling real-time telemetry, self-healing network behaviour and end-to-end visibility from the physical layer through to the application layer.

These developments will further erode the boundary between active and passive infrastructure.

The most significant constraint on the horizon, however, is power. The electricity demand generated by large-scale AI infrastructure is growing at a rate that grid capacity in most markets cannot absorb quickly.

Facilities of 550 MW are already under construction for single customers and announced capacity figures continue to grow. If the cabling industry can help improve power efficiency alongside connectivity performance, that may represent the next meaningful frontier of innovation.

A second longer-term constraint is data readiness. The ability of organisations to realise returns on AI infrastructure depends on their data being structured, accessible and machine-readable. Organisations with fragmented or siloed data environments will find that even well-designed AI infrastructure delivers limited results until the underlying data architecture is addressed. Infrastructure investment and data readiness must advance in parallel.

For the cabling and connectivity industry, the direction of travel is unambiguous. The decisions made at the infrastructure planning stage, on topology, fibre specification, connector quality, route diversity and modularity, will define the operational and commercial performance of data centres for the next decade and beyond.

Getting those decisions right at the outset, rather than correcting them under operational pressure, is the defining challenge of the current build cycle.