As AI and GPU fabrics replace traditional workloads, the network underpinning data center operations is rapidly shifting from a quiet, background support system into a central pillar of performance and scalability.

Modern facilities are congested with more of everything: more density, fiber, cables, and more connections – all sat in and amongst rising demands for power and cooling capacity. But nothing can scale in isolation. As a result, this transition is unfolding on multiple fronts simultaneously.

Each layer introduces its own complexity – but it’s the interaction between these forces that’s creating the most significant challenges. Navigating this hybrid landscape demands a clear understanding of how the network has evolved, the ways in which its components now depend on one another, and how modern networks can align with transformations across the entire data center ecosystem.

For Usman Nasir, director of emerging technologies and strategy at Sumitomo Electric Lightwave (SEL), today’s network can be characterized as a system undergoing intense structural integration:

“The data center is now the computer,” he explains. “Traditional networks were built to connect people to servers, from node to node. AI fabrics are designed to connect accelerators to accelerators, and that changes the assumptions at every layer.”

Nasir views this transition as being shaped by three interrelated forces: tighter integration between components, an architectural requirement for lossless behavior, and the emergence of distinct front-end and back-end networks within the same facility. Together, these advancements are redefining how networks are designed, built, and operated.

The practical challenges

Perhaps the most fundamental development in the shift from traditional to AI-driven networks is the dramatic increase in bandwidth per rack. MPO/MTP (multi-fiber push-on/multi-fiber termination push-on) high-fiber trunk solutions are evolving toward higher strand counts per connector to support massive parallel optics in spine-leaf fabrics. Nasir outlines one of the practical challenges this creates:

“One of the biggest differentiators we require to achieve the cabling density needed is in very small-form (VSFF) connectors. The goal is to deliver three times the density of a standard MPO/MTP connector. This can translate to 2,048 fibers within just one rack when fully populated with 16-fiber MMC connectors – with potential to reach 3,072 fibers with 24-fiber connectors.”

This densification is essential for delivering the speed, quality, and ultra-low latency AI workloads demand. But such a dramatic move away from traditional setups introduces a host of new challenges.

Packing more fiber into the same physical footprint drives the need for increasingly compact connectors – elevating sensitivity to tolerances, handling, and installation quality. As pathways become congested with cables, network architectures must also be reimagined.

“When it comes to connectivity, more fibers per connector is essential,” explains Nasir. “12-fiber MPO was the standard for a long time, and 24-fiber solutions felt new not that long ago. Now, whether it’s MPO or even smaller formats such as MMC or SNMT, we need significantly more fibers per connector.”

The next challenge is how to physically connect and manage the sheer number of connectors going in and out of racks, switches, and patch panels. Traditional point-to-point cabling simply won’t meet the mark.

Breakout complexity has also increased significantly. When a single 1.6T port breaks out into two 800G or four 400G links for example – structured cabling with multi-drop breakouts to connect several racks as efficiently as possible becomes mission-critical.

Scaling up, out – and across

The scale of AI-driven change is so extreme that simply scaling up is no longer sufficient. Instead, networks must scale in three dimensions: up, out, and across.

Scaling up by adding more compute, memory, and power is a familiar process. Scaling out involves expanding horizontal fabrics to connect thousands of GPUs – typically within a single facility. Scaling across, however, represents a fundamental architectural shift.

“Scale across is the third pillar,” explains Nasir. “Technologies like Nvidia Spectrum-X are driving architectures that treat multiple geographically separate data centers as a single AI factory.”

While metro-connected data center clusters aren’t entirely new, they traditionally functioned as independent facilities. Now, workloads may span geographically separate regions, grouping compute resources across sites to act as one powerful computer.

“This kind of resource pooling reduces processing time, improves efficiency, and lowers latency for AI workloads – especially agentic AI,” says Nasir. “But it also adds significant complexity.”

These scale-across connections demand minimal physical latency, more coherent optics, and highly engineered fiber optic cables that are also faster and easier to deploy.

Navigating the bottlenecks

With 1.6T speeds now physically achievable – and increasingly realistic – the risks compound as layers of density, speed, and component interaction become more intertwined at this scale.

“Over the coming year, I think the key bottlenecks are going to shift again,” says Nasir. “What used to take five years to re-design and refresh is now rapidly evolving every 12-18 months. We used to worry primarily about raw bandwidth. Now it’s about operational stability and physical density.”

Two key concerns are emerging as the industry races towards ultra-high-speed connections: tail latency and optics reliability. The former is increasingly being addressed via adaptive routing – ensuring congestion-free, lossless Ethernet that reduces job completion time while lowering power consumption and improving bandwidth efficiency.

When it comes to optics reliability, the industry is steadily moving toward co-packaged optics (CPO) – integrating optical transceivers directly with switching application-specific integrated circuits (ASICs). By reducing distance from inches to millimeters and micros, CPO improves power efficiency and signal strength.

“Then there are linear pluggable optical (LPO) solutions,” continues Nasir. “LPO removes the power-hungry digital signal processor (DSP) from the module, cutting power consumption by up to 50 percent and significantly reducing latency.”

Back to basics

The central question therefore becomes: how can the industry keep up with this rapidly evolving ecosystem? The solution begins with a look at some fundamental principles – such as flexibility.

The ability to adapt and scale for future demand is now mission critical throughout the data center. Sumitomo is delivering on this must-have not only via innovative technology solutions, but with a customer-centric approach, and a commitment to knowledge-sharing.

“We’re the voice of the customer,” says Nasir. “We engineer what’s required, rather than trying to push one-size-fits-all products. And we focus heavily on knowledge transfer through training programs – providing hands-on education and enabling long-term partnerships.”

Sumitomo works closely with hyperscalers, community colleges, and vocational programs to address one of the industry’s most pressing challenges: an imbalance in the supply and demand for skilled labor.

These new networks require not just more workers, but new skills. For instance, the shift from copper to fiber – or the decision not to migrate from single-core to multi- or hollow-core fibers – demands an entirely new approach. Nasir expands:

“As we deploy hundreds of thousands of kilometers of un-terminated cable for scale-across networks, splicing, field termination, and testing all require deeper knowledge and expert precision. As such, spotlighting education is essential.”

As fiber counts rise, and link-loss budgets shrink, building these AI factories demands state-of-the-art products just as much as a force of skilled individuals equipped with the knowledge and experience to deploy and configure modern networks with speed and accuracy.

Speed to design

Another key consideration shared across the industry is speed to design. Not long ago, technology roadmaps were somewhat linear and predictable. But today, rapid innovation and hybrid optical-electrical architectures are making standardization both harder to achieve – and more critical than ever.

“For starters, we need strong collaboration across the ecosystem,” says Nasir. “Who defines what multi-core fiber should look like? Core spacing, geometry, cores layout? These decisions require co-innovation across many silos of the industry.

“When end customers and industry leaders collaborate to create reference designs, we avoid years of iterations – while delivering sustainable and interoperable specifications."

Having established resilient reference designs, the next challenge is to automate the ICT design role, leveraging a plethora of intelligent disparate software currently used in silos for different functions – from structured cabling and port mapping, to cooling and power system designs.

“Predictable performance requires simulation-driven design,” continues Nasir. “That looks like fully integrated modeling that considers power, cooling, connectivity, and resilience together as one holistic system. We need a paradigm shift in order to build the networks with the same approach used to design the chip.”

Nasir points to electronic design automation (EDA) as the central solution. EDA provides a singular platform that incorporates standards, reference designs, product specifications, and performance requirements – allowing ICT engineers to generate a virtual and physical model of the entire network, along with hyper-accurate bill of materials (BOMs) and port maps.

Looking ahead

Across an increasingly interconnected landscape, resilience at the network level must be built-in from the beginning. Today’s racks are host to a wider range of components than traditional setups – including cooling infrastructure.

For cabling, this means airflow paths must be highly engineered – and the risk of dust and condensation must be taken into account when designing, deploying, and maintaining these modern networks.

“At some point, you’re going to have to service these dense liquid cooled systems and introduce contaminations into ultra-high density physical connections.” says Nasir. “As a result, connector resilience is extremely critical. That’s why we’re introducing connectors that are not just smaller, but more resistant to dust, temperature variation, and moisture. Next-gen connectors using non-physical contact methods for coupling represent a silver lining for next-gen connectivity.”

These emerging solutions, with an expanded-beam between two mating connectors, significantly reduce sensitivity to dust and moisture. Instead of core-to-core contact, a lensed mating surface creates an air gap, minimizing wear and sensitivity to contamination, while significantly reducing installation time and maintenance requirements.

As higher fiber counts, smaller connectors, environmental resilience, and geographic scale advance simultaneously, the data center landscape is becoming a place where complexity must be embraced and managed, rather than avoided.

Predictable performance requires Electronic Design Automation (EDA) for a fully-integrated modeling of compute, storage, power, cooling, and resilient connectivity – all together as a real-time digital twin. Get in touch to build a resilient and scalable future together