For years, conversations around AI infrastructure have focused on one critical challenge: scaling out. Inside the data center, this looks like adding more machines and nodes to existing systems, distributing workloads across multiple instances.

But as deployments continue to evolve at pace, the industry must prepare for what comes next.

Beyond adding more racks to the data hall, the next generation of AI systems will require concentrating significantly more capacity inside each individual rack, driving a fundamental shift in how connectivity is designed.

Rather than simply scaling out across more and more racks, network architectures are scaling up, adding more resources to individual nodes within existing footprints. Alongside this shift, the industry is transitioning away from copper towards optical interconnects capable of supporting the bandwidth and latency AI demands.

In a recent DCD>Broadcast episode, Corning’s Shirley Brown explores why connectivity is emerging as one of the defining infrastructure challenges of next-generation AI deployments, and how expanded beam technologies are helping to address many of the operational pressures accompanying this transition.

A new landscape for connectivity

While today's AI infrastructure remains heavily focused on scale-out architectures, scale up is now influencing how future systems are designed.

"It's more density, more network capacity, and more fiber connectivity within a single rack," explains Brown. "Everything becomes physically closer together, and everything – compute, switching, and connectivity – is all in one physical space."

This concentration is dramatically altering the physical environment inside the rack. Fiber routing becomes more complex and access to individual connections more limited, making routine installation and maintenance tasks challenging.

Simultaneously, the network landscape itself is evolving. Copper has long provided a practical solution for many short-reach applications, but AI's rapidly increasing bandwidth requirements are exposing its limitations.

"The story is really about latency," says Brown. "Copper is reaching its limits on bandwidth and latency that fiber is going to be able to solve."

Optical connectivity provides the performance modern AI clusters require, but it also introduces a new set of operational considerations. Unlike copper, traditional physical-contact fiber connectors rely on microscopic fiber cores aligning precisely with one another, making them highly sensitive to contamination – particularly during installation and maintenance.

"We're starting to hit this threshold of needing a new technology to solve a lot of the deployment speed challenges that our customers face," adds Brown.

Scaling for AI

AI infrastructure has already transformed the volume of fiber required inside modern data centers. According to Brown, today's large-scale AI deployments can require roughly ten times more fiber than traditional enterprise environments.

While this delivers the network performance required to support increasingly powerful GPU clusters, it also places new demands on deployment teams.

Installing substantially more optical infrastructure requires more skilled technicians, while maintaining consistency across large-scale deployments becomes increasingly difficult.

"The traditional requirements of inspecting and cleaning connectors are much more challenging whenever you have this massive increase in labor," explains Brown.

This complexity extends beyond installation. When GPUs account for a substantial proportion of overall deployment costs, every additional day spent commissioning infrastructure delays production workloads and postpones revenue.

"Our customers are making massive investments," says Brown. "They really need that return on investment quickly, and that means deployment speed needs to go up."

Increasingly, connectivity represents more than a networking consideration. It directly influences how quickly AI infrastructure can begin delivering value.

Reliability in increasingly inaccessible environments

The issue of maintenance is an important one. Even today's scale-out environments require engineers to carefully disconnect individual links without disturbing neighboring connections. Brown describes this as managing the ‘blast radius’ of maintenance, where a single intervention must avoid creating unnecessary disruption elsewhere in the system.

Future scale-up architectures are expected to make this considerably more challenging. As optical connectivity moves deeper into the rack, technicians may no longer have direct access to individual connectors. Instead, entire GPU trays could need to be removed, simultaneously disconnecting hundreds of optical links.

"Once that connectivity moves to the backplane, it’s no longer accessible," explains Brown.

This fundamentally changes expectations around connector performance. Connectivity solutions that work correctly from the moment they’re installed – and can continue doing so with minimal intervention – become imperative.

Rethinking fiber connectivity

Addressing these challenges demands an entirely different approach to how optical connectivity functions within increasingly dense AI environments.

Together, Corning and USConnec have developed MMC Connectors with PRIZM TMT Ferrule solutions. Rather than relying on the direct alignment of two microscopic fiber cores, PRIZM TMT uses precision optical lenses to expand and collimate light before transmission, creating a significantly larger optical area.

While physical-contact connectors have supported optical performance for decades, their design means the fiber cores themselves must physically touch. This makes them inherently vulnerable to contamination.

Expanded beam technology actively addresses this constraint. As light exits the fiber, it passes through a collimating lens and travels across a protected air gap before being focused back into the receiving fiber. Because the beam itself is substantially larger, the optical interface becomes far less sensitive to dust and debris than conventional physical-contact connectors.

"The goal here with lens connectivity is plug and play," explains Brown. "It's no longer the world of the last 50 years of fiber connectivity, where you inspect and clean and then mate. It's about plugging it in and it works."

As AI deployments continue to accelerate, predictable installation becomes increasingly valuable, helping operators deploy infrastructure faster while reducing the operational variability that comes with high-stakes, large-scale installations.

Familiar, but fundamentally different

Introducing new technologies into critical infrastructure inevitably presents adoption challenges. Operators have spent decades building processes around physical-contact connectors, and replacing well-established ecosystems comes with some innate risk. Brown outlines how PRIZM TMT’s design helps to address this key concern:

"It feels and looks very similar to physical contact. It's a familiar form factor, but you get this really great benefit from a dust-insensitivity perspective."

Maintaining a familiar interface allows operators to benefit from expanded beam technology without fundamentally changing how they approach optical connectivity. Existing deployment practices remain largely the same and troubleshooting stays straightforward when required.

"If you do have to troubleshoot, you can still view the interface and confirm there's no physical issue before moving on to investigate other parts of the system," adds Brown.

This combination of familiarity and improved operational resilience offers a practical path towards next-generation optical connectivity.

Partnering up with blind mate

While expanded beam technology delivers immediate advantages for modern AI deployments, its greatest significance may lie in supporting next-generation architectures.

One of the defining characteristics of future scale-up systems will be blind mate optical connectivity.

As GPU trays become denser and optical infrastructure increasingly inaccessible, manually connecting individual fiber assemblies becomes both impractical and inefficient.

"When you're dealing with thousands of fibers in a rack," explains Brown. "You can't be sitting there plugging in individual connectors one at a time."

Instead, entire groups of fibers will need to connect automatically as hardware slides into position. This enables manufacturers to assemble complex AI systems more efficiently, while allowing operators to replace hardware with minimal disruption once systems are deployed.

"You need to be able to mate hundreds – or even thousands – of fibers in one shot."

The operational advantages extend throughout the infrastructure lifecycle. Rather than disconnecting and reconnecting individual fiber assemblies during maintenance, engineers can simply remove a GPU tray, install a replacement, and restore service significantly faster.

However, this approach places even greater emphasis on connector reliability: "We need a solution that's essentially trouble-free and works right the first time," emphasizes Brown. "That's really going to be imperative in backplane applications."

Connectivity as the backbone of operations

Behind this conversation sits a broader reality. Many of tomorrow's deployment challenges must be addressed long before new systems ever reach the data hall.

While conversations around AI frequently focus on GPUs, power delivery and liquid cooling, connectivity decisions can slide down the priority list during the earliest planning stages.This approach is increasingly outdated.

"If you're not thinking about what you need to do a couple of years down the road. You're likely to be locked into today's solutions.”

For Brown, the direction of travel is clear: "AI systems are getting denser, faster, and harder to access. The physical connectivity model is reaching its limits."

Technologies such as PRIZM TMT are helping operators address these emerging challenges by simplifying deployment, improving reliability, and enabling the blind mate optical architectures expected to underpin future AI systems.

Perhaps most importantly, this shift reinforces that connectivity must become an integral part of infrastructure planning from the earliest stages of AI deployment.

To hear more about scale up from Shirley Brown, watch the full DCD>Broadcast episode, here. Plus, read the latest MMC Connector with PRIZM TMT Ferrule brochure, here.