AI workloads are growing larger and more complex, demanding faster connections between thousands of GPUs all working together seamlessly from racks located within each and every corner of the data center. This reality demands a rethink of the networking and fiber strategies that will define the future of AI infrastructure. 

The central question becomes: are traditional network architectures equipped to keep up with the latency and bandwidth requirements of next-generation AI workloads?

The solution lies in scale-up networks: high-bandwidth, low-latency interconnects designed to link accelerators, such as GPUs, within a single rack to deliver the high-speed connections required for large-scale AI training. 

During a DCD>Broadcast episode from Corning’s Telco Innovation Day 2025, Christina Chan, Corning’s program manager, optical scale-up, addresses why scale-up networking is the key to enabling next-gen AI, and how fiber is set to become the third utility alongside power and cooling.

The end of the line for copper 

The move from traditional compute demands to high-density, power-hungry AI workloads is powering a dramatic shift in network infrastructure. As the demand for AI increases not only in volume, but in the complexity of its outputs, bigger and better networks become increasingly mission critical. 

Scale-up solutions have typically been associated with copper connections, but the shift from single-rack setups to next-gen multi-rack deployments demanding higher speeds over greater distances is making fiber a non-negotiable. Chan explains: 

“AI models are growing around 100 times every two years for things such as scientific simulations and other critical applications across a range of industries. At Corning, we’ve analyzed the data rate and compared it to the distance of connections between different GPUs to determine that there is a point where copper and electrical signals will not be able to meet the mark.

“The cumulative length of all the interconnects that we're seeing today is around three and a half miles or more of copper in one single rack, which is pretty impressive. But as hundreds of linked-up GPUs become a basic requirement for supporting AI workloads, that same interconnected length of copper would need to be 30 plus miles – so 10 times more at the very least.”

To deal with higher data rates, copper requires heavy shielding, since it’s prone to interference from surrounding equipment. Racks don’t just contain a high-volume of GPUs – they’re home to a range of critical components that work together to power the data center. 

So, while copper acts as a data pathway, it also needs to be protected to ensure it carries the right signals. This requires a thicker cable and utilizes more copper. The sheer volume of materials required to meet today’s networking demands via the traditional routes is inefficient and unsustainable. 

Providing pathways that enable the speed of connection and reach between different racks required by modern data center operations, fiber optics can accommodate the distance, power and complexity that its copper counterpart simply can’t achieve.

This shift also comes down to energy efficiency. When electrons travel along copper wires and collide with atoms, the faster they go, the more energy is lost – causing the signal to become weaker in response. Fiber doesn’t utilize electrons in this way, removing the risk of energy loss during increased data rates. Instead, it uses photons (light). And, critically, light moves faster than electricity. 

A bright future ahead 

The transition from electrons to photons brings a range of benefits. Chan elaborates: 

“A large volume of GPUs working alongside one another under a singular system means that they need to be connected with low latency so that they can move at speed to pass information back and forth.

“Photons remove any issues of physical electrons bumping into one another, or electromagnetic interference, which means you don't need to make the interconnects or the cables as thick. As a result, there’s space for more capacity and more connections inside the cable, and they’re running more efficiently. Plus, they deliver higher bandwidth over a greater distance without any signal degradation.”

Ultimately therefore, building a scale-up network with fiber optic means more points can be connected efficiently and reliably within a smaller footprint, utilizing less materials. Crucially, this approach delivers not just accelerated data rates, but speed of installation to streamline time to market. Chan outlines how Corning is keeping this pace: 

“The ability to get networks up and running via solutions such as blind mate – a way to link up multiple connectors quickly and cleanly through the back of the racks – reduces installation times from months to days.”

The volume of connections required for today’s operations cannot be underestimated. Each new GPU must be linked up to every GPU in the system, compounding the complexity with each addition. Add in the additional hurdle of navigating across significant distances, and the need to simplify and fastrack this process is critical to viable deployment. 

The key challenge is delivering these high-density connections not only at pace, but reliably. Individual fibers and the connectors themselves must be installed accurately to create a fully-functioning, streamlined ecosystem. 

Denser and denser

It’s safe to say that densities inside data centers will continue to rise. With networking underpinning these increasingly critical operations, prioritizing fast, reliable, AI-ready setups is a must-have today in order to meet future demands.

The extent of this change is catapulting networking beyond the gray space of legacy solutions and into the white space of cutting-edge technologies: Chan outlines what this means for data center design, planning, and construction: 

“With so much complexity and multiple layers of design considerations required to deliver scale-up networks, we're seeing the need to bring these connections into early discussions. To ensure all of these fibers – potentially tens of thousands in one rack – fit in the space, it’s essential to consider it alongside the facility layout as a whole. As a result, we’re seeing fiber become the third utility in addition to power and cooling.”

This means thinking about how high-density fiber cabling will interact with other critical systems and, as these complex networks reach into the white space, it’s necessary to consider its impact upon operations and infrastructure as a holistic unit.

The knock-on effect of this shift is also making close collaboration with key design partners across the data center ecosystem an essential component for ensuring what’s built today can scale and adapt for tomorrow. 

“For Corning, what that looks like is being able to mold to all the different sizes and dimensions for multiple generations,” explains Chan. “It’s about having enough interconnects, fiber, and cabling to support whatever the future demands.” 

In practice, this can mean installing more fibers upfront that may not be necessary for operations right now, but will be ready to get up and running when required, or configuring the network in a way that can be quickly and easily switched around down the line.

Meeting these requirements for flexibility and density, alongside quick and easy installation, is both a significant challenge and exciting opportunity for innovators like Corning.

“We’re actively listening to what's going on in the industry and making sure we're tracking key trends and new developments,” explains Chan. “We’re also working on co-packaged optics (CPO) as a central strategy to enable faster scale-up networks.

And what that really means for Corning is that we need to start developing and testing these solutions right now, without delay, to enable full-scale deployment before they’re needed.”

When it comes to next-gen networks, Corning is at the heart of championing this collaboration, establishing long-term strategic partnerships, while keeping a keen eye on the road ahead in order to make high-density, compact, reliable, and streamlined network solutions within reach. 

>> Questions about how scale-up innovations will affect your infrastructure? Corning provides free HSDC engineering support. Click here to contact Corning today!