Since the dawn of the AI age, training models and inference workloads have evolved through series after series of supercharged networks and GPUs. Networks have sprouted more links, more lanes, denser spine-and-leaf fabrics, and transformer architectures with ever more heads – akin to living organisms that grow new limbs and neural pathways to cope with increasing demands.
As AI has penetrated mainstream life, the physical architecture of the data center has scaled up, out, and across, transforming from individual GPU clusters into sprawling, distributed AI ecosystems.
As part of Corning’s Connectivity Innovation Day 2026, the first DCD>Broadcast in the series, “Scale out: Building connectivity foundations for AI infrastructure,” saw Michael Crook, market development manager for data centers at Corning, outline how the active buildout of AI facilities is fundamentally reshaping how operators approach connectivity – not just at extraordinary scale, but at severely unprecedented speed.
When GPUs outgrow the rack
Each new generation of AI silicon delivers significant performance gains, but also introduces new networking and connectivity requirements. Higher port counts, denser architectures, smaller form factors, and evolving optical fiber technologies all influence how modern data centers must be designed.
In approaching this new season of connectivity, Corning works closely with leading chip manufacturers to truly gauge how next-generation connectivity solutions must align with these emerging system architectures.
“We really look at what kind of physical cabling we need in-rack, and then between racks, to connect everything together,” says Crook.
Looking inside a single AI rack, the scale of the challenge becomes apparent:
“If I just look at in-rack compute cabling, many are deploying AI nodes that may each have up to 72 GPUs, which are going to be spread across multiple server trays within that rack. Across those 72 GPUs, you’re going to consume over 1,100 fibers within that one rack.”
That density creates an unfathomable deployment challenge. Every cable assembly must be unpacked, labelled, routed, installed, and tested before the system can be brought online. And the complexity quickly multiplies from there. According to Crook, a single AI compute cluster with multiple compute racks can require more than 8,000 fiber connections.
“The challenge becomes exponentially larger as you build out these clusters,” he says.
As AI infrastructure scales from individual machines to clusters, and ultimately to distributed AI ecosystems, connectivity becomes a critical factor in determining how quickly and efficiently new AI capacity can be deployed.
From designing AI infrastructure to building it at scale
“Scale out” refers to adding more machines and nodes to an AI system, distributing workloads across multiple instances to deliver the parallelism that modern AI training and inference demand.
This scale out model is the primary driving force behind the exponential growth in cabling between racks and rows. Every new compute node added to a cluster requires high-speed fiber connections back into the network fabric, making physical deployment increasingly complex. As Crook explains:
“One of the main challenges is just the sheer number of cables that need to be installed within these AI nodes and the number of network switches that need to be deployed to build these large AI clusters. That’s really a time-consuming process, and hyperscalers want to go as fast as possible.”
Not to mention the actuality of real-world deployments whereby racks may not necessarily be positioned directly alongside the network switches they connect to.
Crook describes scenarios where power constraints have required compute rows to be located in different aisles – or even separate data halls – from the switching infrastructure. As a result, operators must route these high fiber counts through cable pathways that are already heavily congested, adding further complexity to an already taxing task.
Building faster, integrating smarter
The challenge is not solely technological. There is often a disconnect between fast‑moving AI and network architectures and the people, processes, and installation practices needed to deploy them at scale.
For data centers keen to stay in the running for commercial success, deployment timelines have dramatically compressed. Projects that might have taken 18 months to complete just a few years ago are now expected to be delivered in as little as six months. We’re talking larger scale-outs, over shorter timeframes.
As this trajectory appears to be here to stay, Corning’s focus is on making the installation of physical connectivity infrastructure as simple and efficient as possible – both through the design of the cabling architecture itself and the way it is deployed.
“That’s going to enable density, scalability, and ultimately speed to market,” says Crook.
Corning’s approach is to provide end‑to‑end support rather than simply supplying products. Its market development team works alongside hyperscale customers early in the design process to understand emerging architectures, deployment challenges, and operational requirements, feeding those insights directly into product development.
Reflecting on the company’s long history of innovation, Crooks adds:
“Here at Corning, we’re celebrating our 175-year anniversary, and that innovation pipeline is really in hyperdrive as we try to keep up with this new demand.”
The fiber shuffle
As Corning expanded its connectivity portfolio, increasing cable density was only the first step. The next challenge was enabling operators to efficiently distribute these ever-growing volumes of traffic across complex AI fabrics. The result was a set of “fiber shuffling” solutions to split connections, create more efficient physical layouts, and scale clusters.
Where hyperscale AI data centers may have hundreds of thousands of high-speed GPU connections, each MPO interface must be carefully mapped and distributed across multiple network switches. Fiber shuffling provides a deterministic way of breaking out the individual fiber lanes within a multi-fiber connector and routing them to their required destinations.
When considering the scale of modern AI racks, it is clear to see why this is needed.
Crook explains that a typical switch provides around 50Tbps of capacity, equating to roughly 128 400G ports. In a non‑blocking leaf-spine architecture, about half of those ports connect downstream to GPUs and the other half connect upstream into the spine, leaving only 64 ports available to serve compute racks.
A single AI rack, however, can require as many as 144 high‑speed connections – more ports than a single switch can support. Instead, each rack must connect to multiple switches, requiring every multi-fiber MPO cable to be split and fanned out across several switching devices.
“That’s really what shuffling is. Taking those multiple channels within a multi-fiber interface and mapping those individual channels to multiple output interfaces, so I can physically connect those network switches together,” says Crook.
To support different deployment models, Corning provides both a modular “shuffle box” and an integrated “shuffle assembly,” each with different trade-offs in flexibility, density, and insertion loss, so customers can match the shuffling approach to their architecture and performance constraints.
“We’re already seeing designs where next‑generation GPU nodes will have up to 576 GPUs per node, and customers want to connect multiple of these nodes together. That’s why we’re working closely with them on where to place the shuffling function in the network.
“The right choice really depends on the customer’s data center design. If they want to minimize attenuation, a cable assembly can be the better option; for smaller nodes or different design constraints, the hardware solution may be the best fit,” explains Crook.
Packing in the fibers
The growing use of fiber shuffling also creates a practical challenge that comes in the way of deploying and managing these very dense fiber infrastructures. As AI clusters scale, operators need to move enormous fiber counts between racks, rows, and buildings, while installing and splicing those fibers as quickly and efficiently as possible. Ribbon cabling is one of the key physical enablers that makes this possible.
A ribbon cable arranges multiple optical fibers side by side in a flat, wide structure, creating a compact footprint for exceptionally high fiber density. Just as importantly, it allows technicians to splice 12 or 16 fibers in a single operation rather than performing the same number of individual splices. It’s a practical, time‑saving way to deliver the dense fiber that shuffling architectures depend on.
Where ribbon cable is already the de facto standard for high‑density building‑to‑building connectivity – a topic explored in greater detail in Corning’s “Scale across” DCD>Broadcast – Corning believes many of the same advantages can be brought inside the data center, extending the same efficiencies to AI deployments.
What’s next for scale out?
If there is one clear takeaway, it’s that AI infrastructure shows no sign of slowing. GPU nodes are already evolving from today’s 72-GPU configurations toward systems with up to 576 GPUs, while hyperscalers continue to plan even larger AI clusters within single data center campuses.
Meeting that demand depends on connectivity solutions actively designed to reduce installation complexity, accelerate deployment, and improve the practicality of building AI fabrics at scale.
And eventually, “scale out” becomes “scale across”. As individual sites reach their power and space limits, operators will need to extend operations across multiple buildings – and ultimately entire campuses – as the next evolution in AI infrastructure.
For Crooks, the not-so-secret secret lies in connectivity designed into AI infrastructure from the outset. That means working with customers early enough to align physical infrastructure with evolving AI architectures.
“Bring us in early. Let our engineers have conversations with your engineers. Let us look at our current portfolio, or let’s look at whether we need to undertake additional innovation and development to help our customers make the right cabling and hardware decisions. That’s what we’re here for,” Crook concludes.
Watch the full DCD>Broadcast episode on the fiber scale out here.
More from Corning
