On an architectural diagram, an AI network appears as clean lines connecting racks and switches. On the data center floor, those lines become thousands of fibers that must reach the correct GPU, work first time, and remain accessible when the system changes.

AI has raised the stakes for that physical infrastructure. Users expect immediate answers, while operators need to bring expensive compute capacity online quickly. A dirty connector, incorrectly routed fiber, or delayed delivery can prevent GPUs from joining the cluster.

Operators therefore need infrastructure that can be deployed repeatedly at scale without becoming unmanageable, while retaining enough flexibility to support new hardware.

In a recent DCD>Broadcast episode, CommScope’s Ken Hall, Scott Eischens, and Rudy Musschebroeck explore how AI is changing network design from initial installation through future expansion, and why planning for day two now matters as much as delivering day one.

From overnight processing to immediate answers

Traditional applications could tolerate jobs running overnight. AI has created an expectation of immediacy, placing greater importance on low latency and time to first token.

“We need to enable low-latency networks so we can access data and manage the compute capabilities in a much faster, much larger environment,” says Hall. “The expectation today is immediate, or close to it.”

AI networks are built by repeating modular building blocks. GPUs and CPUs form rack-scale systems, grouped into pods and connected through layers of switching. Keeping the network flat reduces latency, but physical connectivity multiplies rapidly as clusters grow.

Musschebroeck divides that growth into three dimensions:

• Scale-up: Connects the GPUs and servers within a rack so they behave as one high-performance computer, today predominantly using copper.
• Scale-out: Links racks and pods across the data hall, largely through fiber.
• Scale-across: Connects separate halls, buildings, or sites when power or space constraints prevent an AI system from occupying one location.

“Every single one of those neurons needs to talk to the others,” he says, comparing the GPUs to cells in a human brain. “It’s very important that everything is tightly interconnected.”

When clean lines become thousands of fibers

In a typical deployment, hundreds or thousands of GPUs can translate into tens of thousands of individual fiber connections. Each must be mapped, labeled, routed, and connected correctly while remaining manageable within an already crowded facility.

“When you step into a data center and see it, it’s super dense,” says Eischens. “Those nice, neat lines are a whole bunch of fiber connections going all over the place.”

For Musschebroeck, this creates three immediate deployment challenges. The first is speed: operators need materials delivered and installed quickly so substantial infrastructure investments can begin generating revenue. The second is first-time-right quality, ensuring every expected GPU appears when the system starts. The third is logistics.

“We’re looking at a 2D blueprint, but the reality is three-dimensional, with poles, different ceiling heights, and cable ladders at different heights,” he says. Cable lengths and routes must reflect the building rather than simply the network schematic.

Taking work off the data center floor

Point-to-point cabling may appear simple, with one connection at either end, but high cable volumes can stitch cabinets together and obstruct access when equipment needs replacing or upgrading. Structured cabling introduces an additional connection point but provides greater flexibility between cabinets and pods.

It also allows more integration, inspection, and testing to take place offsite. Trunks and utilities can be installed in advance, while cabinets are assembled and tested in a controlled environment before being rolled onto the data center floor for final connection.

This is increasingly important as different trades compete for space around AI hardware. Power, liquid cooling, and connectivity all converge on GPU cabinets. Cooling hoses and large power connections increasingly occupy the rear, pushing fiber toward the front so each system can remain accessible.

CommScope’s Rapid Fiber Connect platform, explored in a recent DCD Tech Showcase, allows GPU cabinet connections to be integrated and tested offsite, then joined through fewer Mass Insert MMC connectors.

“It pushes fiber toward the same mindset as power and cooling,” says Eischens. “Drop the cabinet into place, make a couple of clicks, and the whole cabinet is ready to go.”

Building resilience into the physical layer

AI traffic cannot be oversubscribed in the same way as conventional cloud or telecom networks. GPUs may need to communicate with any other GPU at full speed, creating non-blocking architectures in which bandwidth is preserved through each switching layer.

The links themselves can contain multiple lanes. A 400Gbps connection, for example, may consist of four 100Gbps lanes. A multiplane architecture shuffles these lanes across different switches rather than routing them all through one. If a connection fails, the network may slow down, but the entire link does not disappear. The same approach can support more GPUs without adding another switching tier and its associated latency.

That logical resilience creates physical complexity. CommScope’s Propel shuffle modules place the lane distribution inside a replaceable module while retaining standard trunks and patch cords. If requirements change or something is damaged, operators can replace the module rather than an entire shuffled cable assembly.

Connector technology is evolving too. CommScope’s FastSelfClean technology uses V-grooves instead of conventional ferrules, cleaning the fiber end faces each time the connector is inserted. Its first planned implementation, FSC144, brings 144 fibers together in one connector with insertion loss of less than 0.1dB.

“You want to connect many fibers at the same time, and you don’t want to worry about cleaning or inspection,” says Musschebroeck. “Reliability needs to go up in these systems.”

Following fiber into the future

The boundary between equipment and network infrastructure is also beginning to shift. Co-packaged optics integrates electrical-to-optical conversion more closely with switching silicon, reducing the power consumed by pluggable transceivers and bringing fiber deeper inside switches.

Musschebroeck expects a similar transition within rack-scale systems. Their scale-up networks are currently connected predominantly through copper, but higher speeds, greater performance, and rising GPU counts are likely to increase the role of fiber.

The definition of a cabinet may change too. Eischens notes that GPU counts per system are progressing from 72 toward 144, with 576 or more under discussion for future designs. Accommodating those systems could require multiple cabinets operating as a single unit, or an enclosure approaching the size of a small room.

Operators therefore need to consider maximum capacity, materials, labor, and expansion from the beginning. They must decide what belongs offsite and what must happen on the data center floor, while enabling repairs and upgrades without wholesale recabling.

“You’ve got to be looking at where this is going and what the next step will be,” Hall concludes. “If the next step doesn’t fit inside those walls, you’ve got to look at how you add to that.”

For AI networks, getting to the first token may be the immediate goal, but reaching it efficiently depends on planning well beyond day one.

To hear more about how fiber infrastructure is evolving for AI, watch the full DCD>Broadcast episode with Ken Hall, Scott Eischens, and Rudy Musschebroeck, here.