The Forgotten Plumbing That Keeps Every AI Data Center Running

In September 2023, a single fiber-optic cable fault in a northern Virginia data center campus knocked a major financial services firm offline for eleven hours. The outage cost an estimated $9 million in lost trading revenue. Nobody wrote about the cable. They wrote about the cloud provider. They always write about the cloud provider.

This is the great misdirection of the AI infrastructure boom. Every analyst on Wall Street can quote you Nvidia's data center revenue to the penny. They can tell you how many megawatts Microsoft is contracting from nuclear plants and how many custom chips Google is fabbing. But ask them who made the fiber trunks connecting those GPU clusters, or which company's copper connectivity sits inside the cable trays routing power and data between racks, and you'll mostly get blank stares.

The physical layer of the internet has always been boring. And boring, in markets, means mispriced.

Here's what actually happens when a hyperscaler builds a new AI training facility. Before a single H100 or TPU gets racked, somebody has to run tens of thousands of fiber-optic cables through conduit. Somebody has to install the patch panels, the structured cabling, the antennas for private wireless networks inside the building. Somebody has to make sure the switching infrastructure can handle 400-gigabit and soon 800-gigabit Ethernet without packet loss that would crater a multi-week training run. This isn't glamorous work. It is, however, the work that determines whether a $2 billion data center actually functions or sits dark.

The numbers tell the story better than any analogy. Global data center construction spending is on pace to exceed $350 billion in 2026, according to Synergy Research Group. That figure was $150 billion just three years ago. Roughly 15 to 20 percent of any data center build goes to what the industry calls "structured cabling and connectivity" (the physical stuff). That's a $50 to $70 billion annual market that barely existed as a category a decade ago, when most enterprise networking was copper Cat6 running to office cubicles.

The shift from training to inference makes this even more interesting. Training clusters are dense and concentrated. Inference, which is what actually serves ChatGPT responses and Gemini queries to a billion users, needs to be distributed. More locations. More edge deployments. More fiber runs between smaller facilities. Alphabet's Q1 2026 call made this explicit: their Cloud backlog nearly doubled to $462 billion, and Sundar Pichai described a build-out plan touching dozens of new regions globally. Each one of those regions needs physical network infrastructure from the ground up.

Meanwhile, telecom operators (who were supposed to be a dead-end market) are quietly spending again. Open RAN deployments, 5G densification, and fiber-to-the-home expansions across Europe and Latin America are all driving demand for indoor cellular solutions, Wi-Fi access points, and the switching gear that ties it all together. One company that was spun off from a much larger conglomerate has been sitting at the intersection of all three of these demand waves (data center, telecom, broadband) for nearly fifty years. It changed its name earlier this year, which tells you something about how hard it's working to shed an old identity.

Founded in 1976 in Hickory, North Carolina, this business started making coaxial cable for the early cable television industry. For decades, it grew by following wherever wires needed to go. Cable TV in the 1980s. Enterprise networking in the 1990s. Fiber broadband in the 2000s. Each wave brought new products but the same core competence: connecting things physically, reliably, at scale.

Then came the AI data center wave. And suddenly a company that most tech investors had mentally filed under "mature telco supplier" found itself holding products that every hyperscaler and colocation provider needed yesterday. Fiber optic connectivity. High-density patch panels. In-building wireless systems. Network switches designed for the kind of east-west traffic patterns that GPU clusters generate.

The stock trades at a fraction of where it was five years ago. The balance sheet carries real debt from a leveraged acquisition. And the name change in January 2026 was an explicit attempt to signal that this is no longer your father's cable company. Whether the market listens is an open question. But the order book doesn't care about perceptions.