The infrastructure nobody wanted is the infrastructure everyone uses

This is a sponsored article brought to you by Find Better Internet.

Gary Winnick built Global Crossing into a $47 billion company in three years. He did it by selling a story: that internet traffic was doubling every 100 days, and whoever owned the pipes when demand caught up would own the future. Investors believed him. So did Bernie Ebbers at WorldCom, and the dozens of smaller carriers racing to lay cable under oceans and across continents before anyone else got there.

By 2002, Global Crossing had filed for bankruptcy with $25.5 billion in assets and Winnick’s fortune had evaporated along with it. WorldCom’s collapse took Ebbers down with a 25-year federal sentence for accounting fraud. More than 60 telecom companies failed between 2001 and 2003, and the fiber they had spent hundreds of billions of dollars burying in the ground sat dark, unused, waiting for traffic that was growing far slower than anyone had promised investors.

That fiber didn’t disappear. It got bought, repriced, and quietly folded into the infrastructure IT teams now depend on without a second thought, and the story of how it got there says more about enterprise technology planning than most vendor pitch decks ever will.

The math that sank an industry

The 100-day doubling claim wasn’t just optimistic. It was wrong by an order of magnitude, and the person who proved it was Andrew Odlyzko, a mathematician then at AT&T Labs and later the University of Minnesota. Odlyzko’s research on internet traffic growth found that actual demand was increasing 75% to 150% annually, not doubling every three months. The gap between those two numbers is the gap between a healthy build-out and a bubble.

Carriers had already committed to the faster number. Dense wavelength-division multiplexing let a single fiber strand carry dozens of separate light signals at once, so the supply side of the equation was also expanding fast. Overbuild layered on top of an inflated demand forecast, and the combination produced a network built for a version of the future that was still years away.

Some of the accounting used to paper over that gap turned criminal. WorldCom’s fraud ran through a mechanism regulators later called capacity swaps: carriers traded fiber capacity with each other and booked the trades as revenue on both sides, inflating growth numbers that were already inflated by bad math. Bernie Ebbers eventually faced federal charges tied to $11 billion in fraudulent accounting built substantially on that pattern.

The result: by 2002, roughly 97.5% of installed long-haul fiber sat dark, carrying no traffic at all. Level 3 Communications, one of the few carriers that survived the crash intact, eventually carried 70% of global internet traffic at its peak, largely by buying up capacity and assets from competitors who didn’t make it.

Assets nobody wanted, bought at a discount

Bankruptcy court is where the real story starts. When Global Crossing, WorldCom, and dozens of smaller carriers went under, their fiber networks didn’t get ripped out of the ground. Creditors sold them, often for a fraction of what it cost to build them, to whoever was still standing and willing to write a check.

Level 3’s $34 billion acquisition price when CenturyLink bought it in 2017 traces back to exactly this pattern of distressed-asset accumulation. The company built its network largely by absorbing capacity from failed competitors during the crash years, then rode the fiber it had picked up cheap into the streaming and cloud era. One detailed account of the buildout puts the total market value destroyed at two trillion dollars, a figure that dwarfs what the carriers themselves had spent laying the cable in the first place, and traces how that same discarded capacity ended up carrying Netflix, Zoom, and enterprise cloud traffic a decade later.

Nobody planning enterprise network procurement in 2002 was thinking about Netflix. The fiber existed because a generation of telecom executives bet wrong about demand, went bankrupt, and left behind physical assets that outlasted the companies that built them by twenty years and counting.

Dark fiber never really stayed dark

The unused capacity from the crash years became known as dark fiber, and it turned into one of the more overlooked line items in enterprise connectivity procurement. Carriers that survived the shakeout could light up existing strands for a fraction of the cost of trenching new cable, which meant bandwidth prices kept falling for two decades even as demand climbed.

That price collapse is a large part of why cloud computing works as a business model at all. Data centers depend on cheap, abundant fiber to move traffic between facilities and to customers, and the glut from the crash years subsidized exactly that for far longer than anyone expected in 2002.

The demand side eventually caught up, and then some. Data Center Knowledge reported that bandwidth purchased for data center connectivity surged nearly 330% between 2020 and 2024, with metro dark fiber purchases alone climbing 268% in a single year. Just ten buyers, mostly hyperscalers and large carriers, accounted for 62% of all bandwidth purchases in 2024. The same asset class that sat idle and worthless in 2003 is now the thing hyperscale operators compete to acquire.

The pattern enterprise IT keeps missing

Every enterprise IT team that signs a connectivity contract, provisions a leased line, or negotiates data center interconnect pricing is transacting in a market shaped by a bankruptcy wave most of them have never heard of. The infrastructure is invisible until something breaks or a bill spikes, and by then the history behind why prices are what they are has usually been forgotten.

It’s happening again, and the players have changed. An IEEE Communications Society analysis comparing today’s AI infrastructure spending to the dot-com fiber buildout found that Big Tech companies plan to spend $364 to $400 billion on AI infrastructure, with quarterly data center spending jumping from $9.5 billion in early 2020 to $40.4 billion by the second quarter of 2025. Bain and Company estimates the industry needs $2 trillion in annual AI revenue by 2030 to justify that spending, against roughly $45 billion in AI revenue today.

That same analysis points to research suggesting the return on that spending isn’t showing up yet inside the companies buying it. MIT Media Lab found that 95% of custom enterprise AI tools fail to deliver measurable impact. Gartner puts the failure rate for AI initiatives at 85%, and McKinsey found 80% of companies using AI report no significant effect on their bottom line. Even Bret Taylor, OpenAI’s board chair, has said publicly that the industry is in a bubble and that a lot of people will lose a lot of money.

The gap is large enough that even people inside the industry are naming it out loud. The difference this time, according to the same analysis, is that today’s hyperscalers generate far more cash than 1990s telecom carriers did, which gives them more room to absorb a slowdown without the kind of bankruptcy cascade that hit Global Crossing and WorldCom. Whether that cushion holds is a separate question from whether the spending itself is rational, and IT leaders evaluating vendor contracts and long-term infrastructure commitments right now are the ones who will find out either way.

Reading the contract behind the contract

Enterprise IT teams don’t need to stop building on AI infrastructure or stop signing data center contracts. They do need to treat today’s pricing and availability as a snapshot of a market still mid-cycle. The carriers offering rock-bottom bandwidth pricing in 2003 were selling off assets built by companies that no longer existed. The cloud providers and colocation vendors offering aggressive AI infrastructure pricing today are making a similar bet on future demand, funded by balance sheets that are stronger but not infinite.

Procurement teams that assume today’s pricing and capacity commitments are permanent are making the same mistake Global Crossing’s customers made in 2000, just with better financing behind it. Contracts negotiated during a build-out phase tend to look very different once the market is repriced, and the companies that survive a correction are rarely the ones who signed the most aggressive expansion deals at the peak. Reading the fine print on capacity commitments, exit terms, and pricing floors matters more during a boom than it does once the market has already settled.

What actually lasts

The fiber crash punished the people who caused it. Winnick and Ebbers lost their companies. Thousands of employees lost their jobs. Investors lost the two trillion dollars mentioned earlier, money that never came back regardless of what the fiber eventually got used for.

What survived was the physical layer, because glass in the ground doesn’t go bankrupt even when the company that buried it does. That’s the distinction worth holding onto: the capital gets destroyed, the infrastructure gets reallocated, and the entities that end up controlling it are rarely the ones who paid to build it. Enterprise IT teams evaluating today’s AI infrastructure boom are, whether they realize it or not, placing a bet on which side of that split they’ll end up on when the current cycle turns.