The Global Crossing of Compute

In 1997, Global Crossing was founded to lay a transatlantic fiber-optic cable, funded by massive piles of debt. The infrastructure they built was entirely real, and entirely transformative. The business model, though, was structurally doomed by a fundamental economic reality: the marginal cost of bandwidth was zero.

Laying submarine cable is an exercise in astronomical fixed capital expenditure. A company must manufacture thousands of miles of specialized fiber, charter maritime fleets, and trench the ocean floor, all before a single dollar of revenue can be recognized. Once that cable is finally lit, however, the marginal cost of routing an extra byte of data from New York to London is functionally zero. In a competitive market where multiple heavily indebted players were racing to lay their own proprietary cables, the price of transit inevitably fell to meet that marginal cost.

The math was unyielding. When revenue crashes toward zero, carrying a $12.4 billion debt load becomes unserviceable. The inevitable result was an industry-wide wipeout, culminating in the historic bankruptcy of WorldCom. The fiber remained in the ground, but the equity of the companies that put it there was completely wiped out.

This telecom prelude is the most instructive parallel we have for the current macroeconomic environment surrounding artificial intelligence. Today, we are witnessing another massive, debt-fueled buildout of foundational infrastructure. Instead of dark fiber, the physical substrate is compute — specifically, vast clusters of graphics processing units, heavily weighted toward Nvidia's high-end architectures. And instead of telecom upstarts, the vehicle of choice is the middle-layer AI cloud provider.

Consider CoreWeave. The specialized cloud provider recently raised a $7.5 billion debt facility, collateralized largely by its fleet of highly sought-after GPUs. This is an extraordinary amount of leverage for a startup, and it is not an isolated incident; it reflects a broader explosion of private credit flooding into the infrastructure layer of the AI boom. Financial authorities are increasingly tracking the private credit systemic risks tied up in these aggressive, high-yield borrowing facilities.

The compute infrastructure these companies are purchasing is undeniably real, and likely just as foundational to the next decade of software as broadband internet was to the last two. The vulnerability lies entirely in the balance sheet.

To understand why this leverage is fatal, you have to follow the profits across the value chain. If we map the 1999 telecom stack alongside the 2024 AI stack, the locus of vulnerability becomes immediately apparent.

In 1999, the value chain was straightforward:

• The Monopoly Supplier: Cisco extracted massive margins by selling the physical networking gear.
• The Middle-Layer Distributors: Global Crossing and its peers bought that gear with high-interest debt to build out transit capacity.
• The End-Users: Pets.com and a thousand other venture-backed dot-coms rented the bandwidth.

Today, the actors have changed but the structural leverage is identical:

• The Monopoly Supplier: Nvidia extracts the vast majority of the margin by selling the GPUs, boasting gross margins north of 70 percent.
• The Middle-Layer Distributors: Specialized cloud providers like CoreWeave buy that gear using private credit.
• The End-Users: A myriad of AI wrapper startups and foundational model builders rent the compute.

Running alongside this stack are the hyperscalers — Amazon, Google, and Microsoft. They are also distributors, but with a critical difference: they possess fortress balance sheets and generate immense free cash flow from their legacy software and traditional cloud businesses.

The fundamental issue is this: compute, much like transit bandwidth, functions as a fungible utility. There is simply no structural margin in the undifferentiated middle layer to service billions of dollars in high-yield debt. If an AI startup is training a model on ten thousand GPUs, they generally do not care whose logo is on the data center door, so long as the uptime is reliable and the network interconnects are fast. The only reason these middle-layer providers appear financially sound today is because they are capturing a temporary scarcity premium. Over the last two years, every generative AI startup in the world was desperate for Nvidia's latest chips, and the major hyperscalers could not build out data centers fast enough to satisfy the demand. The middle layer stepped into this gap, charging sky-high hourly rental rates to desperate buyers.

The problem with building a business model on a scarcity premium is that it requires ongoing scarcity. In a market where the supplier is scaling production as fast as physics allows, that scarcity is fleeting. And yet, the data suggests the acute shortage is already evaporating.

Recent reports indicate that Amazon Web Services' AI GPU wait times have plummeted from up to 11 months to just weeks. This heavily tracked institutional data is directly corroborated by ground-level developer sentiment. Across the ecosystem, engineers are noting that securing large-scale compute allocations — once the existential bottleneck for any serious AI project — is now a standard, frictionless procurement process.

To put it another way, the acute shortage that allowed middlemen to charge massive premiums is over.

When wait times drop from nearly a year to a matter of weeks, the negotiating leverage shifts entirely back to the buyer. Compute is not literally zero marginal cost in the way a lit fiber strand is, because you still have to pay for the incremental electricity to run the server. But relatively speaking, the fixed capital expenditure required to buy an Nvidia H100 cluster so dwarfs the variable cost of the power to run it that the economic effect is identical. Once the server rack is installed and networked, the imperative is to rent it out at almost any price above the cost of electricity. A GPU sitting idle is burning cash, especially when it was financed with a term loan carrying double-digit interest rates. A middle-layer cloud provider cannot afford to let those servers sit empty, so they cut prices. Their heavily levered competitors, facing the exact same debt covenants, immediately do the same.

This dynamic leads to the inevitable structural squeeze. The fundamental lesson of the telecom bust remains undefeated today: when a highly levered middle layer attempts to sell a fungible, rapidly depreciating utility, debt is not a moat — it is a time bomb.

As capacity floods the market, compute rental prices will crash toward their absolute floor: the cost of electricity plus the baseline depreciation of the hardware. In a market where supply rapidly outstrips demand, the price of a purely fungible utility always trends toward its marginal cost. When rental yields collapse, the highly levered middle-layer companies will find themselves unable to service their debt. This will trigger a wave of defaults mirroring the 2002 telecom crash, leaving private credit lenders holding the bag.

But the physical servers, much like the transatlantic fiber of the late 90s, will not simply disappear into the ether; they will be repossessed. And waiting patiently on the other side of those bankruptcy proceedings will be the hyperscalers. Flush with cash, Amazon, Microsoft, and Google will simply buy the repossessed server racks out of bankruptcy for pennies on the dollar, turning a temporary competitor's private credit disaster into their own permanent infrastructural moat.

The structural direction is entirely clear: the middle layer of AI infrastructure will collapse under its own debt as the marginal cost of compute races toward zero. What is less certain, though, is exactly when the private credit music stops playing; as always in tech, the shift from paradigm to panic takes longer than you think.