The myth of infinite demand
Every bubble needs one number that makes the spending look conservative, and this one had a very good number: internet traffic was doubling every hundred days. That works out to roughly a tenfold increase every year. At that rate no amount of fibre is too much fibre, and any carrier that hesitates is simply choosing to be run over.
The figure originated with UUNet, the backbone operator later owned by WorldCom, and travelled outward from there as industry messaging. It was repeated by WorldCom, echoed in a 1998 U.S. Department of Commerce report, and then absorbed wholesale by analyst notes, earnings calls, investor presentations and the press until it stopped being a claim and became a background fact.
It was wrong. In October 1998 Kerry Coffman and Andrew Odlyzko published The size and growth rate of the Internet in First Monday, in which they actually went and measured the thing. Public internet backbone traffic was growing at about 100% per year from late 1996 through April 1998 — a doubling once a year, not three and a half times a year. The spectacular growth rates everyone was quoting had been real for a brief window in 1995 and 1996, when the network was small enough that any absolute increase looked enormous in percentage terms. Their conclusion was that sustained 1,000% growth claims were inaccurate.
The paper is free, was published four years before the collapse, and has been cited a few hundred times since. The distance between "doubling every 100 days" and "doubling every 365 days" is the entire buildout.
The physical and financial excess
On the strength of the phantom number, U.S. carriers raised roughly $1.9 trillion in debt and equity between 1996 and 2001, a figure that traces back to American Enterprise Institute work on the Telecommunications Act and has been recycled in later commentary. Estimates of the same episode vary a lot depending on what you count: Cuban, comparing the AI buildout to the fibre glut, put the late-1990s spend at about $500 billion for laying fibre nationwide. That is a narrower measure — cable in the ground rather than every dollar of capital raised across the sector — and the gap between the two is a useful reminder that nobody has a tidy audited total for this.
What is not in dispute is the outcome. By 2000, less than 5% of installed fibre was actually being used. The rest was dark: glass in a conduit, fully paid for, carrying nothing. Bandwidth prices collapsed by nearly 90% as supply outran demand, which is exactly what happens when every carrier simultaneously builds the network that one company's forecast said the country would need.
The emblematic case is Global Crossing. Its network cost roughly $15 billion to build. The company filed for Chapter 11 on 28 January 2002, listing total assets of $22.4 billion and debts of around $12.4 billion, at the time the fourth-largest bankruptcy in U.S. history. Founder Gary Winnick took roughly $734 million out of the company before it collapsed. The fibre stayed in the ground. The $734 million did not.
The vendor financing loop
The boom lasted longer than the underlying demand because the equipment makers propped up their own customers. Lucent, Nortel and Cisco lent carriers the money to buy Lucent, Nortel and Cisco gear. The sale booked as revenue immediately; the credit risk stayed behind on the vendor's balance sheet as a receivable or a loan, where it looked like an asset.
The sums were not small. Lucent committed about $8.1 billion. Nortel committed about $3.1 billion, of which roughly $1.4 billion was drawn. Cisco had about $2.4 billion in customer loans. A McKinsey estimate put the combined vendor-financing exposure of nine suppliers at about $25.6 billion by the end of 2000.
The terms tell you what the competition for these deals was like. Loans reportedly reached as much as 130% of the cost of the equipment — the vendor handing over more money than the gear was worth, so the customer had working capital left over — and were often unsecured or tied to promises of future purchases. This is a structure that works beautifully as long as the borrower can refinance somewhere else later, and only ever needs to fail once, at the end.
When outside capital dried up in 2000 and 2001, it stopped working all at once. The carriers failed, the loans became bad debts, and the market for selling those receivables onward disappeared, so the vendors were holding the credit risk and facing collapsing demand at the same moment. Lucent took more than $2 billion in bad-debt provisions in a single year. Nortel's customer financing and related receivables reached $5.2 billion at the end of 2000, and its bad-debt allowance rose to $890 million by the end of 2001. Lucent, one of the great American industrial names, was absorbed into Alcatel in 2006.
The aftermath, and the dark fibre
Between 2000 and 2003, 47 telecom firms went bankrupt. Estimates of the shareholder value destroyed run to about $2 trillion. The equity was wiped out, the debt was wiped out, and the executives who had already sold were fine.
The fibre did not evaporate. Distressed investors bought long-haul networks out of bankruptcy for, in the phrase that recurs in every retelling, a few cents on . Freed of the debt that had been raised to build it, the glass became economically viable at prices the original owners could never have made work, and it became the backbone that carried streaming video, cloud computing and everything after. Who exactly bought which network, and for how much, is harder to pin down than the story implies — we could not establish reliable buyer-by-buyer figures for Global Crossing, 360networks or Williams Communications, and we are not going to guess.
The lesson people draw from this is usually the cheerful one: the infrastructure survives, so the boom was worth it. That is true for the infrastructure. It is not true for the capital. The fibre was a gift to whoever bought it second. The people who paid for it got nothing.
Why it matters to this crash
This is the comparison being made, explicitly and by name, about the .
Torsten Sløk at Apollo Global Management is the central source. His framing, in the widely circulated Daily Spark note "The AI capex boom is building twice as fast as the housing boom", is that the telecom and fibre buildout topped out at 1.2% of GDP in 2000 before collapsing and contributing to the recession that followed, and that datacentre capex is on a much steeper path: 0.3% of GDP in 2019, 0.6% in 2023, 1.4% in 2025, and a projected 3.1% in 2027. Consensus expects capex to run at roughly 3% of GDP every year from 2027 to 2029. On Apollo's cumulative-change measure, telecom capex rose only 0.4 percentage points of GDP across the late 1990s; the AI buildout rises 1.7 points in the two years from 2025 to 2027 alone.
Michael Burry has shared Sløk's charts. Mark Cuban has made the fibre-glut comparison in his own terms. When people who agree on almost nothing else all reach for the same 1999 analogy, it is at least worth knowing what actually happened in 1999.
The honest version of the parallel is not "therefore this ends the same way". It is narrower and harder to dismiss: the telecom buildout was rational given its demand forecast, the demand forecast came from a vendor's marketing, and the person who published the correct number in 1998 was ignored for four years.
What would make this dangerous
Each of these has a 1999 twin.
A single unverified demand metric doing all the work. If the case for a given datacentre rests on a compute-demand growth rate that originates with someone selling compute, and no independent measurement of end-user demand exists, that is the doubling-every-100-days structure exactly. The tell is not that the number is high. It is that nobody outside the vendor can check it.
Vendor financing that shows up as revenue. Suppliers lending customers the money to buy their own product, booking the sale immediately and carrying the credit risk as a receivable. In 2000 that reached $25.6 billion across nine suppliers, with some loans at 130% of equipment cost. The equivalent today would be chip and infrastructure vendors funding their own buyers' purchases through equity stakes, prepayments or backstops. The number that matters is not the headline commitment but how much has been drawn, and what happens to it if the borrower cannot refinance elsewhere.
Utilisation. The single most damning statistic of the fibre era was that less than 5% of installed capacity was lit. The AI equivalent is the share of built GPU capacity actually running paid workloads rather than sitting idle or on internal research. If that figure is being reported at all, it is being reported by the people who own the GPUs.
Price collapse in the underlying commodity. Bandwidth fell nearly 90%. Falling per-token or per-GPU-hour prices are wonderful for buyers and fatal for anyone who financed the capacity on the assumption of stable pricing.
Insider extraction ahead of the reckoning. $734 million left Global Crossing before the filing. Concentrated founder or executive selling into a buildout financed with other people's debt is not proof of anything, but it is the pattern.
What would make it not dangerous is equally specific: independently verified demand growth, capacity that lights up as fast as it is built, and capex funded by operating cash flow rather than by the seller of the equipment. Those are all observable. They are worth observing before 2027, when the spending is projected to hit more than twice the GDP share the telecom boom ever reached.