Research Paper: Fiber 1999, Compute 2026

What the Telecom Overbuild Tells Us About the AI Infrastructure Cycle

The AI build-out rhymes with the fiber overbuild. The losses will land somewhere else.

The closest historical analogue to the AI build-out is the 1996–2002 fiber overbuild, not the consumer dot-com bubble. That cycle built capacity ahead of demand and financed it with debt, vendor credit and creative accounting. It ended with more than 60 carrier failures, while the fiber itself carried two decades of internet growth. Every fiber-era leader fell further than the Nasdaq’s 78%: Cisco fell 89%, and the leveraged builders went to zero.

The AI cycle repeats three of fiber’s features: debt-funded capacity, circular supplier financing, and accounting that flatters near-term earnings. Its core is far stronger, though. Four hyperscalers guided to roughly $700–725 billion of 2026 capex, against about $97 billion of US carrier capex in 2000, and the buyers are among the most profitable companies in history. The fragility sits at the edges. The bottleneck is also reversed: 30–50% of the 16 GW of US data-center capacity scheduled for 2026 faces severe delay or cancellation. Unlike dark fiber, a GPU waiting for a building loses value every quarter.

The paper does not call the timing of a correction. It maps the order in which the triggers are likely to fire (credit first, chips last), the tripwires that signal each step, and where a long-horizon investor should stand.

Key findings

• Three rhymes with 1999: debt-funded capacity ahead of proven demand, circular supplier financing, and depreciation assumptions that flatter earnings [Abstract].

• Two decisive differences: the core buyers are highly profitable, and demand is measured in revenue rather than projections.

• One reversal: the constraint is power and buildings, not demand. Chips risk arriving faster than data centers can house them.

• Where losses land: on leveraged intermediaries, lenders against short-lived collateral, and capital priced for perpetual growth. OpenAI is the central node, and Oracle is the large credit most exposed to it.

• Thinner shock absorbers: with inflation at 3.7%, a Fed raising rates and US interest costs above $1 trillion a year, policy cannot cushion a correction as it did in 2001.

• A conditional Gulf advantage: the Gulf’s advantage is real as an allocator of capital abroad, but uncertain as a host of compute while the war continues.

What would prove us wrong

If AI revenue grows faster than capex through 2027, and the leveraged builders refinance at lower spreads, there is no reset to buy.

Table of contents