AI Capex vs. the Dot-Com Buildout: A Visual Comparison

By 2002, most of that fiber was dark. Unused. The companies that built it were bankrupt or restructuring. Over $2 trillion in market capitalization evaporated. The technology was real — we use that fiber today, it carries the internet you’re reading this on — but the capital deployed to build it vastly exceeded the revenue it could generate in any reasonable timeframe. The fiber got used eventually. The investors who funded the buildout mostly didn’t recover.

The parallels to AI infrastructure are uncomfortable. Not because AI is fake (it isn’t) but because the financial dynamics of the buildout follow a pattern that has a name in economic history: the capital cycle. Transformative technology attracts capital. Capital funds buildout beyond what near-term demand can absorb. The gap between investment and revenue creates financial stress. The stress eventually corrects, often violently.

The scale comparison makes the discomfort concrete. Adjusted for inflation, the telecom/fiber buildout peaked at roughly $150 billion per year in today’s dollars. AI infrastructure capital expenditure is on track to exceed $450 billion in 2026. That’s 3x the peak rate of the dot-com buildout at the equivalent point in the cycle — roughly 5 years after the initial wave of investment began.

The revenue comparison is equally stark. At the peak of the telecom bubble, fiber network utilization was estimated at 2.5–5% of installed capacity. The demand existed in projections, not in contracts. Today, JPMorgan estimates that AI capex requires $650 billion per year in revenue to earn a 10% ROI. Current generative AI revenue is $37 billion. The utilization rate is higher than telecom’s was — AI products have real customers paying real money — but the gap between what’s needed and what exists is structurally similar.

The debt dynamics rhyme too. The telecom buildout was heavily debt-financed, with companies like WorldCom and Global Crossing leveraging future revenue contracts to borrow against projected growth. Today, AI infrastructure debt issuance exceeded $200 billion in 2025, with companies like CoreWeave carrying 9x EBITDA leverage against revenue concentrated in a handful of customers. The structured credit parallels are even more direct: JPMorgan projects data center paper will reach 7-10% of combined ABS/CMBS issuance by 2026-2027, creating the same kind of concentrated sector exposure that amplified the 2008 crisis.

There are important differences. The hyperscalers funding the AI buildout are vastly more capitalized than the telecom companies of 2000. Alphabet, Microsoft, and Amazon have balance sheets that can absorb losses the telecom operators never could. The technology deployment cycle is faster — AI products reach customers in months, not the years it took to light fiber. And the customer base for AI is broader, spanning enterprise software, consumer products, and cloud services rather than telecom carriers selling to other telecom carriers.

These differences mean the AI cycle is less likely to produce the kind of sudden, catastrophic collapse the telecom bubble did. What they don’t mean is that the capital deployed today will generate the returns being priced into the market. The fiber got used. The investors lost their money anyway. The question is whether AI infrastructure follows the same path: technology that succeeds on a timeline that bankrupts the capital structure built to fund it.

History doesn’t repeat, but capital cycles rhyme. The AI Infrastructure Stress Index tracks whether this cycle is rhyming with its predecessors. Right now, the data says it is.