Friday, November 20, 2026

Compare AI debt to dot-com bubble.

AI debt differs from the dot-com bubble primarily in financing structure and scale of physical infrastructure, even as both involve over-optimism about a transformative technology. The ECB and others explicitly separate the comparisons: the scale of AI capital spending resembles (or exceeds) the dot-com/internet boom, while the debt-heavy, opaque financing more closely echoes the subprime crisis or the telecom debt binge with in the broader late-1990s tech cycle.

Financing Mix : Equity vs. Debt

  • Dot-com era (late 1990s–2000) : Heavily equity-financed. Much of the speculation involved internet software/services startups with light physical assets, funded by venture capital and public equity offerings. Losses fell mainly on shareholders (many stocks fell 60–90%). The associated telecom infrastructure build-out did involve substantial debt (~$500+ billion in U.S. telecom bonds/capex in the late 1990s, peaking with sector debt around $300–1 trillion by the early 2000s), including vendor financing (equipment makers like Cisco lending to customers). Still, the overall tech boom’s equity dominance helped limit systemic fallout—the recession was mild.
  • AI today : Shifted rapidly toward debt and off-balance-sheet commitments as hyperscaler free cash flow turned negative. On-balance-sheet debt for major AI firms is heading past $800 billion by end-2026; off-balance-sheet leases, take-or-pay contracts, residual-value guarantees (especially on GPUs), and SPVs exceed $3–4 trillion in notional value. Circular financing is prominent (e.g., Nvidia equity stakes/guarantees enabling customers to buy its chips). Private credit plays a larger role. This creates more leverage, maturity mismatches (short-lived chips vs. longer debt), and opacity—features the ECB links more to subprime than pure equity-funded tech.

Key implication : Dot-com losses were mostly equity wipeouts. AI debt raises the chance of credit losses, forced sales, and wider spillovers if revenues lag.

Scale of Investment Relative to the Economy

AI infrastructure spending is larger as a share of GDP than the internet/telecom boom :

  • Dot-com/internet boom (roughly 1996–2003) : ~1.1% of U.S. GDP annually on average.
  • AI projections (2025–2032) : Often estimated at 3–3.6%+ of U.S. GDP annually (data centers, power, chips), exceeding even the 19th-century railroad boom in some measures (~2.2%). Hyperscaler capex alone is on track for $1.2–1.4 trillion in 2027.

Much AI hardware is imported, so the direct domestic GDP boost is smaller than headline spending, but the financial claims are still large.

Nature of the Assets and Circular Risks

  • Dot-com : Mix of pure software/platforms (low tangible assets) and telecom fiber/networks. Overcapacity left “dark fiber” that was later reused cheaply. Vendor financing created some circularity.
  • AI : Extremely capital-intensive physical build-out (GPUs with short useful lives of ~1–3 years economically, data centers, power). Residual-value guarantees and supplier-backed financing amplify circularity and collateral risk if secondary GPU markets weaken. Concentration is higher (a handful of hyperscalers + Nvidia dominate).

Valuations and Market Concentration

Both eras saw elevated tech valuations and concentration. Dot-com peak S&P 500 CAPE/P/E was higher in some metrics, today’s top AI names form an even larger share of the index (~40% for the top 10 vs. lower in 2000). However, current leaders generate substantial real earnings and cash flows from non-AI businesses, unlike many unprofitable 1999–2000 internet pure-plays.

Potential Consequences of a Correction

  • Dot-com bust : Sharp equity crash, mild recession, limited banking-system damage because equity absorbed most losses. Productivity gains from the internet continued.
  • AI scenario : A pure equity correction could resemble a milder version of 2000–2002. But heavy debt + opacity raises the risk of credit-market stress, project cancellations, and broader tightening—more subprime-like transmission, though hyperscalers’ balance-sheet strength and diversified revenues provide buffers that weaker telecom CLECs lacked. Regulators (Fed, ECB) flag this as a financial-stability watch item rather than an imminent crisis.

Summary Table of Key Differences

Aspect

Dot-Com / Telecom (late 1990s–early 2000s)

AI Boom (mid-2020s)

Primary financing

Mostly equity (VC/public markets) - telecom had heavy debt

Increasingly debt + large off-balance-sheet commitments

Investment scale (GDP)

~1.1% annually

~3–3.6%+ annually (projected)

Asset type

Software + fiber networks

GPUs, data centers, power (short asset lives)

Circularity

Present (vendor financing)

More pronounced (supplier guarantees/equity)

Loss absorption

Mainly equity holders

Equity + credit/private markets

Recession impact

Mild

Potentially broader if credit tightens

Bottom line : The AI debt build-up shares the overinvestment optimism of the dot-com era but carries greater leverage and opacity risks that make pure equity-style outcomes less certain. A correction could still be manageable given the strength of the core players, but the shift to debt financing is the feature that most distinguishes it—and the one regulators are watching most closely. Outcomes will hinge on whether AI revenues materialize at the scale currently priced into both equity and the debt supporting the infrastructure.