---
type: Mechanism
title: Infrastructure Overbuild
description: Every great infrastructure boom built more than the world needed; the assets' long lives let demand catch up. GPUs break that pattern.
resource: https://aibubblequestion.com/building-ahead-of-demand
tags: [infrastructure, capex, overbuild, depreciation, mechanism]
sources:
  - id: goldman-sachs-ai-buildout-scale
    title: "Goldman Sachs on AI build-out scale"
    resource: https://www.goldmansachs.com/insights/articles/tracking-trillions-the-assumptions-shaping-scale-of-the-ai-build-out
    author: "Goldman Sachs"
  - id: richmond-fed-telecom-capex
    title: "Richmond Fed on telecom capex"
    resource: https://www.richmondfed.org/~/media/richmondfedorg/publications/research/economic_quarterly/2003/fall/pdf/wolman.pdf
    author: "Federal Reserve Bank of Richmond (Wolman)"
  - id: odlyzko-railway-capex
    title: "Odlyzko on railway capex"
    resource: https://www-users.cse.umn.edu/~odlyzko/doc/mania01.pdf
    author: "Andrew Odlyzko"
generated:
  by: "claude/opus-5"
  at: 2026-08-30T00:00:00Z
status: stable
stale_after: 2027-02-28T00:00:00Z
verified:
  - by: "claude/opus-5"
    at: 2026-08-30T00:00:00Z
---

# The pattern across five booms

| Boom | Peak capex vs GDP | Overcapacity at peak | Demand catch-up | Asset life |
| --- | --- | --- | --- | --- |
| UK rail (1840s)[^odlyzko-railway-capex] | authorized capital roughly 40 to 100 per cent of a year's GDP, depending on the accounting (see railway-mania); spend high single digits/yr | 20-30%+ of lines marginal | 20-40 years | 50+ years |
| US rail (1870s-90s) | ~5-10% of GDP in peak years | similar marginal share | 5-40 years | 50+ years |
| Electric utilities (1920s) | several % of GDP | substantial idle capacity | 10-20 years | 30-60 years |
| Fiber (1998-2002)[^richmond-fed-telecom-capex] | ~1.0-1.2% of US GDP (2000) | ~95% of long-haul fiber dark | 5-10 years | decades |
| AI data centers (2023-26)[^goldman-sachs-ai-buildout-scale] | ~5% of US GDP by Goldman's global AI-infrastructure scenario model (explicitly not a forecast; scoped to compute, data centers and power ecosystem-wide, not hyperscaler capex alone) | no clean metric; capex and revenue figures in this space use incompatible scopes (see ai-demand-side), so no single "capex is N times revenue" multiple is defensible | projected 5-10+ years | shells: decades; GPUs: accounting lives now 4-6 years per filings, economic/resale life contested |

# Regularities

* First-wave investors are largely wiped; second-wave buyers of distressed assets and
  the broader economy capture the value.
* The pattern "worked" historically because assets outlived the wait for demand.
* The AI difference: accounting useful lives for AI infrastructure now run 4-6 years
  at the hyperscalers, per their own 10-Ks (see
  [the AI demand side](/current/ai-demand-side.md) for the filings). No market-wide,
  transaction-level dataset establishes a resale or economic depreciation rate for
  GPUs specifically; the unsourced "30-60% a year" figure this page previously
  carried has no document behind it and is dropped rather than restated. If
  accounting life and resale value diverge from economic life the catch-up window
  shortens, but by how much is not established from any primary document located
  this pass.


[^odlyzko-railway-capex]: Odlyzko on railway capex
[^richmond-fed-telecom-capex]: Richmond Fed on telecom capex
[^goldman-sachs-ai-buildout-scale]: Goldman Sachs on AI build-out scale