Capex supercycles
Is the AI buildout the biggest investment wave in history? This third piece sets today's AI capex supercycle against the ones that came before — the Chinese manufacturing supercycle of 1980–2015, the great Western industrialisation of 1870–1913, and the railway and canal manias before that — measured the only way that makes them comparable across two centuries: GDP-to-GDP. It builds on the factory census and capital-stock pieces. Editorial estimates; methodology in the appendix.
The only fair comparison is GDP-to-GDP
A dollar in 2025 and a pound in 1847 are not comparable; a share of GDP is. So the question is not “how many dollars” but “how much of the economy's annual output did this investment wave absorb at its peak?” On that measure — and it is the honest one — the AI buildout is large but not unprecedented, while the Chinese supercycle is in a category by itself.
Peak annual capex as a share of the relevant economy's GDP, log scale, editorial estimates. GDP-to-GDP, the AI buildout sits near the canal- and railway-mania band; the Chinese supercycle is in a class of its own.
The six waves
| Supercycle | Era | Where | Peak capex | Financing |
|---|---|---|---|---|
| AI / data-centre buildout | 2023–2030E | US-centric | 1.3% | Hyperscaler free cash flow, private credit, debt |
| British canal mania | 1790s | Britain | 1.5% | Local equity subscriptions, turnpike trusts |
| US railroadisation | 1870–1900 | United States | 5% | Bonds, land grants, heavy foreign (British) capital |
| UK railway mania | 1840s | Britain | 6.5% | Joint-stock equity bubble; mass retail speculation |
| Western great industrialisation | 1870–1913 | US, Germany, UK | 13% | Retained earnings, universal banks, bond markets |
| Chinese manufacturing supercycle | 1995–2015 | China | 43% | State banks, SOEs, local-government land finance |
What the comparison shows
- AI is a railway mania, not a China supercycle — for now. At ~1.3% of US GDP, the data-centre buildout sits squarely in the historical band of great national investment booms — comparable to the 1840s railway mania and the US railroadisation of the 1880s. It is not remotely the ~43% of GDP that China sustained for two decades.
- But it is faster and more concentrated. The railways were built by thousands of companies over decades; the AI buildout is being financed by a handful of firms in a handful of years, out of their own cash flow. Concentrated capital moves faster — and fails faster.
- China's supercycle is the true outlier. No investment wave in recorded history — not the railways, not Western industrialisation, not AI — comes within an order of magnitude of a whole large economy running 40%+ fixed investment for twenty years. On the log axis above, it is a rung on its own.
- Every supercycle over-builds. Canals, railways and Chinese factories all ended in overcapacity and write-downs — while leaving the durable asset standing. The open question for AI is which part it repeats: the crash, or the network that outlives it.
Why it still matters that AI is “only” a railway mania
Being in the railway-mania band is not small. Those booms reorganised their economies, minted and destroyed fortunes, and left infrastructure that compounded for a century. The claim here is narrower and more precise: measured GDP-to-GDP, the AI capex wave is a serious national investment boom of a historically familiar size — not, as it is sometimes described, the largest capital buildout ever. That title still belongs, by a wide margin, to the Chinese factory supercycle this series has been tracing.
Appendix — methodology
The measure
Each cycle is scored on peak annual capex as a share of GDP — the single year (or short peak window) in which the investment absorbed the largest fraction of the relevant economy's output. GDP is the national economy for the national booms (Britain, US, China) and US GDP for the AI figure, with a world-GDP figure given alongside. Peak share, not cumulative, is used because it is the most consistently recoverable number across two centuries.
The AI figure
Built up from disclosed and guided hyperscaler capex (Amazon, Microsoft, Google, Meta) plus other data-centre and chip investment, on the order of $300–500bn/yr by 2025 and rising. Against ~$29tn of US GDP that is ~1.0–1.5% (the ~1.3% used here); against ~$106tn of world GDP it is ~0.4%. A cumulative $3–5tn through 2030 is a scenario, not a forecast.
The historical figures
- Chinese supercycle. China's gross fixed capital formation ran ~40–45% of GDP for roughly two decades (World Bank); manufacturing alone ~10–13%. The ~43% here is the total fixed-investment share, the figure that makes it the outlier.
- Western industrialisation & railways. Investment shares for the US, Germany and Britain, 1840–1913, from long-run economic-history series (Maddison-tradition national accounts). Railway investment peaked at ~5–7% of UK GDP in the late 1840s and absorbed a large share of US capital formation in the 1880s.
- Canal mania. The 1790s British canal boom is included as the origin point of the speculative-infrastructure template; its GDP share was small (~1–2%) but its structural effect large.
Caveats
Peak-share comparisons flatter short, intense booms and understate long, steady ones; a cumulative or capital-stock comparison would rank the cycles somewhat differently (and would move Western industrialisation up). The AI figure is the softest on the page — it is a live, fast-moving number and the denominator choice (US vs. world GDP) changes it threefold. The qualitative ranking — China in a class of its own, AI in the railway-mania band — is robust to those choices.
Sources consulted
- World Bank. Manufacturing value added (% of GDP), gross fixed capital formation.
- China NBS. Industrial enterprises 'above designated size', fixed-asset investment.
- OECD. Capital stock, consumption of fixed capital, investment by industry.
- Hyperscaler filings. Amazon, Microsoft, Google, Meta capex guidance for the AI buildout.
- Maddison Project / economic history. 19th- and early-20th-century investment shares (railways, industrialisation).