The Evidence
Building Ahead of Demand
Every great infrastructure boom built more than the world needed, then waited for the world to grow into it. The waiting worked because the assets lasted decades. AI's core asset doesn't.
Updated
Infrastructure bubbles have a signature shape. Capital spendingMoney a company spends on long-lived physical things: buildings, machines, chips, data centers. Spent now, paid back (hopefully) over years.Full definition in the glossary climbs to a historically extreme share of the economy; capacity overshoots demand by multiples; first-wave investors are wiped out; and second-wave buyers, who pick the assets up at fire-sale prices, capture most of the value once demand finally arrives. It happened with British railways, American railroads, 1920s electric utilities, and 1990s fiber. The table below puts the AI data-center build-out in that lineage, with sources per column. Historical cells summarize the ranges in the underlying scholarship; see Sources. AI figures are estimates as of mid-2026 and marked as such.
| Dimension | UK rail (1840s) | US rail (1870s-90s) | Electric utilities (1920s) | Fiber (1998-2002) | AI data centers (2023-26) |
|---|---|---|---|---|---|
| Peak capex vs GDP | Authorized capital roughly 40 to 100 per cent of a year's GDP, depending on the accounting (see the Railway Mania case study); spend in high single digits of GDP in peak years (Odlyzko (opens in a new window)) | ~5-10% of GDP in peak build years (W&M (opens in a new window)) | Several % of GDP late-1920s (Kiesling (opens in a new window)) | ~1.0-1.2% of US GDP in 2000 (~$213B) (Richmond Fed (opens in a new window)) | Goldman Sachs models roughly 5% of US GDP by 2026 in its AI-infrastructure scenario, explicitly not a forecast, scoped to compute, data centers, and power spending ecosystem-wide (about $765B/yr in 2026, $1.6T/yr by 2031 in that model) -- not a hyperscaler-capex figure specifically (Goldman Sachs (opens in a new window)) |
| Overcapacity at peak | 20-30%+ of lines marginal or unprofitable | Similar share of mileage financially weak; waves of receivership | Substantial idle generation and transmission pre-WWII | ~95% of long-haul fiber dark c. 2001-02 (LA Times (opens in a new window)) | No clean "dark compute" metric exists; capex and revenue figures in this space use incompatible scopes (see the demand question), so no single "capex is N times revenue" multiple is defensible. Measured utilization outside the hyperscalers is Cast AI's own 5%-average sample (opens in a new window) of about 23,000 clusters it monitors, not a fleet-wide measurement; no audited hyperscaler figures exist. |
| Time for demand to catch up | Trunk lines fast; branches 20-40 years, or never | 5-10 yrs core routes; 20-40 yrs frontier lines | ~10-20 years (full use by 1940s-50s) | ~5-10 years for much of the fiber | Projected 5-10+ years if bull-case revenue arrives (Allianz (opens in a new window)) |
| Asset fate after the bust | Restructurings; assets absorbed at deep discounts | Foreclosure reorganizations; consolidators bought cheap | Holding companies collapsed; plants kept running under new owners | Networks sold for cents on the dollar in bankruptcy | Not yet written; history suggests strong balance sheets would absorb weak builders |
| Who captured the value | Second-wave owners and the whole economy | Consolidators and the economy | Regulated utilities, municipalities, consumers | Surviving telcos, cloud and streaming firms, users | Open question; the pattern favors later buyers and the application layer |
| Asset life vs waiting time | Track and rights-of-way: 50+ years. Long enough. | Same: decades. Long enough. | Plants: 30-60 years. Long enough. | Fiber strands: decades; electronics upgraded. Long enough. | Shells and power: decades. GPUs: 3-5 years. Maybe not long enough. |
Historical cells summarize the ranges in the underlying scholarship; see Sources. AI figures are estimates as of mid-2026 and marked as such.
The pattern in one sentence
Overbuilding is how infrastructure gets built: investors overpay, the economy inherits the surplus, and the people who buy after the crash collect the returns. In every prior case, the physical assets outlived the wait. Rails lasted fifty years. Fiber laid in 1999 carries Netflix today. The asset's patience is what made "demand will catch up" eventually true.
The one structural difference that matters most
AI's build-out splits into two kinds of asset. The buildings, power hookups, and cooling, roughly half the cost, are patient assets like rail beds and conduit. The GPUs inside are not. High-end accelerators are not patient assets: accounting useful lives now run 4 to 6 years per the hyperscalers' own filings (see the demand question for the individual companies), and 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.
That compresses the window. Fiber could wait a decade for demand; depreciatingSpreading the cost of an asset over its useful life in the accounts. Stretch the assumed life and reported profits go up, with no change in the business.Full definition in the glossary GPUs must earn their cost back inside a single hardware generation, or be replaced with fresh capital spendingMoney a company spends on long-lived physical things: buildings, machines, chips, data centers. Spent now, paid back (hopefully) over years.Full definition in the glossary before the first round has paid off. If demand arrives on the bulls' schedule, the distinction never matters. If demand arrives on the historical schedule, 5 to 20 years, it is the difference between owning dark fiber and owning warehouses of obsolete silicon. This is the single strongest structural argument on the bear side of the ledger, and it is why the GPU depreciation accounting controversy is more than a bookkeeping spat.
Where the AI numbers stand, mid-2026
Combined hyperscaler capex was roughly $159 billion in 2022. Bank estimates for 2026 cluster around $600 to 750 billion a year. The $1.6 trillion by 2031 figure sometimes paired with it is Goldman Sachs' own AI-infrastructure model (opens in a new window), explicitly framed as a scenario, not a forecast, and scoped to compute, data centers, and power spending ecosystem-wide, anchored to Nvidia's forward data-center revenue as a proxy -- not to hyperscaler capex specifically; using it as a hyperscaler-capex number overstates what the model measures. No named, dated report with disclosed methodology was located behind a $140 billion global generative-AI-revenue figure for 2026; estimates in that space vary by more than double depending on scope and horizon, so the figure is dropped here rather than restated without a source. Identifiable AI revenue, company by company, is on the demand question page; whichever way the ecosystem estimate is scoped, it is a small fraction of the spending side. Spending several times current revenue is not proof of a bubble; it is what building ahead of demand looks like, every time, in both the stories that ended well for investors and the ones that didn't. Which way this one breaks depends on the demand side, and on nothing going wrong in the meantime.