MTS

Wiki/History

AI Capex in Historical Context — Prior Buildouts Compared

Last updated: May 2026

The pattern recognition challenge

Almost every essay on the AI buildout reaches for one of two analogies. The bullish framing invokes electrification or the railroads — slow-burn general purpose technologies whose returns showed up over a generation, vindicating the early speculators. The bearish framing invokes the 1996–2001 telecom buildout — a debt-financed overbuild that lit a single-digit percentage of installed fiber for years afterward and bankrupted most of the operators that built it.

Both analogies are real, and both are partial. The point of this entry is to put actual numbers next to each one, so the reader can decide which features of the historical cases the AI buildout shares and which it does not. Three variables matter most:

  1. Scale relative to the economy (capex as a share of GDP, of S&P 500 capex, of the total private investment base).
  2. Capital structure (debt vs. equity, balance sheet quality of the builders, recourse on the assets).
  3. Realized utilization (how long it took for installed capacity to be lit, monetized, or absorbed into productivity statistics).

When you sort the prior cycles along these three axes, the AI buildout is not unprecedented in scale — railroads were larger — and it is not unprecedented in absolute dollars — telecom was almost as large after inflation. What is unusual is the combination of (a) very large absolute dollars, (b) concentrated in a handful of hyperscaler balance sheets with low leverage, (c) deployed on equipment with a 4–6 year useful life rather than 30–50 year right-of-way assets. That combination has no exact precedent, which is what makes the historical exercise interesting rather than dispositive.

Telecom 1996–2001: nearest analog

The 1996–2001 telecom buildout is the cleanest comparison, partly because it is recent enough that the numbers are well measured and partly because the underlying logic — bandwidth as the substrate for an emerging digital economy — is structurally similar to compute as the substrate for AI.

Aggregate spend. In the five years following the Telecommunications Act of 1996, US telecom companies invested more than $500 billion, the great majority of it debt-financed, in fiber optic plant, central office switching, and new wireless networks. Capital spending by facilities-based US carriers nearly tripled between 1996 and 2000, reaching a peak of roughly $120 billion in 2000 (about $213 billion in 2025 dollars). The cumulative network buildout produced roughly 80–90 million route-miles of fiber in the United States alone, with comparable buildouts in Europe and Asia.

Capital structure. US telecom companies issued more than $500 billion in new bonds between 1996 and 2001, and FCC Chairman Michael Powell testified in 2002 that the industry owed roughly $1 trillion, "much of which will never be repaid." Bondholders in the subsequent restructurings recovered a little over 20 cents on the dollar on average.

The collapse. Capex peaked in 2000 at roughly $120 billion and fell by more than 50% over the next three years. Average broadband capex in 2002–2008 was about $65.8 billion. The equity wipe-out was even more dramatic: telecom stocks lost more than $2 trillion in market value between 2000 and 2002, a CAGR of roughly –50% over the three-year period.

The bankruptcy list. Global Crossing filed Chapter 11 in January 2002 with $22.4 billion in assets and $12.4 billion in debt — at the time the fourth-largest bankruptcy in US history. WorldCom followed in July 2002 with $107 billion in assets and $41 billion in debt, the largest bankruptcy on record up to that point. 360networks, Williams Communications, McLeod USA, Adelphia, and dozens of CLECs failed. Qwest survived but was investigated by the SEC; Level 3 survived but its equity went to roughly two dollars from a peak of $130.

The utilization shortfall. This is the number most often cited in AI capex commentary, and the underlying source matters. Estimates of dark (unlit) fiber as a share of installed plant in 2001–2002 cluster around 95% — that is, roughly 5% of the strands installed in the boom were carrying traffic three years post-peak. By late 2005, four years after the bust, roughly 85% of fiber strands were still dark. The bandwidth glut compressed transit prices by about 90%, which in turn made the next generation of bandwidth-hungry services — YouTube, Netflix streaming, Skype, the migration of enterprise IT to the cloud — economically viable in the latter half of the 2000s. By 2010 the overhang was largely worked through.

Was it justified in retrospect? This is the heart of Carlota Perez's "Schumpeterian over-investment" thesis. In Technological Revolutions and Financial Capital (2002), Perez argues that every great surge of technology proceeds in two halves — an installation phase, dominated by financial capital and ending in a bubble, and a deployment phase, dominated by production capital that uses the cheap infrastructure left behind. On Perez's reading, the telecom bust was not a verdict that fiber was a bad idea; it was the price of compressing thirty years of physical-layer build into five. The fiber strands lit between 2005 and 2015 powered a real economy. The first wave of operators paid the bill; the second wave got the assets at ten cents on the dollar.

The Perez frame matters for AI because it suggests the right question is not "will the equipment be used" but "will the equity holders who paid for the equipment see the use." Those are very different questions, and the telecom case decisively answers them differently.

Railroads, 1865–1893: the deepest precedent

Railroads are the canonical "speculative infrastructure that ultimately served the economy" case. US rail mileage grew from roughly 53,000 miles in 1870 to nearly 200,000 by 1900. In peak construction years rail investment is estimated to have absorbed 10–20% of US GDP — a figure that dwarfs anything since, including the AI buildout at roughly 2% of GDP.

The financial carnage was repeated and severe. The Panic of 1873 took down 55 US railroads in November and another 60 by September 1874. The Panic of 1893 — caused in significant part by railroad over-leverage — bankrupted more than 150 railroads in a year, including the Atchison, Topeka and Santa Fe, the Northern Pacific, the Union Pacific, and effectively every major western trunk line. By 1894, roughly a quarter of US rail mileage was operating under receivership.

Robert Fogel's Railroads and American Economic Growth (1964) used a counterfactual "social savings" methodology to argue that the direct economic contribution of railroads, even at the 1890 mileage peak, was less than 5% of GNP — much smaller than the prevailing historical wisdom assumed. Albert Fishlow's parallel work put the 1890 figure closer to 15% of GNP. Modern reworkings of Fogel's exercise (Donaldson and Hornbeck, 2016; Hornbeck and Rotemberg, 2024) using county-level data find substantially larger general-equilibrium gains than Fogel's partial-equilibrium estimate — agricultural land values fell roughly 60% in counties that lost rail access in counterfactual exercises, and aggregate productivity gains were on the order of 25%.

For our purposes, the relevant point is the dissociation between the operators' financial returns and society's economic returns. The first generation of railroad equity holders was largely wiped out. The mileage was not. By 1920 the network was carrying the freight that built American industry, on tracks paid for by 1880s bondholders. The Perez deployment-phase analogy was first articulated about this cycle.

Electrification, 1890–1930: the productivity-lag case

Electrification is the canonical "the productivity gains showed up two decades late" case. Paul David's 1990 paper The Dynamo and the Computer — the founding document of this literature — observed that the incandescent lamp was patented in 1880, but by 1900 only about 3% of US residences had electric lighting and electric motors accounted for under 5% of factory mechanical drive. Both figures crossed 50% only in the 1920s. Aggregate manufacturing productivity growth did not visibly accelerate from electrification until the second half of the 1920s, when the gains finally showed up as a roughly 5% per year acceleration in manufacturing TFP.

David's central insight is that the lag was not a measurement artifact. Early electrified factories simply replaced their central steam engine with a central dynamo and kept the existing belt-and-shaft layout. The productivity dividend required a redesign — individual motors on individual machines, single-story layouts replacing multi-story buildings stacked around a central drive shaft, materials flow rebuilt around the new freedom of placement. That redesign took twenty years. Erik Brynjolfsson, Daniel Rock, and Chad Syverson generalized this into the "productivity J-curve" (NBER WP 25148, 2018) and applied it explicitly to AI: GPTs require large, often unmeasured, intangible complementary investments; measured productivity dips during the build, then accelerates when the intangibles bear fruit. The historical lag has been on the order of two decades, though there is reason to think the lag for software-mediated GPTs may be shorter.

Direct dollar comparisons for the electrification cycle are harder to construct because national accounts before 1929 are reconstructed rather than measured. Between 1902 and 1930 US electricity generation increased by more than a factor of 20, and the utility industry was the second-largest sector of US capital formation behind only the railroads. By 1930, roughly 90% of urban and non-farm rural homes had electrical service; only about 10% of farms did, which is what motivated the 1935 Rural Electrification Administration.

Semiconductor fabs: the only cyclical analog that ever fully repeats

Semiconductor capex is the closest analog to AI capex in physics — both are massive capital outlays on equipment with relatively short useful lives, building capacity that exhibits dramatic learning-curve cost declines. It is also the only one of these cycles that has repeated multiple times within living memory, so it provides usable priors on cyclical behavior.

The 1980s Japan-Korea DRAM buildout established the basic pattern: simultaneous capacity additions across Samsung, Hyundai, Toshiba, NEC, and Hitachi produced a 1985–1986 memory glut that wiped out US DRAM producers and forced the 1986 US-Japan Semiconductor Agreement. The 2000s Taiwan-Korea logic foundry buildout (TSMC, UMC, Samsung) produced a similar cycle around 2001 and again in 2008. The DRAM industry has run a roughly four-year boom-bust cycle since the 1990s with double-digit revenue swings in both directions.

Per-fab capex. A leading-edge fab today costs $15–25 billion. TSMC's three-fab Arizona complex represents more than $65 billion in committed capex; Intel's two-fab Ohio site is over $20 billion in private investment alongside CHIPS Act support; Samsung's Taylor, Texas site is around $17 billion; Micron's Clay, NY plan is over $100 billion across multiple fabs over fifteen-plus years. CHIPS Act direct funding awards include $6.6 billion for TSMC, $7.86 billion for Intel, $4.75 billion for Samsung, and $6.2 billion for Micron.

Industry-wide capex. Global semiconductor capex was roughly $185 billion in 2024 and is forecast at the $200 billion range for 2025, with TSMC alone at $38–42 billion in 2025 capex. AI demand has shifted the cycle: HBM (high-bandwidth memory) supply is tight enough that Samsung and SK Hynix have publicly committed to underbuild rather than chase the cycle, in marked contrast to prior cycles.

The lesson from the semi cycle for AI is less about absolute scale than about behavior: the operators who survive cycles are the ones who invest counter-cyclically, fund capex from operating cash flow rather than debt, and treat the asset as a multi-cycle business rather than a single-bet wager. That maps onto a specific subset of AI infrastructure investors.

AWS, Azure, and the first cloud buildout (2010–2020)

The "first cloud buildout" is the most underrated comparison in current AI capex commentary, because it is the case where the original investors did capture the returns, and at a scale very large for its era.

AWS capex was roughly $4 billion in 2014, scaling to about $20 billion in 2019. Microsoft Azure and Google Cloud capex grew on similar trajectories with a lag. Combined hyperscaler cloud-related capex from 2014 through 2020 totaled in the range of $400–500 billion. At the time, this was widely characterized in the same terms as the AI buildout is today: "they are spending more than they can ever earn back," "the unit economics don't work," "this is a return-on-invested-capital disaster."

By 2024 AWS was generating roughly $40 billion in operating income on segment revenue of $108 billion, with operating margins above 36% — a return on the cumulative cloud capex base that, by any reasonable accounting, retired the entire skeptic case. Azure and Google Cloud trail AWS by a few years but show the same trajectory.

What this teaches for AI is specifically that massive recurring capex by a small number of hyperscalers, funded from operating cash flow, is consistent with very attractive long-run returns — provided the underlying service achieves utility-like recurring usage. The question for AI capex is not "can a hyperscaler capex program be a good investment" — AWS settled that. The question is "is AI inference more like S3 storage (high utilization, recurring, sticky) or more like dark fiber in 2002 (commodity, oversupplied, no pricing power)."

The 2025–2026 AI buildout in context

Aggregate capex for the top five US hyperscalers (Microsoft, Google, Amazon, Meta, Oracle):

  • 2022: ~$162 billion
  • 2023: ~$127 billion (the AWS-led capex pause)
  • 2024: ~$256 billion (+63% YoY)
  • 2025: ~$443 billion (+73% YoY)
  • 2026 (consensus): ~$602–725 billion (+36–63% YoY)

Roughly 75% of 2026 hyperscaler capex is directly tied to AI infrastructure — GPUs, networking, datacenter shell, power. The remainder is conventional cloud and other infrastructure.

Scale benchmarks for 2026. At roughly $646 billion, total hyperscaler capex is about 2% of US GDP, comparable in real-dollar terms to the peak of telecom capex in 2000 (~$213 billion in 2025 dollars), but at three times the peak telecom level. Investment in tech equipment and software reached approximately 4.4% of US GDP in 2025, close to the dotcom peak share. Deutsche Bank estimated in late 2025 that AI-related capex contributed roughly half of measured US real GDP growth in the first half of 2025.

Capital structure. This is the variable that most cleanly distinguishes the AI buildout from telecom. The five hyperscalers entered 2025 with combined net cash of roughly $200 billion and trailing-twelve-month operating cash flow of more than $500 billion. Their 2025 capex was approximately 90% funded from operating cash flow with the balance from a mix of debt issuance and finance leases. By contrast, the 1996–2001 telecom buildout was approximately 80–90% debt-financed by issuers whose operating cash flow could not service the debt at any reasonable utilization assumption.

Where leverage is showing up in the AI cycle is in the second tier: neocloud GPU operators (CoreWeave, Crusoe, Lambda, Nebius), AI data-center developers, and special-purpose vehicles funding individual sites. These look much more like 1999 telecom in financial structure. They are also a small fraction of total AI capex.

Useful life. A 2000-vintage fiber strand is still in the ground and still usable in 2026 — call it a 30-year asset. A 2025-vintage Hopper GPU has an accounting useful life of 5–6 years and may be commercially obsolete in 3. This shortens the period over which any individual capex dollar must earn out, and it raises the depreciation drag on operating margins. It also means an AI capex overhang clears far faster than a fiber overhang did: a 50% glut in GPUs is mostly worked off in five years by physical retirement, whereas a 50% glut in fiber required nearly a decade of demand growth.

Side-by-side: the data table

Cycle Period Peak annual capex (nominal) Peak annual capex (2025 $) Peak as % of GDP Capital structure Operators' fate Long-run social return
US railroads 1865–1893 n/a (aggregate ~$10B+ in 1880s $) ~$300B+ in peak years 10–20% in peak years Mostly debt and equity, recurring panics ~25% of mileage in receivership by 1894 Very high (Donaldson-Hornbeck ~25% TFP); operator returns much lower
Electrification 1890–1930 n/a (utilities 2nd-largest sector) n/a est. 1–2% Mostly private utilities + bonds Industry consolidated; survivors profitable Very high but lagged ~20 years (Paul David)
US telecom 1996–2001 ~$120B (2000) ~$213B ~1.2% ~80–90% debt; ~$500B in new bonds WorldCom, Global Crossing, 360networks, etc. bankrupt; >$2T equity wiped High after 2005; second wave captured
DRAM/foundry recurring varies; ~$60–80B per cycle peak n/a <1% globally Mixed, cycle-driven Repeated consolidation; 2–3 survivors per generation High; enabled every downstream computing wave
First cloud (AWS-led) 2010–2020 ~$50B combined (2019) ~$60B <0.5% ~100% from operating cash flow Three winners; no major bankruptcies Very high; AWS alone ~$40B operating income on capex base
AI / hyperscaler 2023–? ~$443B (2025) → ~$646B (2026 est.) same ~2.0% (2026) ~90% from operating cash flow at top 5; debt concentrated in neoclouds TBD; balance sheets strong at hyperscaler tier TBD; productivity J-curve in early phase

What the comparison tells us about AI capex

Five conclusions worth holding simultaneously:

1. The scale is large, but it is not unprecedented. AI capex at 2% of GDP is a meaningful share of the economy, but it is below the railroad peak (10–20%) and similar in real-dollar terms to the dotcom-era IT capex share of GDP (~4.4% for tech equipment and software broadly in 2025). The framing "this has never happened before" is wrong on the data. The framing "this is bigger than telecom" is correct in absolute dollars and similar to telecom as a share of GDP.

2. The historical hit rate on infrastructure overbuilds being eventually utilized is very high. Fiber, rail mileage, electrical generation, semiconductor fab capacity, and cloud capacity have all, in their turn, been eventually absorbed and then exceeded by demand. The skeptic position "this capacity will be stranded" has a poor historical track record on a 10-year horizon. It has a much better track record on a 3–5 year horizon: dark fiber was a real overhang for half a decade before clearing.

3. The historical hit rate on the first generation of operators surviving is much lower. Across every major infrastructure cycle, the financial returns to the first cohort of builders have been bad, even when the social returns were excellent. The mechanism is consistent: debt-financed leveraged builders are wiped out in the gap between buildout and utilization, and the assets are then re-priced into the hands of operators who can carry them through the deployment phase. This is the Perez "double movement" — installation phase ending in financial crisis, deployment phase delivering the gains.

4. The capital structure of the AI buildout is unusually defensive by historical standards. The top five hyperscalers funding 90% of their capex from operating cash flow, with low net leverage and dominant market positions in adjacent businesses, is structurally unlike the 1996–2001 telecom operators. It is more like the post-2010 cloud buildout, which produced unambiguously good returns. Where leverage is concentrated — in the neocloud tier and SPV-financed data centers — that is the segment where the historical analogy to the telecom bust is most apt.

5. The productivity dividend is likely to lag the capex by years, possibly a decade. The Paul David / Brynjolfsson J-curve has been right twice in modern economic history. The implication is that the period 2026–2030 may show large measured capex, weak measured productivity, and a widening gap between AI capability benchmarks and AI economic statistics. This is what the historical pattern looks like in its early phase; it is not necessarily a signal that the buildout was wrong.

The honest summary is that the AI buildout sits at an unusual point in the historical distribution: closer to telecom in absolute dollars and to electrification in general-purpose-technology character, but with a capital structure closer to the cash-funded cloud buildout than to either. That combination is genuinely novel. The historical analogies bound the range of outcomes but do not settle which one we are in.

Sources

Telecom buildout

Railroads

Electrification and productivity lag

Perez framework

  • Perez, Technological Revolutions and Financial Capital: The Dynamics of Bubbles and Golden Ages (Edward Elgar, 2002). PDF
  • Carlota Perez, publications

Semiconductors

Hyperscaler / cloud capex