Analytical Essay

How Big Is the AI Buildout, Really?

Evaluating the global artificial intelligence infrastructure supercycle against history's core U.S. megaprojects and wartime mobilization.

By Karl Svartholm & co • June 4, 2026 • 15 min read

TL;DR

At roughly $665 billion a year, the AI infrastructure buildout is the most capital-intensive peacetime sprint in modern history — bigger, per year, than Apollo, the Interstate Highway System, or the Manhattan Project. But it is still a fraction of U.S. wartime mobilization and of the global oil market; much of the spend (the GPUs) goes obsolete within a few years; and it carries a real power, water, and community bill. Enormous — but not unprecedented.

The usual way to compare large historical projects is to adjust their cost for inflation. That gives us a first rough scale. But inflation-adjusted dollars are a strange, somewhat fragile measure. They answer a question like, “What would this specific amount of money feel like in today's prices?” which is fundamentally different from asking, “How much of society's actual, real-resource capacity did this project command?”

To evaluate whether the contemporary artificial intelligence buildout represents a typical software-driven speculative bubble or the most capital-intensive physical infrastructure supercycle in modern human history, we must subject it to a multi-tiered analysis: comparing inflation-adjusted capital commitments, peak shares of Gross Domestic Product, shifts in national surplus capacity, resource bottleneck grabs, and comparative consumer-spending equivalents.

Scope

A note on what is being compared. Every historical benchmark here — the Manhattan Project, Apollo, the Interstate Highway System, the Bell System, and World War II — is a U.S. program or U.S. wartime spending, measured against U.S. GDP. The AI buildout figure, by contrast, is global hyperscaler capex. So throughout, a worldwide spend is set beside a single nation's economy and its national projects — which, if anything, understates how large the AI number looms over any one country. In the closing section the comparison flips to global everyday spending (coffee, oil, and the rest); those are all global totals, tagged inline.

The Chronological Context (1935–2035)

TL;DR

A visual map of when each comparison project was active, 1935–2035. Hide any bar to re-scale every chart in the essay at once.

Below, we trace the active developmental windows of each comparative era. Click the eye icon next to any horizontal bar to hide it and dynamically auto-scale the charts throughout the essay.

Inflation-Adjusted Cost & Annualized Velocity

TL;DR

Total cost hides speed. Spread over decades, the U.S. Interstate ran ~$18B/yr; the global AI buildout runs ~$665B/yr — only U.S. WWII spending ever ran faster.

When evaluating infrastructure, we often fall victim to the cumulative illusion: celebrating projects because their total absolute budget was large. But this obscures the dimension of temporal velocity (how much money was injected per unit of active buildout time).

For example, the U.S. Interstate Highway System consumed over $634B in total resources. Yet, because that outlay was spread smoothly across 35 years, its annualized velocity sat at a manageable ~$18.1B/year. Contrast this with the current global AI buildout: a three-year (2024–2026) initial sprint carrying an annualized velocity of approximately $665B/year.

Chart 1: Inflation-Adjusted Cost comparison

Compare Absolute Cumulative Costs vs. Annualized Velocity

Project / Buildout	Years Active	Cumulative Cost (2026 $)	Annual Velocity (2026 $/yr)
Manhattan Project	1942–45 (3 yrs)	~$40B	~$13.3B/yr
Apollo Program	1960–73 (13 yrs)	~$309B	~$23.8B/yr
Interstate Highway	1956–91 (35 yrs)	~$634B	~$18.1B/yr
Bell System Assets	1939–74 (35 yrs)	~$470B	~$13.4B/yr
Late-1990s Telecom Buildout	1996–01 (5 yrs)	~$1.0T+	~$200.0B/yr
AI Buildout (Cumulative 24–26)	2024–26 (2 yrs)	~$1.33T	~$665.0B/yr
AI Buildout (To 2030 Spec)	2024–30 (6 yrs)	~$5.00T	~$833.3B/yr
World War II (U.S.)	1941–45 (4 yrs)	~$4.1T+	~$1,025.0B/yr

GDP Share: How Big Relative to the Economy?

TL;DR

As a share of U.S. GDP, the AI buildout (~2.1%) towers over Apollo or the Manhattan Project — but sits nowhere near U.S. WWII's ~40%.

A clearer lens is GDP Share—measuring a project relative to the size of the contemporary U.S. economy to see what percentage of output it absorbed at its absolute peak. (Every share in this section is against U.S. GDP; the AI buildout is global capex measured the same way.) This shows the true magnitude of the U.S. war effort's entire-society shift in World War II versus the focused speartip of the Manhattan Project.

Chart 2: GDP Share comparison

Compare Total Amortized GDP share vs. peak annual intensity (default: Peak GDP Mode) — all shares are of U.S. GDP

Project / Buildout	Cumulative GDP share	Peak Annual GDP Share	Cognitive Scope Interpretation
Manhattan Project	~0.25%	~0.4%	Inside the broader war mobilization; physically concentrated.
Apollo Program	~0.25%	~0.4%	Peak NASA budget (1966) was ~4.4% of federal outlays, but 0.4% of U.S. GDP.
Interstate Highway System	~2.0%	~0.4%	Low peak annual drag, but extreme physical and geographical duration.
Bell System (AT&T assets)	~5.0%	~5.0%	Accumulated asset stock of a nation-wide communication utility by 1974.
Late-1990s Telecom Buildout	~1.0%	~1.1%	Highly intense, private market infrastructure deployment over a condensed era.
AI Buildout (2024–2026 Wave)	~2.1%	~2.1%	Exceeds the peak GDP footprint of several legendary peacetime projects combined.
AI Buildout (To 2030 Spec)	~2.5%	~2.5%	Projected macroeconomic footprint if global cloud expansions hit peak modeling.
World War II (U.S.)	~35.0%	~40.0%	The ultimate bound; complete reorganization of the sovereign state.

Note: While U.S. WWII spending redirected over a third of the nation's productive capacity, the global AI buildout represents the most massive single industrial shift coordinated in modern peacetime history.

The Money-Pie Illusion

TL;DR

Private AI capex isn't a slice of a fixed budget — belief in future profits pulls real resources into the present. The binding constraint is physical, not financial.

A private AI buildout is not just some planner taking slices from a fixed national budget. Much of it is financed directly by belief—belief in future profits, in monopoly rents, in generative productivity, in high-margin SaaS multiples, and in global “intelligence-as-a-service.” Capital markets pull those expected futures directly into the present.

“Money is not the resource. Money is the claim. The real question is: what does that claim let you command?”

Under state-directed programs, the government commands by fiat. In private cycles, command is distributed. But the ultimate constraints are physical. When tech conglomerates commit nearly $700 billion in annual expenditure, they are not buying imaginary slices of a money-pie. They are commanding physical materials, supply lines, packaging facilities, and structural labor away from the rest of the civilian economy.

Surplus Capacity: Going Beyond Raw GDP

TL;DR

Measured against U.S. discretionary capacity (GDP minus household consumption), the global AI buildout absorbs ~7.4% — more than any peacetime U.S. project, still far below U.S. WWII's 90%+.

To approximate how much of a society's discretionary capacity is redirected to a project, we can isolate Surplus Capacity (approximated as Gross Domestic Product minus basic personal consumption). This filters out non-discretionary survival spending to focus on the active investment pool.

$$\text{Surplus Capacity} \approx \text{GDP} - \text{Household Consumption}$$

Chart 3: Surplus Capacity comparison

Full analysis controls: select any metric (default: Annual Surplus Share)

Evaluating historical projects through their allocation of surplus capacity highlights how intensely private capital is concentrating the economy's engineering machinery into computational facilities. At its peak, the Manhattan Project utilized roughly 2.0% of surplus. In 2026, the broad global AI buildout consumes roughly 7.4% of U.S. discretionary capacity—making it macroeconomically more demanding than any peacetime U.S. state initiative, while the U.S. WWII effort commandeered an astronomical 90%+ of the nation's surplus as consumption was rationed and reorganized.

Bottleneck Share: What the Project Captures

TL;DR

What a project monopolizes matters more than what it costs. AI corners advanced packaging, HBM memory, high-voltage transformers, cooling, and grid-connection queues.

Instead of dollars, we should measure what a civilizational project monopolizes. By evaluating which core physical assets are captured, we understand how a buildout fundamentally reorders the material landscape.

1942–1945 • Fissile Enrichment

The Manhattan Project

Monopolized uranium/plutonium isotopes, micro-porous barrier tube technologies, top quantum physicists, and massive dedicated regional electrical grids (Oak Ridge).

1960–1973 • Systems Engineering

The Apollo Program

Captured high-end telemetry, liquid rocket propellant production, systems engineering talent, and a massive aerospace contractor supply network.

1956–1991 • Civil Commodities

The Interstate Highway System

Captured regional concrete supply, asphalt refineries, steel structural production, heavy earthmovers, and local planning rights-of-way.

1939–1974 • Centralized Platform

The Bell System Grid

Commandeered global copper mining, utility pole manufacturing, electromechanical relay components, and over 1% of the total US labor force.

1941–1945 • Total Mobilization

U.S. World War II Effort

Captured total steel, oil, and aluminum fabrication capacities; introduced systemic consumer rationing and marshaled over 12 million active military personnel.

2024–2026+ • Silicon & Power

The Modern AI Buildout

Monopolizes TSMC advanced CoWoS packaging, HBM3e fabrication, high-voltage transformers, liquid-cooling manifolds, and regional power grid queues.

Why the Bottleneck Method Fails

TL;DR

Bottlenecks move. Solving one (chip packaging) just exposes the next — memory, then cooling, then the power grid itself.

The bottleneck view feels closer to reality than raw inflation index numbers or Gross Domestic Product. But it has structural flaws too: bottlenecks are highly dynamic and move over time.

Under extreme demand pressure, capital markets and engineering teams do not encounter static checkpoints. Solving one resource shortage immediately triggers downstream constraints.

Sequence of Modern Infrastructure Constraints

Phase 1: Advanced Packaging. CoWoS packaging limitations at TSMC during 2023–2024 restricted H100 GPU assembly.
Phase 2: Specialty Memory. As wafer packaging capacity expanded in late 2024, High-Bandwidth Memory (HBM3e) supply chains became the core pricing driver.
Phase 3: Thermal Infrastructure. By late 2025, datacenter space, liquid closed-loop systems, and high-voltage distribution units constrained deployments.
Phase 4: Utility Connections. The ultimate bottleneck has settled at utility grid connection lead times, permitting delays, and regional baseload generation limits.

Architectural Legacy: What Remains?

TL;DR

The durable layer — grid, fiber, cooling, building shells — outlives any bubble. The GPUs don't, and the power, water, and community bill is real. A mixed legacy, not “permanent foundations.”

The size of the AI buildout is not best measured by how much speculative money is directed at it. It is measured by what it makes unavailable to everything else, how long those bottlenecks last, and whether the constructed physical assets remain useful if the financial story breaks.

The Telecom Legacy: While fiber operators like WorldCom and Global Crossing faced severe financial distress in 2001, the millions of miles of transcontinental glass they laid remained in the ground. The marginal cost of data transmission fell to near-zero, directly enabling the Web 2.0 era, global cloud computing, and modern digital platforms.

The AI Infrastructure Legacy: Here the picture splits in two. The long-lived layer — substations, high-voltage transmission lines, nuclear and renewable power-purchase agreements, fiber interconnects, cooling plant, and the data-center shells themselves — is durable physical capital that will outlast any single financial cycle, much as the fiber did. But the most expensive part is not. The GPUs that absorb the bulk of the spend are effectively obsolete within three to five years; unlike Bell's copper or the Interstate's concrete, they do not sit in the ground quietly earning their keep for decades. A telephone pole from 1965 still carries a line. An H100 from 2024 will be e-waste before the substation feeding it is even fully amortized.

And these foundations are not free to the people who live beside them. The campuses draw enormous baseload power, push up local electricity prices, and reject vast quantities of heat; many are cooled with millions of gallons of water in regions already short of it; and they tend to land in communities that absorb the noise, the construction, and the grid load without automatically sharing in the upside. None of this is the subject of this essay — but none of it belongs in a tidy ledger of “permanent foundations” either.

So the true legacy of the trillion-dollar sprint will be a mixed one: a durable inheritance of power and connectivity, a fast-depreciating mountain of silicon, and a real environmental and community bill — settled in megawatts, water, and concrete long after the stock charts are forgotten.

The Burke-Style Comparison: Coffee, Popcorn & Crude Oil

TL;DR

$665B/yr is unimaginable until you scale it: more than the world's coffee, games, and cinema combined — yet under half its drinks bill and barely a quarter of its oil.

James Burke's trick, in Connections, was never to quote a number — it was to set it beside something you already carry in your head. A figure like $665 billion a year is unimaginable. The same money expressed as "every cup of coffee on Earth, three times over" is not. So before we close, let us run the modern AI buildout against the things humanity already spends its money on without a second thought — the ordinary, non-discretionary streams of everyday life.

The reference figure throughout is the AI infrastructure spend of roughly $665 billion in a single year (the annualized 2024–2026 run-rate from Section 1). Here is what that one year of GPUs, substations, and concrete buys, measured in the small comforts of the species:

Every figure below is a global annual total — one year of worldwide spending, set against one year of AI.

The Box Office Global ~22×

Every cinema ticket sold on the planet totals roughly $30B/yr worldwide. The AI buildout spends, annually, about twenty-two years of humanity going to the movies — the entire theatrical century, in fast-forward.

The Coffee Market Global ~3.6×

The world drinks about $185B/yr of coffee. One year of AI infrastructure equals every espresso, latte, and gas-station drip the species buys — more than three times over.

The Game Industry Global ~3.6×

Every console, PC, and mobile game on Earth adds up to about $185B/yr worldwide. One year of AI capex would bankroll the entire global games industry — all of it — for more than three and a half years.

The Drinks Bill Global 0.42×

The first jolt in reverse. The world spends about $1.6 trillion/yr on beer, wine, and spirits — globally. A full year of the entire AI buildout is under half the planet's bar tab — about 0.42× what humanity drinks.

The Oil Market Global 0.25×

And the bigger one. The world burns through roughly $2.7 trillion/yr of crude oil — globally. For all its scale, a full year of the AI buildout is barely a quarter of the oil bill: the entire trillion-dollar sprint vanishes into the global oil market about four times over.

Stack those streams up against a single year of the buildout and the shape of the thing becomes plain. The bars below are all one year of each — no megaprojects, no decades-long tallies, just what the world spends annually on each, set beside the annual AI spend:

One Year of the AI Buildout vs. One Year of Everyday Spending

Annual spend, same scale. Every figure is a global annual total.

And there is the whole comparison in one image. One year of the AI buildout outweighs the planet's coffee, its entire games industry, and a year of global cinema-going — combined. Yet line it up against the world's drinks bill or its oil market and the spell breaks: the same trillion-dollar sprint is under half of what humanity drinks, and barely a quarter of what it burns as crude, every single year.

But Burke always ended on the catch, and here is ours: these comparisons illustrate the sheer volume of capital moving through financial networks, yet they carry a structural limitation — you cannot easily convert a round of drinks, popcorn, coffee beans, or barrels of crude into high-purity silicon fabrication, cryogenic rocket engines, or uranium enrichment barriers. Some ordinary habits become enormous because they are repeated by millions. Some state projects become enormous because they concentrate authority. Some speculative booms become enormous because belief in the future lets them command resources in the present. The dollar figures match; the things the dollars buy do not.

Citations and Core Metrics

Selected Core Metrics

DOE gives the Manhattan Project's peak workforce as ~130,000 and cost as ~$2.2B.
The Planetary Society gives Apollo as ~$25.8B, or ~$309B in 2025 dollars.
FHWA gives the final 1991 Interstate Cost Estimate as ~$128.9B, with ~$114.3B federal share.
Construction Physics summarizes AT&T / Bell scale as ~$5B in assets in 1939, ~$74B in 1974, and ~1M employees.
DOE/LBNL says U.S. data centers used ~4.4% of U.S. electricity in 2023 and could reach ~6.7–12% by 2028.
Reuters cites Morgan Stanley's estimate that Amazon, Microsoft, Alphabet, and Meta may spend ~$630B on data centers and AI chips in 2026, about 2.2% of U.S. GDP.
Goldman Sachs gives a broader baseline of ~$765B annual AI capex in 2026.
Grand View Research sizes the global alcoholic-beverages market at ~$1.5–1.7T in 2024 (global).
Gower Street Analytics puts the 2024 global theatrical box office at ~$30B, with projections near ~$33B for 2025 (global).
Research and Markets sizes the global coffee market at ~$185B in 2024 (global).
Newzoo puts the global games market at ~$184–188B in 2024 (global).
The U.S. EIA and IEA put global oil consumption near ~101–103 million barrels/day in 2025; at ~$70–75/bbl that is roughly $2.6–2.8 trillion/yr (global).

A note from the author

I made this to understand the AI buildout better — and I think I do now. Whether the result is useful to anyone else, or whether every figure holds up, I genuinely can't say. The numbers are sourced and the reasoning is laid out so you can check it yourself. If something's wrong, tell me.

Colophon Concept & most thoughts by Karl Svartholm; everything else by Gemini 3.5 Thinking & Claude Opus 4.8. — Karl Svartholm