Without AI Spending, US Corporate Investment Would Be Negative: The $725 Billion Concentration Risk Propping Up the American Economy

The United States economy grew at an annualised rate of 2.0% in the first quarter of 2026. That number, taken alone, sounds unremarkable — a mid-cycle expansion, slightly below trend, nothing to alarm or excite. But beneath that headline figure is a structural dependency that has no precedent in the modern American economy. Approximately 75% of that growth — 1.5 percentage points of the 2.0% total — came from a single category of spending: artificial intelligence infrastructure. Computer and peripheral equipment investment, software investment, and data-centre construction collectively drove three-quarters of all economic expansion. Pantheon Macroeconomics, one of the most closely followed independent economics consultancies on Wall Street, calculated in February that without AI spending, total US corporate equipment investment would be contracting. Not slowing. Contracting.

This is not a statement about AI's promise or peril. It is a statement about the arithmetic of the world's largest economy. Five companies — Amazon, Microsoft, Alphabet, Meta, and Oracle — are collectively spending approximately $725 billion on capital expenditure in 2026. That figure is larger than the GDP of all but roughly 20 countries on Earth. It represents about 2% of US GDP flowing through the balance sheets of five firms. And it is accelerating.

The Numbers: Five Companies, $725 Billion, One Technology

The scale of AI capital expenditure in 2026 requires careful enumeration because the numbers are large enough to distort intuition. Morgan Stanley projects that the five largest hyperscalers will spend approximately $805 billion in total capex this year. Goldman Sachs puts the figure at over $500 billion for AI-related spending specifically. Apollo Global Management estimates hyperscaler capex at $646 billion, or roughly 2% of US GDP. The differences reflect methodological choices about what counts as “AI-related” versus general infrastructure, but the order of magnitude is consistent: somewhere between $650 billion and $800 billion, with approximately 75% directed at AI-specific infrastructure.

To put $725 billion in context: it exceeds the entire global defence spending of every country except the United States and China combined. It is more than the GDP of Switzerland ($906B), Saudi Arabia ($1.1T), or the Netherlands ($1.2T). It dwarfs the combined spending of the dot-com era, the mobile revolution, and the initial cloud computing buildout. This is not incremental growth. It is the largest single-year capital expenditure surge in the history of the technology industry.

The GDP Dependency: Strip Out AI and the Economy Looks Very Different

The most consequential finding in recent US macroeconomic research is not about inflation or interest rates. It is a simple counterfactual from Pantheon Macroeconomics: if you remove AI-related spending from the national accounts, US corporate equipment investment is declining. Not flat. Declining.

The BEA's GDP breakdown for Q1 2026 tells the story. Real GDP grew at 2.0% annualised. Computer and peripheral equipment investment and software investment together contributed approximately 1.09 percentage points to that figure — more than half the total from just two line items. Add in the construction spending for data centres (classified under nonresidential structures) and the AI-adjacent semiconductor equipment demand, and the AI ecosystem accounts for roughly 1.5 percentage points of the 2.0% headline number. The remaining 0.5 percentage points comes from everything else: consumer spending, government expenditure, net exports, residential investment, and all non-AI business investment combined.

A Fortune analysis in February went further, calculating that AI capex plus the wealth effect from tech stocks — the mechanism by which rising share prices make tech employees and equity holders feel richer and spend more — now accounts for roughly one-third of all US economic activity. Former Dallas Fed President Robert Kaplan, now at Goldman Sachs, publicly warned the Federal Reserve in June that it should treat the $725 billion AI capex cycle as a macroeconomic variable requiring monitoring, not merely an industry trend.

What $725 Billion Actually Buys

The money is going into three categories of physical infrastructure, each with its own supply chain and bottleneck profile.

GPUs and custom AI chips account for the largest single share. Nvidia's data-centre revenue alone exceeded $130 billion in its fiscal year 2026, and both Amazon (Trainium) and Google (TPU) are investing billions in proprietary silicon. The semiconductor supply chain runs through Taiwan (TSMC fabrication), South Korea (Samsung and SK Hynix high-bandwidth memory), and the Netherlands (ASML lithography). A disruption to any of these nodes — whether from geopolitics, natural disaster, or trade restrictions — would throttle the entire AI buildout.

Data-centre construction is the second category. The 14 largest publicly owned data-centre operators globally are expected to spend close to $750 billion in 2026, nearly double the $450 billion spent in 2025. New facilities are being built from northern Virginia to Singapore to the deserts of Saudi Arabia. The construction boom is visible in cement demand, steel orders, and specialised construction labour markets. In northern Virginia, the largest data-centre cluster in the world, local officials report that data-centre permits now exceed all other commercial construction categories combined.

Power infrastructure is the third and increasingly binding constraint. A Motley Fool analysis from June 2026 identified three bottlenecks for the $725 billion capex cycle: power, memory, and optical bandwidth. Of these, power is the hardest to solve quickly. A new data centre requires reliable electricity measured in hundreds of megawatts. A single large AI training cluster can consume as much power as a small city. The grid was not built for this, and permitting new generation capacity takes years.

The Energy Dimension: If Data Centres Were a Country

The International Energy Agency projects that global data-centre electricity consumption could approach 1,050 terawatt-hours (TWh) by the end of 2026. In 2024, data centres consumed approximately 415 TWh — about 1.5% of global electricity demand. The projected 2026 figure represents more than a doubling in just two years.

To put 1,050 TWh in perspective: if data centres were a country, they would rank as the fifth-largest electricity consumer on Earth, behind China, the United States, India, and Russia, but ahead of Japan. Electricity consumption from AI-focused data centres specifically surged 50% in 2025 alone, and the growth rate is accelerating as new GPU clusters come online. The US Energy Information Administration projects that total US power demand will hit record levels in both 2025 and 2026, driven primarily by data centres.

This creates a set of second-order economic effects. Natural gas demand rises (gas still provides the marginal megawatt in most US power markets). The copper supercycle intensifies (data centres and grid connections require enormous quantities of copper wiring). Utility companies in data-centre-heavy regions are raising capital expenditure budgets and, in some cases, delaying the retirement of coal and gas plants to meet demand. In a bitter irony for climate policy, the AI revolution that its proponents describe as a tool for solving climate change is, in the short term, increasing fossil fuel demand.

The Concentration Risk: Five Companies as Macro Variable

The United States has experienced investment-driven growth cycles before. The railroad boom of the 1870s, the electrification buildout of the 1920s, the interstate highway system of the 1950s, and the fibre-optic buildout of the late 1990s all channelled enormous capital into physical infrastructure. But those booms were distributed across dozens or hundreds of companies. The current AI capex cycle is concentrated in five.

Amazon, Microsoft, Alphabet, Meta, and Oracle together account for roughly 60% of all US nonresidential fixed investment growth in 2026. Their spending decisions are made by a handful of executives and boards, driven by competitive dynamics within the AI industry rather than by broad-based economic demand signals. When Amazon commits $200 billion to AI infrastructure, it is partly because Microsoft committed $150 billion. When Meta announces $115–135 billion, it is partly because Alphabet announced $175–185 billion. The investment decisions that drive most of America's GDP growth are partly a competitive arms race among five firms, not a response to widespread demand growth across the economy.

This creates an unusual fragility. If the AI investment thesis were to be reassessed — if revenue growth from AI products disappointed, if a major hyperscaler cut its capex guidance, if financial conditions tightened enough to make $725 billion in annual spending unsustainable — the impact would not be confined to the technology sector. It would show up as reduced GDP growth, lower construction activity, weaker semiconductor demand (affecting South Korea and Taiwan), lower copper prices (affecting Chile and Peru), and reduced power-plant orders. The AI capex cycle has become a macroeconomic transmission mechanism that connects Silicon Valley boardrooms to commodity markets in Santiago and grid operators in Tokyo.

The Two-Track Economy

The AI capex boom is not lifting all boats. It is lifting the boats that are docked in the AI harbour and leaving the rest to navigate choppier waters. The K-shaped American consumer dynamic — where the top 10% of earners, who hold 87% of equities, drive half of all spending — is partly an AI story. The wealth effect from surging tech stocks (Nvidia alone has added over $1 trillion in market capitalisation since 2024) flows disproportionately to the holders of those stocks, who then spend more on housing, travel, and services.

Meanwhile, the non-AI economy tells a different story. Business investment in non-technology equipment is declining. The Conference Board Leading Economic Index, while recovering slightly (+0.1% in May to 99.3), has six- and twelve-month growth rates that remain negative, suggesting continued deceleration ahead. The Conference Board projects 1.8% GDP growth for full-year 2026, down from 2.1% in 2025. If AI capex were to plateau at current levels rather than continue growing, that 1.8% forecast would fall to well below 1%.

The Federal Reserve faces a unique challenge. Nine of 18 FOMC members now project rate hikes in 2026, with the dot plot's median at 3.8%. But the inflationary impulse from AI spending is not the kind that higher rates typically cure. Raising rates from 3.5–3.75% to 4.0% or above would increase borrowing costs for consumers and small businesses — the non-AI economy — while doing little to slow the capex of cash-rich hyperscalers who fund their spending from operating cash flow, not debt. Meta, Alphabet, and Microsoft hold hundreds of billions in cash and short-term investments. Their spending decisions are not interest-rate-sensitive in the way that a small manufacturer's are.

The Global Spillovers

The AI capex boom is not purely an American phenomenon, but the United States captures a disproportionate share of the first-order spending. The second-order effects, however, ripple globally.

Taiwan's GDP is expected to grow over 4% in 2026, propelled by semiconductor exports that are overwhelmingly driven by AI chip demand. TSMC alone accounts for over 90% of the world's advanced-node fabrication. A slowdown in AI capex would be felt in Taiwan's national accounts within a single quarter. South Korea faces similar exposure through Samsung and SK Hynix, whose high-bandwidth memory (HBM) chips are essential for AI training clusters. Malaysia has attracted billions in data-centre investment from US hyperscalers, transforming its economic growth prospects.

Commodity exporters feel the pull differently. Chile, the world's largest copper producer, benefits from data-centre wiring demand that has contributed to copper prices exceeding $10,000 per tonne. Energy exporters in the Gulf and the United States see incremental natural gas demand from power generation for data centres. Even renewable energy deployment is being accelerated by corporate power-purchase agreements from hyperscalers seeking to offset their carbon footprint.

What Could Go Wrong

Every previous technology investment cycle has eventually cooled. The railroad boom produced overbuilding and the Panic of 1873. The fibre-optic buildout of the late 1990s ended with the dot-com crash, leaving thousands of miles of dark fibre under city streets. The question is not whether the AI capex cycle will slow but whether the broader economy can absorb the impact when it does.

Three scenarios merit consideration. In the best case, AI revenue growth accelerates enough to justify the spending, capex stabilises at high levels rather than growing further, and the technology diffuses into the broader economy through productivity gains — the St. Louis Fed is already tracking AI's contribution to productivity growth. This is the scenario where the railroad analogy is apt: the buildout is expensive and disruptive, but the infrastructure pays for itself over decades.

In the middle case, AI revenue disappoints relative to the $725 billion being invested, and one or more hyperscalers cut capex guidance. Markets reprice AI expectations downward. The wealth effect reverses, consumer spending weakens, and the Fed faces a choice between cutting rates to support growth and holding rates to combat still-elevated inflation. This scenario is closest to the dot-com parallel: not catastrophic, but a meaningful growth shock that takes 12–18 months to work through the economy.

In the worst case, a geopolitical event — a trade disruption with China that cuts off semiconductor supply, a conflict involving Taiwan, or a financial crisis that freezes capital markets — forces a sudden stop in AI capex. The economy loses 75% of its growth engine in a matter of quarters. This scenario is unlikely in 2026 but not impossible, and it is the reason that the concentration of economic growth in five companies and one technology represents a systemic risk that policymakers are only beginning to acknowledge.

The Deeper Question

The $725 billion AI capex cycle raises a question that extends beyond economics into political economy: what does it mean for a democratic society when three-quarters of its economic growth depends on investment decisions made by the leadership teams of five private corporations?

The United States has always relied on private investment as its primary growth engine — it is a defining feature of American capitalism. But the degree of concentration is new. During the railroad era, hundreds of competing companies built parallel lines across the continent. During the electrification buildout, thousands of utilities, municipal and private, invested in generation and distribution. Even during the dot-com boom, investment was spread across thousands of startups, telecoms, and hardware companies. The current AI buildout is being funded, at scale, by five companies with a combined market capitalisation exceeding $15 trillion.

David Sacks, the White House AI adviser, has framed this positively: AI could drive 75% of US GDP growth, he told reporters in June, positioning America as the undisputed leader in the most important technology of the century. That may well prove correct. But a GDP growth rate that depends on five companies continuing to spend at rates that have no historical precedent is a growth rate that carries a very specific kind of risk. It is the risk of concentration, and it is not one that the Federal Reserve, the Treasury Department, or the BEA's national accounts framework is currently equipped to measure, monitor, or mitigate.

The US economy is not in crisis. But it is standing on a narrower foundation than the headline numbers suggest. The $725 billion question is whether that foundation is the beginning of a new era of productivity-driven growth or a single point of failure in the world's most important economy.

Explore the data behind this analysis: Largest Economies in the World • GDP by Country • Electricity Consumption by Country • Country Comparison Tool.