Three Data Centers in a Trench Coat: 4% of GDP, 92% of Growth

“Our economy might just be three AI data centers in a trench coat.”– Rusty Foster, Today in Tabs

He’s not wrong. And the trench coat is getting threadbare.

AI’s 92% GDP Contribution in Context

How unusual 92% concentration really is

To appreciate how unusual 92% concentration is, compare it to the dot-com era. AI-related categories contributed 0.97 percentage points to real GDP growth in the first three quarters of 2025, higher than the 0.69 points that identical IT categories contributed during the dot-com peak in 2000. AI-linked investment drove 39% of GDP growth across the first nine months of 2025. During the dot-com peak, the equivalent figure was 28% (St. Louis Fed, January 2026).

By August 2025, something happened that had no precedent: AI data center expenditure’s contribution to GDP growth surpassed the total impact of all U.S. consumer spending. Consumer spending is two-thirds of GDP. A category representing 4% of the economy was outgrowing it.

The numbers stacked fast. AI-related capex contributed 1.1 percentage points to GDP growth in H1 2025 (J.P. Morgan), outpacing the consumer as an engine of expansion. Hardware investment was up 41% year-over-year. Data center construction hit a record $40 billion annual rate by June. Capex among the top cloud companies had quadrupled to nearly $400 billion annually, with the top 10 spenders accounting for nearly a third of all U.S. business spending (Morgan Stanley).

The 92% figure has a real asterisk. MRB Partners analyst Shaireen Bhide argues it overstates AI’s net contribution: much of the hardware going into data centers is imported (GPUs from Taiwan, networking equipment from Asia), and imports subtract from GDP. After adjusting for AI-related imports, Bhide estimates the net contribution drops to 40–50 basis points, or roughly 20–25% of real GDP growth. Bespoke Investment Group reached a similar conclusion, noting that Q1 2025 was an outlier.

Both analyses are methodologically sound. But even the adjusted numbers tell a concerning story. A single investment category driven by a handful of companies accounting for a fifth to a quarter of all economic growth is not normal. And the unadjusted figures, the ones that showed up in the BEA data and shaped policy, created a GDP headline that masked what was happening underneath. Manufacturing was stalling. Retail was weak. Job creation was slowing. The rest of the economy was barely expanding.

The headline GDP said “strong economy.” The composition said “one-legged stool.”

The AI Infrastructure and Housing Bubble Parallel

Structural dependency then and now

In 2005, residential investment reached 6.7% of U.S. GDP, its highest level in half a century. The Federal Reserve documented how residential investment had surged 40% above its long-run average share of GDP. Mortgage debt climbed from 61% of GDP in 1998 to 97% by 2006. Between 2001 and 2005, roughly 40% of net private-sector job creation came from housing-related sectors.

The economy looked great. GDP was growing. Employment was up. The problem was that the growth was structurally dependent on a single sector, and that sector was fueled by financial engineering that disguised the true risk. Sound familiar?

The AI infrastructure boom shares an uncomfortable structural similarity. Not in the specific mechanism (nobody is packaging subprime data center leases into CDOs yet), but in the concentration pattern. A narrow sector is generating a disproportionate share of GDP growth. The rest of the economy is underperforming. And financial engineering is making the true exposure difficult to measure.

There are real differences, and they matter. Housing had a direct wealth effect on 69% of American households. AI infrastructure investment flows to a handful of companies and their shareholders. The dot-com bust wiped out roughly $6 trillion, about 60% of GDP at the time. Oliver Wyman’s January 2026 analysis estimates a comparable AI equity correction would erase approximately $33 trillion. That is more than total U.S. GDP. The WEF argues the consumption impact would be more limited precisely because AI wealth is more concentrated than housing wealth was. Cold comfort if you hold the stocks.

The housing bust triggered a financial crisis because the risk was embedded in the banking system through mortgage-backed securities. The AI boom’s financial plumbing looks different. Not necessarily safer.

Where $120B in AI Data Center Debt Is Hiding

Off-balance-sheet debt, SPVs, and hidden leverage

Tech companies have moved more than $120 billion in data center debt off their balance sheets using special purpose vehicles, according to the Financial Times. Oracle leads with $66 billion, followed by Meta at $30 billion, xAI at $20 billion, and CoreWeave at $2.6 billion. The structures involve private credit firms (PIMCO, BlackRock, Apollo, Blue Owl Capital, JPMorgan) providing debt and equity through entities designed to keep liabilities off the hyperscalers’ books.

Paul Kedrosky describes the mechanism plainly: companies create SPVs they indirectly control but don’t have to consolidate on their balance sheets. Meta’s $27 billion Hyperion data center deal with Blue Owl, structured through an SPV named “Beignet Investor,” has just $2.5 billion in equity against $27 billion in debt. That’s a 10% equity cushion. Kedrosky calls it “wildly insufficient if projected AI workloads stall or margins compress.”

UBS reports that tech companies had borrowed approximately $450 billion from private funds as of early 2025, up $100 billion year-over-year. Morgan Stanley estimates $800 billion in private credit will be required between 2025 and 2028 to finance AI data centers alone. In 2025, the five major hyperscalers issued $121 billion in bonds, more than four times their five-year average. Their combined free cash flow is forecast to shrink by 43% between late 2024 and early 2026.

“In 2008, banks discovered they owned far more US housing risk than their internal reports suggested. They might soon discover the same about data-center and digital infrastructure risk.”– Oliver Wyman, January 2026

AI Companies Paying Each Other

Cross-investments, round-tripping, and inflated demand

The financial engineering extends beyond SPVs. Some of the AI revenue being counted as economic growth is companies paying each other. Bloomberg mapped what it called AI’s “circular deals,” the web of cross-investments where companies invest in each other, creating revenue that circles back to the investor. Microsoft invested $13 billion in OpenAI, which spends most of it on Microsoft Azure. OpenAI signed a $300 billion cloud deal with Oracle, which must buy Nvidia GPUs to fulfill it. Nvidia invested in OpenAI’s funding rounds. Nvidia took a 7% stake in CoreWeave, then agreed to purchase $6.3 billion in cloud services from CoreWeave, effectively guaranteeing CoreWeave’s revenue. CoreWeave bought its GPUs with borrowed money collateralized by the value of the GPUs themselves.

OpenAI has committed to over $1.15 trillion in long-term computing contracts, against projected 2025 revenue of $13 billion. Goldman Sachs cited “the increasing circularity of the AI ecosystem.” Morgan Stanley’s Todd Castagno warned it was becoming “increasingly circular” in ways that “inflate demand and valuations without creating economic value.”

“Isn’t it a bit strange when the demand for compute is ‘infinite,’ the sellers keep subsidizing the buyers?”– Jim Chanos, 2025

Data Centers Are Crowding Out the Grid

Data centers vs. housing, labor, and electricity

In central Ohio, a couple opened their electricity bill and found it had risen 60%. They hadn’t changed anything. But 130 data centers had moved in around them. Virginia’s Dominion Energy proposed its first base-rate increase since 1992. Bloomberg’s analysis of 25,000 electricity pricing nodes found wholesale costs up as much as 267% over five years in areas near data centers. The boom isn’t an abstraction. It’s showing up in people’s utility bills.

The system-level numbers are worse. Electricity prices jumped 6.9% in 2025, more than double the headline inflation rate (Goldman Sachs). Data centers make up 40% of electricity demand growth. PJM Interconnection, the largest electric grid in the U.S. serving 65 million people across 13 states, reported that consumers will pay $16.6 billion between 2025 and 2027 just to secure power supplies for data centers that haven’t been built yet. PJM’s independent market monitor called it a “massive wealth transfer” from consumers to the data center industry. Households will see prices rise an additional 6% through 2027, dragging down consumer spending growth by 0.2%.

The Council on Foreign Relations argues the AI bubble may not burst from circular financing or debt levels, but from the mundane reality that data centers and housing construction are competing for the same electricians, welders, and HVAC technicians. Tariffs and immigration restrictions are shrinking the labor pool at precisely the moment both sectors need to expand.

AI Productivity: Where Are the Returns?

If it’s transformative, show me the numbers

If AI infrastructure investment is transformative and not just a capex sugar rush, it should show up in productivity data. U.S. nonfarm business productivity grew at roughly 2% year-over-year through Q3 2025, in line with the post-pandemic average but showing no meaningful acceleration from the hundreds of billions flowing into AI. The Fed’s Kansas City branch found gains concentrated in a handful of industries, not the broad-based uplift you’d expect from a general-purpose technology.

MIT’s Nanda Lab reported that despite $30–40 billion in enterprise AI investment, 95% of organizations are getting zero return. The Penn Wharton Budget Model projects the AI productivity boost will peak at an additional 0.2 percentage points of annual growth . Meaningful, but a fraction of what current investment levels imply. Data centers employ few workers once built, limiting the multiplier effect through wage-driven consumption (J.P. Morgan).

This matters for the GDP dependency story. If the economy isn’t getting more productive from AI investment, then the GDP growth it generates is pure spending, not productivity-driven expansion. The growth lasts exactly as long as the spending does, and not a quarter longer.

“Everybody thought it was going to require more computing power and more bandwidth than it actually did.”– Jerry Kaplan, on the 1990s. The infrastructure always gets overbuilt.

AI Capex Bubble: Industrial Bubbles Leave Real Wreckage

Even Bezos calls it what it is

Even Jeff Bezos called the AI data center buildout an “industrial bubble” at the New York Times DealBook Summit in December 2024. He insisted the long-term benefits will justify it. Maybe. But the distinction matters. An industrial bubble means real physical assets get built that eventually find uses. The fiber-optic cables from the telecom boom carried the internet for two decades. The railroad bubble of the 1800s left behind a continental transportation network.

But industrial bubbles still cause pain. The builders go bankrupt, the investors lose capital, and the construction workers lose jobs when the building stops. When the spending represents a massive share of GDP growth, the withdrawal can tip the broader economy into recession.

The WEF’s Chief Economists Outlook acknowledged this: “Economic growth during the bubble phase depends on continually building infrastructure, not using infrastructure.” As long as the hyperscalers keep spending, GDP grows. When they slow, whether from disappointing revenue, rising debt costs, or simple overbuilding, the contribution reverses.

And the slowdown signals may already be appearing. Alphabet’s free cash flow is projected to plummet roughly 90%. Bond spreads on AI-related debt have widened by as much as 40 basis points since September, per Oliver Wyman. CoreWeave’s stock has swung from a $187 peak to $75, a reminder of how volatile debt-fueled growth models. (For a deeper look at CoreWeave’s $29 billion in debt, the power stocks, and where the stress shows up first.)

Every scenario in that table shares one feature: the economy without AI investment is barely growing. The headline says 1.8%. The foundation says 0.1%.

What the Stress Index Is Tracking

12 signals that measure whether the buildout is sustainable

This GDP concentration is exactly the kind of fragility the AI Infrastructure Stress Index is designed to detect. The composite score aggregates 12 signals across credit risk, operational stress, and market sentiment. Several of them connect directly to the dependency pattern described above. The Hyperscaler Capex/Cash Flow Ratio tracks whether the companies driving GDP growth are spending sustainably or burning through cash. The AI Infrastructure Credit Spreads signal monitors the cost of borrowing for the buildout. When capex ratios climb above 100% and credit spreads widen simultaneously, it’s the same pattern that preceded the telecom correction: spending that depends on continued access to cheap capital.

When a single category of business investment carries most of GDP growth, whether it’s housing in 2005 or AI infrastructure in 2025, the economy becomes fragile in a specific way. Not because the investment is wasteful but because the growth is not self-sustaining. It depends on continued spending at the same pace or faster. The moment the rate of investment decelerates, GDP growth doesn’t just slow. The contribution turns negative.

We’re not predicting a financial crisis. The AI hyperscalers have stronger balance sheets than the banks did in 2007, and the debt is less embedded in the consumer financial system. But the structural pattern is the same: an economy that looks healthy on headline GDP while growing dependent on a single investment thesis that must keep accelerating to sustain itself. Miss the moment it decelerates, and you miss the turn.

“Honey, AI capex is eating the economy.”– Paul Kedrosky

He meant it descriptively. We’re tracking it diagnostically.

Watch the capex guidance. When “accelerating investment” becomes “capital discipline,” you’ll know the trench coat has come off. And underneath it: 0.1%.