America's $10tn AI Infrastructure Boom Is Starting to Fray

The numbers describing America's artificial intelligence buildout have long since left the realm of ordinary corporate spending. A landmark study published last month puts the total investment in AI data centres and related infrastructure at $10.3 trillion between 2025 and 2032 — a sum that would make it the largest investment boom in American history relative to the size of the economy, dwarfing the railroads, the interstate highways, and the electrification of the continent. Now, the Financial Times reports, that boom is starting to fray: financing is tightening, investors are questioning whether the revenues will ever justify the debts, and the physical and financial strains of building at this speed are showing everywhere at once.

The biggest bet in American economic history

The figure comes from economist Stijn van Nieuwerburgh of Columbia Business School, in a paper presented at a Brookings Institution conference in late September and reported by The Wall Street Journal. His central estimate: $10.3 trillion of spending on data-centre buildings, power systems, networking, specialised chips, and related equipment over eight years — an annual average equivalent to about 3.6 per cent of US gross domestic product.

To grasp the scale, compare it with the great buildouts of the past. The railroad boom of the late 1800s — the previous record — absorbed roughly 2.2 per cent of GDP per year. The interstate highway system took about 1.1 per cent from the 1950s onward. The telecoms and fibre expansion of the mid-1990s took a similar share, and nationwide electrification a mere 0.5 per cent. The AI buildout, at 3.6 per cent, is in a different league — and it demands roughly 183 gigawatts of new data-centre capacity against about 57 gigawatts installed today. Van Nieuwerburgh described the financial arrangements knitting the boom together — AI firms, Big Tech, banks, private credit lenders, and real-estate companies — as fiendishly complicated, and warned that the combination of unprecedented scale, unproven revenue streams, and layered debt had created the conditions for a serious downturn.

Real money — and real debt

This is not a story of frothy equity alone. The five big hyperscalers — Alphabet, Amazon, Meta, Microsoft, and Oracle — are expected by FactSet estimates to spend $4.2 trillion in capital expenditure over the four years through the end of 2029, and a growing share of it is being funded with borrowed money. Hyperscaler capital spending has surged from roughly $97 billion in 2020 to more than $400 billion in 2025, and firms are now leaning heavily on bond markets, bank loans, and private credit.

At least $1.3 trillion of debt has already been committed to the data-centre boom, with more capital flowing in from insurers and pension funds — often through special-purpose vehicles and private credit deals that obscure exactly who holds the risk. Morgan Stanley data cited by the FT puts global AI-related debt issuance at $450 billion from January through early September 2026, more than double the total for the whole of 2025, with companies raising an estimated $466 billion in AI-linked debt so far this year against $101 billion in the same period last year. Track how credit markets are pricing this risk on our markets page.

Where the fraying shows

The FT's macro analysis finds the strain in three places: financing pressure, investor scepticism, and execution risk. On financing, the pace of new borrowing has already cracked. AI-related debt issuance collapsed from a record $113 billion in June to just $23 billion in September, as lenders grew more selective about which projects deserve funding. Morgan Stanley attributes part of the drop to the sheer volume already raised earlier in the year — but the FT also reports deepening investor doubt about whether the spending will ever pay off.

That doubt has a number attached. Gartner forecasts that AI providers will spend $7 trillion on infrastructure through 2030 while data-centre-attributable AI revenue reaches no more than $3.6 trillion over the same period — leaving more than $4 trillion of asset value exposed to write-downs through 2032 if the current pace holds. Investor Michael Burry has warned that roughly $3 trillion of firm commitments by top tech firms could force asset write-downs in 2028–2029 if demand lags supply, and Gartner predicts more than 80 per cent of specialised AI cloud providers — the so-called neoclouds — will cease operating independently by 2030. Allianz chief investment officer Ludovic Subran captured the mood by calling this a “Darwinian moment” for data-centre debt, in which weaker projects face a fight for survival.

Equity markets are voting too. In mid-October, shares of debt-heavy AI firms tumbled as bets on further US interest-rate rises intensified: neoclouds Nebius and CoreWeave fell 15.1 and 9.5 per cent in a single session, Oracle dropped 7.2 per cent, and the cost of insuring AI-company debt against default surged to record highs. Higher rates are poison for a buildout funded on borrowed money.

The financial plumbing nobody fully understands

What worries regulators most is not the headline debt but the plumbing underneath it. The Bank of England's Financial Policy Committee concluded at its September meeting that surging AI-related borrowing had spread loss exposure across capital markets: as a widening circle of investors funds the boom, exposure to any disappointment in earnings expectations spreads with it — and could reinforce sell-offs in a self-accelerating loop, as happened during July's sharp slide in AI and semiconductor shares.

Particularly troubling are the guarantees greasing the machine. The FT reported that residual value guarantees provided by Big Tech companies over the preceding year to support AI infrastructure investment amounted to as much as $300 billion. These arrangements underpin borrowing by special-purpose vehicles holding chips or data centres by guaranteeing the assets' future value stays above an agreed level. The Bank of England warned that the opacity of such financing — combined with what it called “circular trading” between companies — could amplify losses if returns disappoint. When the ultimate risk-bearer is hidden three layers deep in an SPV, nobody can price the system accurately, including the people running it.

Physics gets a vote

Even if the financing held, the physical world is pushing back. Data centres devour electricity, water, and land, and communities are increasingly reluctant hosts. Electrician wages are climbing, electricity prices are rising, and chip and equipment costs are being bid up by insatiable demand. In some regions, manufacturing projects have begun competing directly with data centres for power, workers, and land — a contest the factories, with their thinner margins, are losing.

Permitting has become a chokepoint: at least one giant project has resorted to force majeure proceedings after failing to secure approval for the power infrastructure it needs, a sign of how America's creaking approval processes collide with trillion-dollar timelines. Goldman Sachs estimates AI investment alone will amount to 1.9 per cent of GDP in 2026; at that scale, every substation queue and water permit becomes a macroeconomic variable.

Why this matters far beyond Silicon Valley

The buildout has already rewired parts of the real economy. AI has generated more than 750,000 net new jobs in the US since 2023, data centres have added 117,000 jobs since 2024, and the AI-fuelled stock rally has boosted household wealth and consumer spending — making the boom a genuine engine of American growth, not just a tech-sector story. That is precisely why its fraying matters: a concentrated bet this large transmits both its gains and its losses economy-wide.

The transmission channels run through ordinary balance sheets. Pension funds and insurers reaching for yield in private AI credit tie retirement savings to data-centre revenues. Rising electricity demand is already nudging up power bills in buildout states. And the interest-rate angle cuts globally: if US borrowing keeps Treasury yields elevated, the dollar stays strong and emerging-market reserves come under pressure — a dynamic worth watching on our reserves tracker alongside the markets page. When 3.6 per cent of GDP is riding on one technology's payoff, there is no such thing as a contained disappointment.

What steadies it — and what snaps it

The boom does not need to collapse to cause damage; it only needs to disappoint. The benign path is that AI revenues grow into the infrastructure — enterprise adoption accelerates, inference demand compounds, and today's overbuilding looks in hindsight like the fibre glut of the dotcom era, which eventually found its uses. The malign path is a financing strike: lenders pull back faster than revenues arrive, weaker SPVs and neoclouds fail, guarantees get called, and forced selling turns a repricing into a rout.

Which path prevails hinges on three variables: the trajectory of AI revenues over the next three years, the path of interest rates, and whether the physical buildout — power, permits, people — can keep pace with the financial commitments already signed. None of the three is within any single company's control. That is what makes this a genuine macro gamble rather than a collection of corporate bets — and why, as the FT's analysis makes clear, the fraying at the edges deserves to be taken seriously long before it reaches the centre.

FAQs

Is this just another dotcom bubble?

There are real differences and uncomfortable similarities. Unlike 1999, today's buildout is anchored by the world's most profitable companies spending actual cash flow, much of the capacity is pre-contracted, and AI revenues — while smaller than the spending — are real and growing. But the structure rhymes: capital expenditure is running far ahead of revenue, leverage is rising, and the most aggressive borrowers are the least diversified. The comforting precedent for bulls is that the dotcom fibre glut eventually got used; the warning is that it took a decade and wiped out most of the companies that laid it.

What are residual value guarantees, and why should I care?

A residual value guarantee is a promise — in this case largely from Big Tech firms — that an asset such as a data centre or a fleet of AI chips will still be worth at least an agreed amount in the future. That promise lets a special-purpose vehicle borrow cheaply to buy the asset, because lenders treat the guarantor's balance sheet as backstop. You should care because roughly $300 billion of these guarantees now underpin AI borrowing, and if chip or property values fall short, the losses flow straight back to the guarantors — concentrating supposedly diversified risk in a handful of household names.

How does a data-centre boom touch ordinary households?

Through three channels: bills, jobs, and savings. Surging electricity demand is already pushing up power prices in buildout states, while local debates rage over the tax breaks lavished on data-centre developers. On the positive side, construction and operations have created well over 100,000 direct jobs plus hundreds of thousands more across the supply chain. And through pension funds and insurers that have bought into private AI credit, millions of retirement accounts now carry quiet exposure to whether data-centre revenues materialise.

What would a disorderly unwind actually look like?

It would start in credit, not equity: spreads on AI-linked debt widening, weaker special-purpose vehicles unable to refinance, guarantees being called, and neoclouds failing or being absorbed. The Bank of England's warning sketches the amplifier — investors who borrowed to build concentrated AI positions selling into falling prices, deepening the decline and triggering further deleveraging. Because insurers and pension funds now sit inside these structures, the pain would travel from Wall Street to retirement statements faster than in a pure equity bust.

Can the buildout simply stop?

Almost certainly not — too much is already sunk and contractually committed, from take-or-pay power agreements to multi-year chip orders, and the political stakes of ceding AI infrastructure leadership are enormous. The likelier outcome is triage rather than a halt: capital gets rationed, the strongest projects proceed, and weaker ones are cancelled, merged, or sold for cents on the dollar. That is what a “Darwinian moment” means in practice — not the end of the boom, but its sorting into winners and wreckage.