The United States has made a wager of historic proportions. It has bet its economic growth, its stock market, the retirement savings of its households and a growing share of its credit system on a single proposition: that artificial intelligence will pay for itself. I believe the proposition is probably true. I also believe that "probably" is doing far more work in that sentence than markets currently admit.
Consider the scale of what is happening. Data-centre investment is on course to rise from 1.4 per cent of US GDP in 2025 to 3.1 per cent in 2027 — a build-out running at roughly twice the pace of the housing boom at its frenzied peak, as Torsten Slok of Apollo has documented. Jason Furman of Harvard has calculated that investment in information processing equipment and software, though only 4 per cent of GDP, accounted for 92 per cent of measured growth in the first half of 2025. Strip it out and the world's largest economy grew at an annual rate of 0.1 per cent. In brief: America does not have an economy with an AI boom inside it. It has an AI boom with an economy attached.
Is this a bubble? The honest answer is: yes and no. Mohamed El-Erian and the Nobel laureate Michael Spence call it a "rational bubble", and the phrase is apt. When a technology promises returns of transformational scale, it is rational for each individual actor to overinvest, even though the aggregate result must be that some capital is destroyed. This was true of railways in the 1840s, electricity in the 1920s and fibre optics in the 1990s. Each of those manias built infrastructure that transformed the world. Each also incinerated the wealth of those who financed it. The two outcomes are not alternatives. They arrive together.
Yet this cycle differs from its ancestors in one crucial respect: it is being financed with debt, at scale, and increasingly in the dark. JPMorgan now estimates total AI capital spending of $5.5tn through 2030, of which $4.1tn will be debt-financed, with loan-to-cost ratios averaging above 85 per cent. The five great hyperscalers issued roughly $121bn of US bonds in 2025, against an average of $28bn a year over the previous five. Meta raised $30bn in a single offering, the largest corporate deal of the year. Oracle, having issued some $43bn of notes in a single fiscal year, was downgraded in July to one notch above junk. Alphabet — Alphabet! — reported negative free cash flow in the second quarter of 2026.
The optimists reply, correctly, that these are profitable, investment-grade companies with enormous order backlogs. Oracle points to $638bn of contracted obligations; Microsoft to $627bn. The debt markets, they note, are functioning smoothly. All true. But the case for concern was never really about the giants. It is about what has assembled itself around them.
Three features of the periphery deserve attention. The first is private credit, which Morgan Stanley expects to supply some $800bn of the $1.5tn external financing gap for data centres through 2028. This is a market of stale marks, opaque collateral and, as the Tricolor and First Brands frauds demonstrated, occasionally pledged-twice collateral. Jamie Dimon warned that where one finds one cockroach, there are usually more. The Federal Reserve's own survey found that concern about private credit among market contacts jumped from 22 per cent to 50 per cent within a year. Business development companies saw redemptions exceed inflows for the first time on record. These are not the vital signs of a healthy corner of finance.
The second is circularity. Nvidia invests in OpenAI, which contracts with Oracle, which buys from Nvidia, while a growing share of the chipmaker's sales flows to entities it has itself financed. Jim Chanos, who has seen this film before, notes that today's vendor financing already far exceeds the roughly $100bn of the dotcom era, when Lucent's circular sales ended in write-downs and ignominy. Michael Burry goes further, alleging that hyperscalers flatter earnings by some $176bn over three years through generously stretched GPU depreciation schedules. His estimate may be wrong. The question he asks — how long does a chip that is obsolete in three years get depreciated over six? — is not.
The third, and least discussed, is what the boom does to everyone else. AI-related issuance has risen from 1 per cent of investment-grade supply in 2024 to roughly 18 per cent this year. The Dallas Fed estimates the resulting duration supply at $360bn in ten-year equivalents — about an eighth of what the Treasury itself supplies. This crowds out weaker borrowers precisely as the 2026-27 high-yield maturity wall arrives. Meanwhile AI is manufacturing a new cohort of the walking dead: software and services firms whose revenues the technology quietly devours faster than their debts can be refinanced. The Bank for International Settlements found the zombie share of listed firms in advanced economies had already risen from 4 per cent in the late 1980s to 15 per cent by 2017, in an era of ever-cheaper money. That era is over. The zombies now face higher coupons, scarcer capital and a predator.
What, then, should we expect? The base case — and it is genuinely the base case — is that the boom continues, the hyperscalers remain solvent, and the accidents stay idiosyncratic: cockroaches, in El-Erian's taxonomy, rather than termites. Even Nouriel Roubini, of all people, has declared the crash thesis mistaken over the medium term. But Slok has identified the asymmetry that should concentrate minds: a cycle that builds at 0.85 percentage points of GDP a year can unwind at a similar pace. The telecom bust of 2001 produced the mildest of postwar recessions. The housing bust, which unwound at comparable speed from a comparable share of GDP, produced the worst. The difference lay not in the technology but in the leverage and the opacity of the financing. On both counts, this cycle has been drifting in the wrong direction.
The signals worth watching are unglamorous: hyperscaler free cash flow, which has already turned negative at two major firms; the AI credit basket, whose spreads have doubled in a year; depreciation disclosures; and above all the companies closest to final demand, whose stumble preceded the capex crash in 2000 while the equipment-makers still soared.
I do not know when this cycle turns, and neither does anyone else. What I do know is that the last two decades taught a costly lesson: the danger is never the technology, which usually works, but the financial structure erected upon it, which usually doesn't. AI will very likely transform the world economy. Whether it first transforms the credit system into rubble is a separate question — and it is the one on which the next few years of prosperity now rest. Thus we arrive, once again, at the oldest truth in finance: a good idea is not the same thing as a good investment. Alas, entire economies keep having to relearn it.
Part two of this series: America is becoming what it used to lecture.