Artificial Finance

Markets

AI Capex Debt Is Reshaping Credit Markets

The AI build-out has quietly become a debt story. Hyperscaler bond issuance is up roughly 970% this year, an estimated $1.65tn sits off balance sheet, and investment-grade credit is starting to feel the strain.

By The Artificial Finance Desk 11 min read

The AI trade has moved from the stock market to the bond market#

For three years the story of artificial intelligence in finance was a story about equity: Nvidia's market value, the "Magnificent Seven," the price of a GPU. That framing is now out of date. The most important development at the intersection of AI and markets this week is not a model release or an earnings beat. It is the growing realisation, spelled out by central banks and sell-side credit desks alike, that the AI build-out is increasingly financed with debt, and that a large part of that debt is not sitting where investors can easily see it.

The timing matters. Markets opened the week of 10 August with the US dollar near a two-month low after a soft July payrolls report showed the economy shedding 23,000 jobs, its first monthly decline since February, according to CNBC. Into that backdrop lands a slower-burning question credit analysts have raised all summer: what happens to investment-grade credit, and the dollar funding under it, when a handful of technology companies become the single largest force in new bond supply?

What happened#

Two data points crystallised the shift. First, a Nikkei analysis, reported by Fortune and Tom's Hardware, estimated that Alphabet, Amazon, Meta, Microsoft and Oracle carry roughly $1.65tn of AI-related obligations that do not appear as debt on their balance sheets, about 122% of the debt these companies actually report. Meta alone accounts for an estimated $420bn against $140bn of recorded debt; Oracle's hidden total is put at $273bn, up nearly 2,900% since 2022. These are estimates built from lease commitments and financing structures disclosed in footnotes, not audited liabilities, so treat the precise number with care. The direction of travel is not in doubt.

Second, the official sector has started to name the risk. The Bank of England's July 2026 Financial Stability Report noted that the five AI hyperscalers made up just 3% of outstanding US investment-grade debt at the end of 2025 but accounted for more than 15% of year-to-date issuance by early May. Across all currencies, the report said, hyperscaler investment-grade issuance this year is broadly comparable in scale to UK gilt issuance over the same period. Goldman Sachs, summarised by 24/7 Wall St, warned of an AI "debt tsunami" and estimated hyperscaler leverage roughly doubled to 1.8 times in six months.

How a cash-rich sector became a borrower#

For most of the past decade, big technology was the opposite of a credit story. These firms generated so much cash that the debate was about buybacks, not funding. AI infrastructure broke that pattern. Training and serving frontier models requires physical plant on an industrial scale: chips, servers, networking, buildings and, above all, power. Morgan Stanley estimates the five largest hyperscalers will spend around $805bn on capital expenditure in 2026, up from $261bn in 2024, with a projection above $1tn for 2027 (all estimates). CreditSights puts 2026 hyperscaler capex near $602bn, roughly three-quarters of it AI-related.

Cash flow can no longer cover that. Alphabet turned free-cash-flow negative in the second quarter for the first time since its listing, a direct result of the capex surge, according to commentary compiled by Intellectia. When internal cash falls short, companies borrow. The sector that used to supply capital to markets has become a large consumer of it.

The mechanics explain why the number is so hard to pin down. Much of the borrowing does not take the form of a plain corporate bond. Instead, developers and cloud providers set up special purpose vehicles (SPVs), separate legal entities that own a data centre and raise the debt against it. Meta's Hyperion project in Louisiana is the template: a roughly $30bn private-credit deal placed inside an SPV called Beignet Investor, structured with Blue Owl Capital, with about $27bn of loans from PIMCO, BlackRock, Apollo and others and $3bn of equity, as detailed by Global Data Center Hub. None of that debt landed on Meta's balance sheet, which left room to issue $30bn of ordinary corporate bonds on top. Reporting collated by Longbridge suggests Oracle, xAI, Meta and CoreWeave have together moved around $120bn of AI-related debt off their books this way.

Market implications#

The effect that matters most is on investment-grade credit. When a small group of issuers floods a market, price concessions rise. Sage Advisory reports that the coverage ratio on hyperscaler deals, a measure of how many times an offering is oversubscribed, fell from nearly five times in February to below two times by July. New-issue concessions have averaged around 12 basis points against roughly 2.5 for the broader market. Demand has not vanished, but investors want more compensation to keep absorbing supply. Fortune framed the mid-July picture bluntly: issuance rising, appetite thinning.

For equities, the read-across is a change in the risk itself. AI exposure used to be a volatility problem you could size through your equity book. Now it is also a credit and duration problem. If spreads on hyperscaler paper widen, the pain reaches pension funds, insurers and investment-grade ETFs that bought it as safe ballast rather than an AI bet.

The foreign exchange angle is underappreciated. The BoE flagged that the heaviest footprint is in US dollar credit. A concentrated wave of dollar issuance interacts with the dollar's own funding dynamics, and it arrives just as a weak labour market pulls the dollar lower and reprices the Fed path. That is one more variable in an already crowded FX picture across the G10 and, through funding costs, into emerging markets.

For private credit, this is both an opportunity and a concentration risk. Blackstone, Blue Owl, Apollo, PIMCO and BlackRock are originating much of the SPV debt, and the 144A market is filling gaps that bank balance sheets cannot, as Leech Tishman describes. The same names keep appearing on the same projects, exactly the kind of correlation that looks fine until it does not.

Reading the structure#

Two ideas do most of the work here. The first is off-balance-sheet financing. If a data centre is owned by an SPV the technology firm does not consolidate, the SPV's debt sits outside the parent's reported liabilities, even though the parent may still be economically on the hook through purchase commitments, leases or guarantees. Analysts rebuild a "true" leverage picture from the footnotes, and the gap between reported and economic leverage is the whole point of the $1.65tn estimate.

The second is credit-spread supply and demand. A bond's spread over government yields is the price for taking that credit and liquidity risk. When supply rises faster than the pool of willing buyers, spreads widen to clear. The coverage ratio and the new-issue concession are the tells: a 12-basis-point concession on a large, frequent issuer says the market will keep buying, but only at a discount that grows with each deal. For quantitative credit investors, the signal is that hyperscaler paper now behaves less like a rates product and more like a sector with its own supply beta, which changes hedging and factor exposures built on the old assumption that these were among the safest corporate names available.

Critical analysis#

The bear case is straightforward. Leverage is rising, much of it is obscured, buyer appetite is thinning, and the underlying assets are GPUs and buildings whose economic life and returns are unproven at this scale. If AI revenue disappoints, the debt does not care.

The bull case deserves equal weight. The absolute stock of AI debt is still modest against these companies' cash generation, and the BoE judged that the limited outstanding stock at the start of 2026 had helped contain the immediate risk. The off-balance-sheet number mixes genuine leverage with ordinary lease and supply commitments that are not the same as a margin call. Deals are still clearing, often several times oversubscribed. Wider spreads are the market working, not breaking.

Two things could turn a manageable build-up into something worse. One is crowding out, where AI issuance raises the cost of credit for everyone else, a risk the BoE named while stressing it has not materialised yet. The other is correlation: the same handful of lenders, the same SPV structures, the same power-constrained regions. Diversification on paper can hide concentration in practice.

Off-balance-sheet vehicles funding a hot asset class, arranged by a tight club of financiers, is not a new plot. Enron used SPVs to move debt off its books, and the 2008 crisis was amplified by structured vehicles holding assets whose risk was mispriced. None of that means AI infrastructure is the next subprime. The cash flows behind these borrowers are real and enormous, which was not true of a synthetic CDO. The closer historical frame is the telecoms build-out of the late 1990s, when genuine demand for capacity was financed with debt on utilisation assumptions that proved years early. This looks less like a bubble in search of a pin than a classic capex cycle, with the twist that the financing has become unusually complex.

Key takeaways#

The AI trade is now a credit trade as much as an equity one, and investors positioned only in equities are watching half the board.

Off-balance-sheet structures make true leverage hard to measure; the $1.65tn estimate is a floor for the debate, not a settled figure.

Investment-grade credit is absorbing the supply, but at rising concessions, an early-warning gauge worth tracking deal by deal.

The heaviest footprint is in US dollar credit, tying the build-out to dollar funding and FX at a delicate moment for rate expectations.

Central banks have moved from watching the risk to naming it, which usually precedes closer supervisory attention.

Frequently asked questions#

Is this an AI bubble? It is better described as a debt-financed capex cycle. Whether it becomes a bubble depends on whether AI revenue justifies the spending and whether the financing stays orderly. The evidence so far points to strain, not collapse.

Why does off-balance-sheet debt matter if the companies are so profitable? Because it changes the true leverage picture and the market's ability to price risk. Profitable firms can still face funding stress if lenders reassess terms all at once.

How would trouble here affect ordinary investors? Through investment-grade bond funds, pension and insurance holdings, and index ETFs that own hyperscaler paper as supposedly safe assets, plus any knock-on effect on the cost of credit.

Does the weak jobs data change the picture? Indirectly. A softer labour market and a lower dollar shift the funding conditions into which this issuance is landing, affecting both spreads and FX.

Is private credit the villain here? No. It is filling a real financing gap that banks cannot. The concern is concentration among a few large lenders and repeated exposure to similar structures.

What should analysts watch next? Coverage ratios and new-issue concessions on hyperscaler deals, the pace of SPV formation, and any sign of crowding-out in non-AI investment-grade credit.

References#

This article is for information only. It is not investment advice, a recommendation, or an offer to buy or sell any security. Figures described as estimates or projections are not certainties. Verified facts, market interpretation and forward-looking views are distinguished in the text.

Related coverage

Markets

AI's Debt Machine Meets a Wary Bond Market

Meta's $12.55bn data-centre bond priced near 7.5% this week as investors began to balk at the record wall of AI debt. Why the financing of the AI build-out — not the technology — is now the market's central question.

10 min read

Quantitative Trading

When the AI Trade Turned on Itself

Leopold Aschenbrenner's Situational Awareness LP lost roughly two-thirds of its value in July and sold a $16bn book to Citadel. The episode is a case study in AI-era crowding, leverage and the fragility of a consensus trade.

11 min read