Who is lending to AI, and can the borrowers repay?

Debt now funds a growing share of the artificial intelligence (AI) build-out: Goldman Sachs expects about a third of the largest companies’ 2026 capital spending to be borrowed.¹ Morgan Stanley estimates that AI-related debt issuance reached about $450 billion worldwide by early September 2026, more than double the total for all of 2025, according to the Bank of England’s Financial Policy Committee.² Bloomberg data count at least $334.5 billion of bonds and loans raised for AI infrastructure by late July, against $185.5 billion in the whole of 2025.³ The largest technology companies, the so-called hyperscalers, now account for almost 10% of global investment-grade corporate bond issuance, up from 0.5% in 2024.⁴

The borrowing follows the spending. The five largest hyperscalers, Alphabet, Amazon, Meta, Microsoft and Oracle, invested $416 billion in 2025, almost six times their 2019 level, and analysts expect $799 billion in 2026 and $1.09 trillion in 2027 (Exhibit 1).⁵ J.P. Morgan analysts expect about $4.1 trillion of AI-related capital expenditure to be financed with debt between 2026 and 2030.⁶

The strongest borrowers can carry this. The hyperscalers’ net debt, including leases signed but not yet started, was 1.3 times their earnings before interest, taxes, depreciation and amortization (EBITDA) in the second quarter, against 2.7 times for the median US investment-grade company.⁷ The pressure sits further out. Oracle, rated one notch above high yield, has $125 billion of borrowings and $288 billion of data center leases that have not yet started.⁸ CoreWeave, the largest of the independent AI cloud providers, pays interest equal to 23% of its revenue,⁹ and the price of insuring its debt against default passed 1,000 basis points in October, a level usually read as distress.¹⁰

Lenders have noticed. Investors and lenders have turned to credit default swaps (CDS), the contracts that pay out if a borrower defaults, and banks to risk-transfer deals that hedge their data center loans; trading in protection on AI names has grown fast.¹¹ The IMF now lists an AI-driven market correction as one of three salient risks to the financial system.¹²

This report sets out who is borrowing and through which channels, how many AI companies there are and where, which lenders have provided the credit and how much they hold, how able the main borrowers are to repay, and how lenders are protecting themselves. It closes by comparing today with the telecom and dot-com bust of 2000 to 2002 and the credit crisis of 2007 and 2008. It describes markets and published filings; it does not value any company, recommend any security, or forecast defaults (see sidebar, “About the research”).

About the research

• Company filings. Form 10-Q and S-1 filings with the US Securities and Exchange Commission, and company announcements.
• Official analysis. The IMF, the Bank for International Settlements (BIS), the Bank of England, the US Federal Reserve and the Financial Crisis Inquiry Commission.
• Market data. LSEG first-half 2026 reviews, BIS derivatives statistics, FRED, the Shiller dataset, J.P. Morgan Asset Management and the Stanford AI Index.
• Reporting. Reuters, Bloomberg, the Financial Times, CNBC and rating agencies, as cited.

The research covers January to October 2026, with data read between October 24 and 27, 2026. Bank estimates are reported as their authors published them and are not combined, because they measure different things. Deal amounts are as announced; individual lenders’ shares of syndicated loans are rarely disclosed and are not estimated here. Market prices move daily and are given with their dates.

A trillion-dollar build-out, increasingly paid for with debt

Until recently the hyperscalers paid for most of their data centers out of their own cash. The Federal Reserve Bank of Dallas puts the investment made since 2023 at $500 billion to $600 billion, much of it funded from retained earnings, and the coming wave at $3 trillion to $5 trillion over three to five years.¹³ As spending outgrew cash flow, free cash flow fell and borrowing rose.¹⁴ The money now arrives through four channels, each with a different lender base.

Bonds: The largest channel

Corporate bonds remain the hyperscalers’ main source of debt. Goldman Sachs counts $108 billion of hyperscaler debt issued in 2025 and $194 billion in 2026 by early August, and nearly $500 billion of AI-related debt overall, of which the hyperscalers were only 40%.¹⁵ AI borrowers supplied 18% of US investment-grade issuance in 2026, up from 7% in 2025 and 1% in 2024, and 40% of bonds with maturities over 15 years.¹⁶ The first half of 2026 brought some of the largest corporate bonds on record: Amazon raised $36.9 billion in March, and Meta, Nvidia and Oracle about $25.0 billion each, with Alphabet at $19.9 billion.¹⁷ Even in sterling, hyperscalers supplied 47% of corporate bond issuance this year.¹⁸

Bank loans: Project finance for the campuses

Banks lend for specific campuses and as bridges to bonds or listings. In April, a $22.75 billion project-finance loan for Vantage Data Centers’ Texas campus, leased to Oracle, was the third-largest syndicated loan of the half-year worldwide.¹⁹ It forms part of a $38 billion package, with a Wisconsin loan of $14.75 billion, led by JPMorgan and MUFG.²⁰ In March, SpaceX, which now includes xAI and X, signed a $20 billion unsecured bridge loan that repaid the debt of xAI and X, before its $86.25 billion stock market listing in June.²¹ SoftBank borrowed $40 billion for a year to fund its investment in OpenAI.²²

Private credit and off-balance-sheet vehicles

The fastest-growing channel keeps debt off the hyperscalers’ balance sheets. The BIS describes the common structure: a joint venture or special-purpose vehicle builds the data center, raises debt from private credit funds and insurers, and leases the capacity to the hyperscaler, which holds a minority stake and may give guarantees. The BIS calls these obligations “shadow borrowing.”²³ Meta’s Hyperion campus in Louisiana is the reference case. Funds managed by Blue Owl Capital own 80% of the joint venture, Meta 20%; Meta leases the campus for an initial four years and guarantees part of its residual value for 16 years.²⁴ The vehicle sold $27 billion of 6.581% notes due 2049 to PIMCO and other bond investors.²⁵ Morgan Stanley puts the financing and leasing structures kept off balance sheets across the sector at some $3 trillion, and expects private credit to fund $700 billion of data center spending from 2026 to 2028.²⁶

Equity and equity-linked money

Borrowers have also turned back to shareholders. Alphabet priced an equity raise of about $85 billion in June, and Oracle plans to raise $45 billion to $50 billion in 2026, about half through shares and convertibles.²⁷ Oracle sold $19.9 billion of shares through its at-the-market program in a single quarter.²⁸ Some of this money circulates within the sector. The IMF reviewed 14 transactions since 2023 that combined about $300 billion of equity and $650 billion of debt in exchange for $1.4 trillion of future revenue commitments, often between suppliers and their own customers.²⁹

Thousands of AI companies, a few dozen borrowers

The number of AI companies is large and rising. The Stanford AI Index counts 3,499 AI companies that received new funding in 2025, up 70.8% on 2024. The United States accounted for 1,953 of them, 56% of the total, ahead of the United Kingdom with 172 and China with 161 (Exhibit 2).³⁰ Money is more concentrated still: private AI investment in the United States reached $285.9 billion in 2025, 23 times China’s $12.4 billion. Among Gulf states, Saudi Arabia drew $2.0 billion.³¹

Debt reaches only a small part of this population. Start-ups are funded with equity; the borrowers are the hyperscalers, Oracle, chip makers such as Nvidia, the independent AI cloud providers known as neoclouds, such as CoreWeave, Nebius and IREN, and the data center developers and vehicles that build for them. The AI laboratories rely mainly on partners that borrow to build their computing capacity; OpenAI, for example, has a $300 billion computing agreement with Oracle.³² These laboratories are valued like large listed companies. Anthropic raised $65 billion at a $965 billion valuation in May,³³ and OpenAI is reported to be raising $30 billion at $1.25 trillion.³⁴

In public markets, AI-related industries make up 53.2% of the S&P 500, and the ten largest companies 40.6% of its value, against 18% to 21% at the start of 1985, 1995, 2005 and 2015.³⁵ That concentration links the credit question to the equity market: the borrowers’ ability to refinance depends on investors’ continued confidence in AI earnings.

Who lends, and how much each holds

Large syndicates of US and Japanese banks dominate the loan market for AI campuses, private credit funds and insurers buy the long-dated debt of the vehicles, and bond investors hold the hyperscalers’ own debt (Exhibit 3). Individual lenders’ shares of a syndicated loan are rarely published, so exposure can be traced reliably only at the level of the deals and of the lender groups. Like the borrowers in Exhibit 2, the lenders come mainly from a few countries, above all the United States and Japan.

Banks

Five names recur. JPMorgan and MUFG led the $38 billion of Oracle-linked Vantage loans, JPMorgan led the $7.1 billion construction loan for Crusoe’s Stargate campus in Abilene, Texas, and Morgan Stanley and MUFG provided $2.6 billion to CoreWeave in 2025.³⁶ Sumitomo Mitsui Banking Corporation (SMBC) led every US digital infrastructure loan league table for 2025, with MUFG and JPMorgan in the top three across several categories.³⁷ The same banks arranged the bridges: Goldman Sachs for SpaceX, and JPMorgan, Goldman Sachs and the three Japanese megabanks for SoftBank.³⁸ Agents for CoreWeave’s 2026 term loans include MUFG, Morgan Stanley and, from August, JPMorgan.³⁹

Holding these loans has proved harder than arranging them. The banks behind the $38 billion of Oracle-linked loans took months to find buyers for their shares, Yahoo Finance reported in October.⁴⁰ Credit Agricole was selling down a Hong Kong data center loan after reaching its own lending cap for the sector.⁴¹ Banks also lend to the private credit funds and vehicles that finance AI, which the BIS identifies as a new channel for shocks.⁴² Under the IMF’s adverse scenario, banks holding 22% of global banking assets would end with capital below regulatory limits, up from 18% a year earlier.⁴³

Private credit funds and insurers

Private credit has become a major lender to technology firms. Worldwide, technology borrowers held more than $1 trillion of direct loans from private credit funds in 2025, 44% of the almost $2.5 trillion outstanding, up from 28% in 2019.⁴⁴ Behind private credit stand insurers. Moody’s counts $85.2 billion of AI-linked investments held by US life insurers at the end of June 2026, 1.4% of their invested assets and 24% more than six months earlier. Of these, $32.2 billion is the debt of the five largest technology companies, about 6.6% of all their debt outstanding, including about $10 billion of Oracle’s and $8.5 billion of Amazon’s.⁴⁵

Can the borrowers repay? Three tiers of credit

Ability to repay varies widely across the borrowers, and the evidence sorts them into three tiers (Exhibit 4).

The hyperscalers: Strong balance sheets, thinner cash flow

The five hyperscalers can service their debt from operating income several times over, and the IMF judges their interest coverage comfortable. But their combined interest bill doubled in a year to nearly $6 billion in the second quarter, their bond spreads have widened by about 30 basis points more than the investment-grade index since April, and their bonds are approaching the weight of banks in the benchmark indices, which the IMF warns could spread any repricing across the credit market.⁴⁶ Goldman Sachs estimates that the hyperscalers need roughly $300 billion of annual AI revenue to break even on their investment.⁴⁷

Oracle: A contracted backlog, a long wait for cash

Oracle shows both the promise and the strain. Its contracted backlog rose 46% in a year to $664 billion, which Oracle attributes to large cloud contracts; the largest reported is a $300 billion computing deal with OpenAI. In the quarter to August, it spent $28.5 billion on capital projects, 3.4 times as much as a year earlier, against $23.1 billion of operating cash flow, leaving free cash flow of minus $5.4 billion. Interest expense rose 55% to $1.43 billion, still covered 4.7 times by operating income.⁴⁸ On top of $125.3 billion of borrowings and $43.8 billion of recognized leases, Oracle has committed to $288 billion of data center leases starting between late 2026 and 2029, for terms of 15 to 19 years.⁴⁹ S&P cut Oracle to BBB-, the lowest investment-grade rating, in July and said it could cut again if leverage stays above four times or if Oracle is not on a path to positive free operating cash flow by fiscal 2029. Analysts’ consensus expects free cash flow of minus $23.7 billion this year and minus $36 billion in 2027.⁵⁰

The neoclouds: Growth funded at high-yield rates

CoreWeave doubled its revenue to $2.58 billion in the second quarter, but its operating loss of $49 million left nothing to pay $592 million of contractual interest, and it lost $626 million.⁵¹ Effective interest rates on its term loans and senior notes range from 7% to 15%; only its convertible notes cost less. It owes $35.6 billion of principal, of which $10.6 billion falls due by the end of 2027, almost twice its $5.5 billion of cash (Exhibit 5). The case for repayment rests on contracts: $103.7 billion of backlog and 98% of revenue from committed contracts.⁵² About two-thirds of the backlog is with investment-grade customers such as Microsoft and Meta, according to Deutsche Bank.⁵³ The risk is concentration, with three customers producing 72% of second-quarter revenue. Analysts expect negative free cash flow of $10.1 billion in the third quarter and estimate that CoreWeave may need to raise about $30 billion in 2027, and credit desks told CNBC that the high-yield market was on hold until after the US midterm elections.⁵⁴

The laboratories: Commitments ahead of revenue

The AI laboratories are the end customers of much of this debt. Their commitments to buy computing power run for many years, and they are funded mainly by equity from their suppliers and investors. OpenAI’s $110 billion funding round in February drew $30 billion each from Nvidia and SoftBank and $50 billion from Amazon, $15 billion of it up front; OpenAI also agreed to spend a further $100 billion on Amazon’s cloud over eight years.⁵⁵ Lenders to Oracle, CoreWeave and the data center vehicles are, in effect, lending against these contracts. If a laboratory’s growth slowed, the strain would pass down the chain of leases and loans that the IMF describes as circular.⁵⁶ For a borrower such as CoreWeave, that strain would arrive as the maturities in Exhibit 5 fall due.

Lenders are buying protection

A credit default swap pays the buyer if a borrower defaults; the price, quoted as an annual spread in basis points, is the market’s price of that risk. Banks buy protection to hedge the loans they hold, investors to hedge bonds or to bet on a decline. In 2026 AI borrowers became some of the most actively traded US corporate names.

Oracle was the most traded US corporate name in March, with 311 trades reported to US swap repositories, more than twice the next company.⁵⁷ Its five-year spread stood at 295 basis points on October 19, up 287% in a year and in the range of BB and B credits, according to S&P Global Market Intelligence.⁵⁸ CoreWeave’s reached 1,145 basis points on October 21, up 179% in a year, while its 9.75% bond due 2031 yielded 15.78% and its market value fell from a peak of about $90 billion to about $31 billion.⁵⁹ The Hyperion notes traded at 91 cents on the dollar in late September, a yield of 7.55% and a spread of about 230 basis points over Treasuries, against 185 at launch.⁶⁰

The hedging is deliberate. Bloomberg reported in November 2025 that the Oracle CDS market had grown into a hedge against an AI crash, and in March 2026 that JPMorgan was offering clients a basket of CDS to hedge AI debt.⁶¹ Banks have also used significant risk transfers, which pass part of a loan book’s losses to outside investors: Morgan Stanley explored one for its data center loans in late 2025, Societe Generale in March 2026, and Deutsche Bank hedged data center risk in a project-finance transfer in September.⁶² In October, the Financial Times reported a surge in default-insurance trading on AI debt.⁶³ The BIS notes that CDS spreads rose most for the lower-rated hyperscalers, reflecting both the volume of new debt and doubts about the projects’ payoffs.⁶⁴

What 2000 and 2007 can and cannot tell us

Two earlier episodes combined new technology, heavy borrowing and credit derivatives. Neither is a template, but each shows how losses travel.

2000 to 2002: The telecom debt behind the dot-com bubble

The dot-com bust is remembered for share prices: the Nasdaq Composite fell 78% from its peak of 5,049 on March 10, 2000, to 1,114 on October 9, 2002.⁶⁵ The credit losses came from telecom companies that had borrowed to lay fiber ahead of demand. US investment in communications equipment more than doubled from about $62 billion a year in 1996 to over $135 billion in 2000, in 1996 dollars, while advances in transmission doubled the capacity of a single fiber about every nine months. The market value of S&P 500 telecom companies fell by about $700 billion from 2000 to 2002.⁶⁶ Telecom firms produced 55% of all defaulted high-yield bond dollars in 2001 and 52% in 2002, more than $50 billion in 2002 alone, and lenders recovered only 16 cents on the dollar, against 25 cents across all defaults.⁶⁷

The parallel with today lies in the assets. Fiber outlasted the companies that laid it. AI chips, by contrast, are depreciated over five to six years by the major technology firms, against ten to 15 years or more for the buildings,⁶⁸ so the collateral behind GPU-backed loans could lose value faster than the loans amortize.

Valuations today are high, though below the 2000 peak (Exhibit 6). The S&P 500 traded at 19.0 times expected earnings at the end of September, against 25.2 times in March 2000, while the Shiller cyclically adjusted price-to-earnings ratio (CAPE), at 40.7 in October, is 92% of its December 1999 record.⁶⁹ One difference is earnings: today’s largest companies earn 36.1% of the index’s profits, which supports much of their 40.6% weight.⁷⁰

2007 to 2008: When credit insurance concentrated the risk

The 2007 crisis began in housing and spread through derivatives. US house prices fell 27% from July 2006 to February 2012.⁷¹ Over-the-counter derivatives had grown to $673 trillion of notional value. AIG sold protection on $79 billion of mortgage securities without being required to hold capital against it and, unable to meet its obligations, received government commitments of more than $180 billion. Of the mortgage securities Moody’s rated triple A in 2006, 83% were later downgraded.⁷² The CDS market itself peaked at $61.2 trillion of notional at the end of 2007 (Exhibit 7).⁷³

Today’s market differs in three ways. It is smaller, at $11.0 trillion at the end of 2025, and 70% of it is cleared through central counterparties that collect margin.⁷⁴ The largest AI borrowers are profitable companies with investment-grade ratings, valued at a lower multiple of their earnings than at the 2000 peak (Exhibit 6), rather than households with stretched mortgages. And much of the demand for protection on AI names comes from lenders hedging their loans; no seller of protection on the scale of AIG has been reported. The resemblances lie elsewhere: debt moved into off-balance-sheet vehicles, ratings resting on the credit of a sponsor, long-dated assets funded by investors who may need liquidity, and exposure concentrated on a few names, as the IMF and the BIS both point out.⁷⁵

Signposts for the months ahead

In order of the timing of the next data point, based on scheduled filings and reported deadlines, these are the indicators we will follow:

1. Hyperscaler results and 2027 capital spending plans. Third-quarter results due from late October will show whether spending guidance for 2027 holds near the $1.09 trillion consensus, and how much of it each company plans to fund with debt.
2. CoreWeave’s third-quarter results and financing plan. Analysts expect negative free cash flow of about $10 billion in the quarter. The $6.2 billion of principal due in 2027 and a possible need of about $30 billion of new capital will test the high-yield market once it reopens.
3. Oracle’s rating. S&P’s triggers are leverage above four times and no path to positive free operating cash flow by fiscal 2029. A cut to high yield could force sales by investors limited to investment-grade debt.
4. Spreads on the most traded names. Oracle’s five-year CDS near 300 basis points and CoreWeave’s above 1,000 are the market’s running measure of AI credit risk; the Hyperion notes show how investors price off-balance-sheet structures.
5. Policy. The IMF has asked authorities to consider sectoral capital requirements and concentration limits on banks’ exposures to AI and private credit.⁷⁶ Any such step would change how much banks can hold.

---

AI borrowing has more than doubled in 2026, and most of it rests on companies that can carry it. The risk lies in the layers further out: the vehicles, neoclouds and project loans whose repayment depends on contracts with a few customers, funded at high-yield rates and secured on equipment that ages quickly. Lenders have begun to price that risk and to hedge it, which is a sign of discipline as much as of worry.

The build-out is likely to continue; the terms on which it is financed may not.

For lenders, investors and companies in the Gulf, where data center projects worth more than $93 billion are active or planned across the GCC,⁷⁷ the lessons are practical. Test the cash flows behind each lease and offtake contract, match the tenor of the debt to the life of the equipment, and watch concentration on a few counterparties. The build-out is likely to continue; the terms on which it is financed may not.