Who funds AI debt, and who is stretched
As at 15 October 2026 · All amounts US$ unless marked A$ · Marcus Today research
CoreWeave (NASDAQ: CRWV) sold a bond at a 9.625% coupon in June. This week it yielded 14.50%. On Tuesday the IMF named the AI financing loop as a top trigger for a market shock. The AI build-out is now running on borrowed money, and the bond market is starting to separate the borrowers who can pay from those who can’t.
In our view the big four hyperscalers – the giant cloud providers – can afford what they are borrowing. The strain sits further down the chain: Oracle (NYSE: ORCL), the neoclouds (specialist GPU cloud providers) and the project loans whose only real customer is an AI lab that still loses money. This is a companion to our AI Ownership Map. That piece showed who owns what. This one shows who owes what, to whom, and at what price.
How much has AI borrowed?
US$274bn
Investment-grade bonds sold by Alphabet, Amazon, Meta, Microsoft and Oracle in 2026, to late September. They averaged about US$30bn a year in 2020–24.
To 29 September 2026
US$1.2 trillion
Lease commitments held off balance sheet. More than US$820bn of it is for data centres that are not yet built.
July 2026
14.50%
Yield on CoreWeave’s 9.625% 2032 bond, sold at par in June.
Close, 14 October 2026
215–260
Oracle’s five-year credit default swap in basis points, up from about 57 in mid-2025. A credit default swap (CDS) is insurance against a default; the higher the price, the higher the perceived risk.
Late September to 7 October 2026
Morgan Stanley expects about US$570bn of AI-related debt worldwide in 2026, counting bonds, loans and private credit. Goldman Sachs expects debt to fund about a third of hyperscaler capital spending this year, up from about a quarter in 2025.
Follow the debt from lender to AI lab
Read the map from top to bottom. The lenders come first. Next are the special purpose vehicles (SPVs) and project companies. These are separate legal entities that borrow for a single site, so the debt doesn’t appear on the tech company’s balance sheet. Then come the borrowers, and at the bottom the AI labs whose contracts are meant to pay it all back.
Tap a type to show or hide it. Tap a company to list its links, and tap it again to clear.
marcustoday. · As at 15 October 2026 · Arrows point to the borrower, the vehicle being backed or the company being paid.
The cost of money has gone up, a lot for some
Every AI bond we tracked now yields more than when it was sold. Part of that is rates: the US 10-year Treasury yield is about 5.24%, near its highest since 2002. What separates the borrowers is the spread – the extra yield over Treasuries. Big Tech bonds have moved up 0.30–1.12 percentage points. Oracle’s have moved 1.69–2.45 points. CoreWeave’s have moved 2.77–4.88 points, into distressed territory.
Oracle and CoreWeave bonds have repriced the most
Yield to worst (%), with the change since issue in basis points. The dashed line is the 10-year Treasury at 5.24% (1 October). Yields as at 12–15 October 2026. The Hyperion SPV is at about 91.6c on 7 October.
The cost of insuring Oracle’s debt has quadrupled
Oracle five-year CDS (basis points). *October readings range from 203 to 260, so treat the top of the range as unconfirmed.
Which borrowers are under real stress?
No major AI data-centre operator or neocloud has defaulted. Character.AI’s Chapter 11 filing this month is an app business, not infrastructure. Bond and CDS markets are already pricing distress in a handful of names. The common thread is reliance on a few customers that don’t make money themselves. Free cash flow (FCF) below is operating cash flow less capital spending. Interest cover is EBITDA (earnings before interest, tax, depreciation and amortisation) divided by interest expense.
Cash flow does not cover the interest
CoreWeave (NASDAQ: CRWV)
CoreWeave has US$35.6bn of debt. Q2 adjusted operating income was US$128m against US$640m of net interest, so operating profit covered about 0.2x of it. FCF was about -US$10.5bn in the first half, and capex guidance is US$35–39bn. Its bonds yield 12.02–14.50%, and in July its CDS implied about a 50% chance of default over five years. Microsoft was about 67% of FY25 revenue. US$4.4bn falls due in the rest of 2026 and US$6.2bn in 2027.
Bonds at about 80c · CDS about 855 basis points (July) · S&P B+
Core Scientific (NASDAQ: CORZ)
The coupon on its US$3.3bn of 7.75% 2031 notes costs about US$256m a year. Annualised adjusted EBITDA is about US$164m. CoreWeave rents about half its capacity, so this is CoreWeave risk at one remove.
Interest cover below 1x (our estimate)
TeraWulf (NASDAQ: WULF) and Cipher Digital (NASDAQ: CIFR)
Both have negative EBITDA against quarterly interest of US$56m and US$67m, on US$5.2bn and US$6.0bn of debt. Their secured notes are backed by Google’s guarantee of their tenant Fluidstack’s rent. That protects bondholders more than shareholders. Cipher fell 18% this week after selling an US$850m zero-coupon convertible bond.
Negative EBITDA · heavy dilution
Project Jupiter loan (Oracle, New Mexico)
The US$18bn bank construction loan trades at 89–91c in the dollar, and the site is about 26% built. Oracle served a force majeure notice on 24 September, citing power delays. Santander and Jefferies hold more of the loan than they planned to.
Loan at 89–91c (18 September)
Galaxy Helios II
US$3.5bn of 9.875% notes, priced at about 550 basis points over Treasuries. The 260MW site is 100% leased to CoreWeave for 15 years. Interest would take about 82% of projected 2028 net operating income.
Single tenant: CoreWeave
Negative cash flow, leaning on markets
Oracle (NYSE: ORCL)
FCF has been negative for five straight quarters, including -US$23.7bn in FY26. S&P expects about -US$42bn in FY27. Debt is US$125.3bn, plus US$288bn of leases signed but not yet started. Remaining performance obligations, or contracted backlog, are US$664bn. OpenAI, the largest customer, doesn’t start paying until 2027. Oracle has raised about US$25bn of new equity to protect its rating and says it will sell no more bonds this year.
S&P BBB- (one notch above junk) · Moody’s Baa2 negative · long bonds yield 8.31–8.55%
SoftBank Group (TYO: 9984)
SoftBank is borrowing to fund its US$64.6bn OpenAI stake. It sold the largest high-yield bond on record, US$11.1bn at up to 9.75%, and a ¥1 trillion retail bond, and it drew a US$40bn bridge loan. It also has margin loans of about US$10bn against its OpenAI shares and US$25bn against its Arm shares. OpenAI is about 30% of its investment assets.
S&P BB+ negative · CDS about 383–400 basis points
OpenAI (private)
OpenAI itself carries almost no debt beyond an undrawn US$4bn revolving credit line. It is, however, the tenant behind Oracle’s Stargate sites and much of CoreWeave’s backlog. It burned US$3.7bn in Q1 and has up to US$665bn of compute commitments. Projections put its FCF at -US$278bn over 2026–30.
Unrated · not expected to be cash-flow positive before about 2030
Applied Digital (NASDAQ: APLD)
US$6.4bn of project debt against Q1 FY27 adjusted EBITDA of US$64m. JPMorgan cut its project bonds to Neutral on 12 October, citing another US$7.8–13bn of supply to come. A roughly US$3.5bn deal for a Meta project was tested in September and never priced.
Trailing FCF about -US$2.8bn
Super Micro Computer (NASDAQ: SMCI)
FY26 operating cash flow was -US$6.8bn on revenue of US$39.1bn. It has US$8.7bn of bank debt and convertibles, plus US$4.2bn of 7% preferred shares. Its internal controls were reported as not effective, and there are subpoenas and an indictment of three individuals.
FCF -US$7.0bn
Nscale (private, IPO filed)
Nscale filed for a US listing on 18 September. It has about US$4.9bn of secured loans at up to 750 basis points over SOFR, the US benchmark lending rate. It has US$24bn of equipment on order. Its Anthropic contracts of up to US$44.6bn have no committed financing yet.
Funding gap
Can pay, but the structure carries risk
Meta’s off-balance-sheet SPVs
Meta Platforms (NASDAQ: META) has two such vehicles. Hyperion (Beignet Investor, US$27.3bn) hit a record low of about 91.6c on 7 October. El Paso (Sopaipilla Investor, US$12.5bn) priced at 7.534%, a yield more typical of junk bonds. Both are rated A+ because Meta guarantees what the sites will be worth (a residual value guarantee). Neither appears in Meta’s reported debt.
A+ · about US$41bn of guarantees
SpaceX (NASDAQ: SPCX), now including xAI
After its June IPO, SpaceX refinanced xAI’s 12.5% junk debt into US$25bn of BBB bonds. Q2 capex was US$18.4bn against EBITDA of US$3.5bn. It is in talks for another US$40bn, led by Apollo, to buy Nvidia chips. Its 2056 bond widened from 175 to 231 basis points over Treasuries within a month of issue.
BBB / Baa1 · about US$100bn of cash
Broadcom (NASDAQ: AVGO) and Nvidia (NASDAQ: NVDA) vendor finance
Broadcom is in talks on US$60–70bn of SPV debt that buys its chips and leases them to customers such as Anthropic, with Broadcom offering partial guarantees. Nvidia has guaranteed up to US$105bn of leases on an OpenAI campus in Ohio, and US$108.5bn in total. Ratings agencies don’t yet count either as debt.
Hidden leverage
Nebius (NASDAQ: NBIS) and IREN (NASDAQ: IREN)
Both fund themselves mainly with convertible bonds; Nebius sold about US$9bn in 2026. Customer prepayments flatter their operating cash flow. Nebius’s Q2 capex of US$5.7bn was about 24 times its EBITDA. IREN’s Microsoft-backed debt is rated A, but its Q4 EBITDA margin was only 14%.
Shares down more than 40% since July
Intel (NASDAQ: INTC)
Net debt is US$20.8bn and FCF is thin, at about US$2.8bn (our estimate). Q2 adjusted FCF was -US$8.4bn. It remains investment grade, with interest cover of about 14x, and the US Government owns about 8.4%.
BBB / Baa2
Low stress: Microsoft, Alphabet, Amazon, Meta (corporate debt), Nvidia, TSMC, Micron, SK Hynix and Samsung hold net cash or close to it. Their bonds have fallen with rates and supply, not on credit worries. Alphabet’s first negative FCF quarter since its 2004 IPO (Q2 2026) and Amazon’s negative trailing FCF are worth watching all the same.
Who holds the debt?
Most of the money doesn’t come from tech. Banks write the construction loans and then try to sell them on. Bond funds and insurers buy the long-dated SPV paper. Sovereign wealth funds come in as equity partners and as investors in the private credit managers that do the lending.
TD, Pimco and Goldman are the biggest US data-centre lenders
US data-centre debt by lender (US$bn), as at March 2026, out of about US$1.3 trillion in total. Later deals, such as the Michigan bonds, are not included.
Banks
JPMorgan and MUFG led the US$38bn Vantage loans for Oracle’s Texas and Wisconsin sites. The sell-down took more than six months, and some of the loans went to non-banks at a discount. SMBC, BNP Paribas, Goldman Sachs and MUFG are agents on the US$18bn Jupiter loan, now at 89–91c. JPMorgan, Goldman, Mizuho, SMBC and MUFG lead SoftBank’s US$40bn bridge loan. Chicago Fed staff put large-bank commitments to “AI-adjacent” borrowers at about US$450bn. Several banks have explored significant risk transfers, which sell the credit risk to outside investors, to cut their exposure.
Private credit and fund managers
Pimco is the largest single buyer. It took about US$18bn of Meta’s Hyperion bonds and about US$10bn of Oracle’s Michigan bonds. Apollo and its insurer Athene lent about US$7bn against xAI’s GPUs, and Apollo co-led the US$35bn Anthropic chip SPV with Blackstone. Blackstone also anchored CoreWeave’s investment-grade US$8.5bn loan and leads Firmus’s US$10bn facility. BlackRock, GIP and HPS own 80% of Meta’s El Paso vehicle. The Bank for International Settlements (BIS) puts private credit loans to AI firms above US$200bn.
Sovereign wealth funds
Gulf and Singapore money comes in mostly as equity, with indirect credit exposure. MGX led the US$40bn takeover of Aligned Data Centers with BlackRock’s GIP. ADIA and GIC put about US$3bn into Vantage. QIA has a US$3bn digital-infrastructure partnership with Blue Owl and a US$20bn partnership with Brookfield. Mubadala is an investor in Apollo’s direct lending. SoftBank is now seeking up to US$100bn from Gulf investors. We found no sovereign fund named as a direct buyer of a specific AI bond.
Insurers, retail investors and governments
Insurers buy the long, fully amortising private placements, such as Lambda’s US$1bn 6.78% loan. Japanese households bought SoftBank’s ¥1 trillion retail bond. Index investors own the rest by default. AI-linked debt is about 14% of JPMorgan’s main US investment-grade index, so most corporate bond ETFs hold it. The Pentagon’s Office of Strategic Capital is in talks to lend Fluidstack about US$5bn. Australia’s National Reconstruction Fund is putting A$200m into Macquarie Technology (ASX: MAQ).
Why the regulators are worried
- IMF (13 October): named the AI financing loop as a top trigger for a market shock. Hyperscalers have put about US$300bn of equity and US$650bn of debt into AI firms, which have committed about US$1.4 trillion of revenue back to them. It warned of a “triple whammy”: equity losses, trouble servicing debt and commitments falling through, all at once.
- BIS: calls off-balance-sheet joint ventures and SPVs “shadow borrowing”. It says bank funding lines to these vehicles can pass shocks on to the banks.
- Bank of England (July): said AI firms’ use of credit has “accelerated rapidly” while hyperscaler free cash flow is falling.
- US Federal Reserve (May): flagged that capital spending is “increasingly funded by debt, creating leverage in the system”.
- Morgan Stanley: sees US$2.9 trillion of data-centre spending to 2028 against US$1.4 trillion of hyperscaler cash flow. That leaves a US$1.5 trillion funding gap.
- Goldman Sachs: estimates AI revenue is about US$230bn a year short of breakeven.
- Bond buyers: hyperscaler order books fell from about 5x covered in February to under 2x by July. AI issuers trade about 115 basis points over Treasuries, against 78 for the investment-grade index. Tech used to trade tighter than the index.
- GPU lifespans: lenders now assume GPU collateral lasts 3–4 years. Nvidia argues up to 10. Michael Burry argues depreciation schedules overstate 2026–28 profits by about US$176bn.
Where are the ASX links?
- NEXTDC (ASX: NXT): A$3.35bn of drawn debt, A$7.1bn undrawn and pro forma liquidity of about A$9.8bn. In September it priced A$1.1bn of 1.75% convertible notes due 2031 at a 32.5% premium. FY27 capex guidance of A$5.25–5.75bn is about 14 times EBITDA guidance of A$385–410m. Liquidity is strong, but the S7 campus depends on OpenAI as anchor tenant.
- Macquarie Technology (ASX: MAQ): a modest A$500m secured revolver and the A$200m NRF hybrid note. FY26 EBITDA was A$115.9m, against FY27 capex guidance of A$485–506m.
- Firmus: Blackstone leads a US$10bn debt facility with Coatue. The ASX float was withdrawn on 9 October.
- AirTrunk: owned by Blackstone and CPP Investments. It refinanced A$16bn with more than 60 banks in 2025 and is in talks for about S$2bn in Singapore.
- Macquarie Group (ASX: MQG): lends to Applied Digital and Nscale. We could not confirm the big four banks’ shares of the AirTrunk, NEXTDC and CDC facilities.
What it means for investors
In our view the AI debt story has two halves. The big four hyperscalers can afford what they are borrowing. Their bonds are down because Treasury yields are near 2002 highs and supply is enormous, not because anyone doubts they will pay. The stress is further down the chain, in Oracle, CoreWeave, the former bitcoin miners and the project loans that depend on AI labs.
The debt is also less visible than it looks. Meta’s two SPVs, Oracle’s Stargate sites, Broadcom’s chip vehicles and Nvidia’s guarantees hold several hundred billion dollars of obligations outside reported debt. Most of it relies on the same two customers, OpenAI and Anthropic. If either stumbles, the losses fall on banks still holding construction loans, on bond funds like Pimco and on private credit funds that investors are already trying to leave.
For Australian investors the direct exposure is small: NEXTDC, Macquarie Technology and, through bank syndicates, AirTrunk. The bigger channel is bond funds, because AI-linked debt is about 14% of the main US investment-grade index. We are watching CoreWeave’s 2027 maturities, the Jupiter loan price, Oracle’s CDS and whether the Anthropic and OpenAI listings go ahead.
Every major AI debt deal we tracked
Each card is a bond, loan or credit facility. “At issue” is the coupon, or the margin over a benchmark, when it was sold. “Now” is the latest yield or price we could find. Figures in brackets are the spread over Treasuries in basis points. Search for a lender such as Pimco or MUFG, pick a sector, or show all deals.
How big are the borrowers?
US$bn unless marked. Market values are around 14–15 October 2026 where available. Negative net debt means net cash. Oracle’s debt is at 31 August 2026; other measures put it at US$117bn or US$167bn on different definitions.
Key terms
- Basis points
- Hundredths of a percentage point. 100 basis points equals 1%.
- Investment grade and high yield
- Bonds rated BBB- or higher by S&P (Baa3 by Moody’s) are investment grade. Anything lower is high yield, or junk.
- Yield to worst
- The lowest yield a bond can deliver if the issuer repays early on the terms allowed. It is the standard measure of a bond’s current yield.
- Credit default swap (CDS)
- A contract that pays out if a borrower defaults. Its annual cost, in basis points, is a live gauge of credit risk.
- Special purpose vehicle (SPV)
- A separate company set up to borrow for one project, which keeps the debt off the sponsor’s balance sheet.
- Residual value guarantee
- A promise by the tenant to cover any shortfall if the asset is worth less than the debt at the end of the lease.
- Delayed-draw term loan (DDTL)
- A loan drawn in stages as equipment is bought. CoreWeave funds its GPUs this way.
- SOFR
- The Secured Overnight Financing Rate, the US benchmark that floating-rate loans are priced over.
- Neocloud
- A specialist cloud company that rents out GPUs for AI, such as CoreWeave, Nebius or Lambda.
- Hyperscaler
- The giant cloud providers: Amazon, Microsoft, Alphabet, Meta and, increasingly, Oracle.
General advice only. This content does not take into account your objectives, financial situation or needs.