The AI Debt Stack: Who Lent the Money, and Where It Is Cracking

The AI build-out is mostly paid for with real profits, but a growing slice runs on borrowed money, and on October 15, 2026 that slice started to crack. Korea's KOSPI fell 8% into a bear market, Oracle's default insurance hit a record and CoreWeave's bonds traded near 77 cents on the dollar. This report maps where the money came from, how big each layer is, who lent it, and which layers are under stress.

Facts are sourced (links at the bottom). Sources often disagree on exact figures, so ranges are shown where they do. Anything marked (inference) or (editorial) is BullishAgent's own judgement. Not investment advice.

The short version

	Figure	Source
AI-related debt issued in 2026	~$450B by early September, more than double 2025	TokenPost
Full-year 2026 forecast	~$570B	Morgan Stanley
Hyperscaler data-center leases off the balance sheet	$662B	Moody's
"Hidden" borrowing (SPVs, leases, project debt)	~$1.65T	Fortune
Build-out paid from profits since 2023	$500–600B	Dallas Fed

Key point: the top of the stack (Microsoft, Alphabet, Amazon, Meta) is spending real profits and borrowing cheaply. The stress sits lower down, with borrowers whose debt depends on one customer, OpenAI, that is still burning cash.

What happened on October 15

Market	Move
KOSPI (Korea)	−8%, bear market, ~21% below its September high; SK Hynix −11%, Samsung −9.5%
Oracle 5-year CDS	288bp, a record (260bp a week earlier)
CoreWeave 5-year CDS	643bp (Bloomberg) to 1,000+bp (LSEG)
CoreWeave 2032 bonds	~77¢ on the dollar
SoftBank 5-year CDS	424bp
US 10-year Treasury	5.25% after a hot CPI, ~5.10% the next morning
September CPI / PPI	CPI 3.7% y/y; PPI +1.3% m/m vs 0.6% expected

A CDS (credit default swap) is insurance against a company not paying its debt. A higher number in basis points (bp) means lenders see more risk. Amazon, Alphabet, Meta and Nvidia CDS also hit record wides, but stayed far below Oracle's.

The stack: seven layers, safest to riskiest

Layer 1 — Hyperscaler profits · stress: none

Size: $500–600B of the build-out since 2023 came from retained earnings.
Who pays: Microsoft, Alphabet, Amazon and Meta from operating cash flow.
Why it matters: 2026 capex guidance totals about $710–750B, up from about $410B in 2025. All four raised guidance during the year and none has cut.

Layer 2 — Hyperscaler bonds · stress: low

Size: the hyperscalers issued $93–120B of bonds in 2025, against roughly $30B a year in 2020–2024. S&P counts $225B in 2026 so far and projects about $400B for the year.
Lenders: investment-grade bond funds, insurers, pension funds.
Signal: CDS at record wides, but low in absolute terms. These borrowers can pay.

Layer 3 — Off-balance-sheet SPVs and leases · stress: medium

Example: Meta's Hyperion data center in Louisiana. About $27–30B of debt sits in a vehicle owned 80% by Blue Owl and 20% by Meta, with Pimco as anchor lender. The debt is off Meta's balance sheet.
Hidden exposure: Meta carries a residual value guarantee of up to $28B, disclosed only in the footnotes of its annual report.
Lenders: private credit funds, which raise money from insurers, pensions and wealthy retail investors.
Why medium: the structures are opaque, so it's hard to see who is exposed.

Layer 4 — Oracle · stress: high

Size: Oracle planned to raise $45–50B of debt and equity in 2026 and raised $30B within days (bonds plus mandatory convertible preferred). Separate project loans include $38B for two sites in Texas and Wisconsin and $18B for one in New Mexico.
Cash: free cash flow was −$23.7B in fiscal 2026. Oracle says $75B of its AI hardware is prepaid or supplied by customers, which reduces what it has to borrow.
Depends on: OpenAI's $300B, five-year compute contract.
Signal: 5-year CDS at a record 288bp.

Layer 5 — CoreWeave and the neoclouds · stress: highest

Size: $35.6B of scheduled principal payments as of June 30, 2026 (10-Q); about $50B including other obligations (Reuters).
Lenders: high-yield bond buyers and lenders who take GPUs as collateral.
Signal: 2032 bonds near 77¢, CDS between 643bp and 1,000+bp depending on the data source. The company is loss-making.

Layer 6 — OpenAI, the hinge

Raised: a round of about $122B, anchored by Amazon ($50B), Nvidia ($30B, largely as chips and capacity rather than cash) and SoftBank ($30B in three $10B tranches due April 1, July 1 and October 1, 2026).
Owes: roughly $400B+ in Stargate commitments, the $300B Oracle contract, plus CoreWeave and Microsoft.
Earns: about $50B in annualized revenue, with about $27B of cash burn expected in 2026.
Why it's the hinge: Oracle, CoreWeave, Stargate's lenders and SoftBank are all repaid, directly or indirectly, out of OpenAI's revenue. (inference) If that revenue falls short, they are hit at the same time.

Layer 7 — SoftBank · stress: high

Structure: margin loans and junk bonds secured against its own OpenAI stake. Its OpenAI money is partly borrowed against OpenAI itself.
Lenders: banks and high-yield buyers.
Signal: 5-year CDS at 424bp.

How the money goes around

The concern people call "circular financing" works like this:

Step	Flow
1	Nvidia invests ~$30B in OpenAI, mostly as chips and capacity. OpenAI buys Nvidia chips. Nvidia is reportedly weighing financing a separate $350B OpenAI chip purchase.
2	Amazon invests $50B. (inference) Some of it is likely spent on AWS compute.
3	SoftBank invests $30B, partly funded by loans secured on its OpenAI stake.
4	OpenAI signs compute contracts with Oracle, Stargate and CoreWeave.
5	Oracle, Stargate and CoreWeave borrow from bond funds, banks and private credit to build data centers, using those contracts as the reason lenders will lend.
6	Lenders are repaid from OpenAI's payments, which depend on OpenAI's revenue.

Why it matters (editorial): these deals make revenue look bigger than it is in cash. Investor money comes back as sales, and those sales justify more borrowing. It works while the end customer keeps growing. If it doesn't, the same loop runs in reverse, and the stress spreads to several layers at once.

Nvidia's own 10-Q says its August memorandums of understanding with capital providers "may not lead to definitive agreements."

Who is actually on the hook

The end lenders are insurers, pension funds, bond funds and wealthy retail investors, reached through corporate bonds and private credit. Banks hold the project loans.

(editorial) Unlike 2008, most of the risk isn't sitting in leveraged banks. It's spread across private credit, which is harder to see and slower to mark down, so losses tend to show up late and all at once.

The 2001 lesson

Internet traffic kept growing straight through 2001–2003. The debt-funded fiber builders still went bankrupt, and Cisco fell more than 80%.

(editorial) AI compute demand looks safe for now. Microsoft is moving some Office AI work to its own models, and Meta launched its own Muse models on October 13. Those in-house models need their own chips and data centers, so total compute keeps growing. What shifts is who pays for it: more from cash-rich hyperscalers, less from debt-funded middlemen. Demand can grow while the lenders to those middlemen lose money.

What to watch, in the order stress would show

Oracle and CoreWeave CDS and bond prices. The first place stress shows, daily.
SoftBank's October 1 tranche. Whether the last $10B reached OpenAI hasn't been confirmed in public reporting.
OpenAI's revenue against its commitments. Several hundred billion owed against about $50B a year in revenue.
Hyperscaler Q3 earnings in late October. Layers 1–2 hold as long as capex guidance keeps rising. A cut would be the real 2001 signal.
Private credit marks. Blue Owl and BDC discounts to their reported value. Slower, but more telling.
The Fed on October 28. Rate-hike odds jumped after the hot CPI. A hike raises the cost of every layer below Layer 2.

The read (editorial)

Layers 1 and 2 are sound: companies with real profits are spending. The risk is concentrated in Layers 4 through 7, where the debt depends on one customer that is still burning cash. That's why chipmakers like TSMC and the cash-rich hyperscalers can hold up even if Oracle, CoreWeave and SoftBank come under real strain.