On 2026-09-30 Microsoft had $35.5B of debt. On the same day it had also signed $364.6B of data-center leases that had not yet started. The second figure is ten times the first, yet it does not appear on the debt line: under the accounting rules, a lease becomes a liability only when it starts, that is, when the building is handed over for use.
So the question “how much does the compute industry owe?” has no single answer, because owing comes in several kinds. A bond is debt. A lease that has not started is not debt today, though the rent is already promised. A take-or-pay contract, in which a customer pays for reserved capacity whether it uses it or not, is not debt at all: for the buyer it is a commitment, for the seller an order. Yet lenders write it into loan agreements as cash they expect to collect. Inside the same building, one company borrows for forty years on a AAA rating, while another pledges GPUs that lose value and borrows for two or three years at double-digit rates.
Why borrow at all? The five largest buyers expect to spend more than $800B on capital expenditure in 2026, about 2.5% of US GDP (BIS, from company guidance). Their own cash no longer covers that everywhere, so they borrow. By the BIS count, $222B of bonds had been sold by September. Our own tally of priced public deals, run through Microsoft’s 2026-11-04 closing, comes to about $253.3B (calc). It ends later than the BIS count and includes Microsoft’s $25B deal, so the two are shown side by side, not as a cross-check. S&P had expected free operating cash flow, that is, operating cash flow minus capital spending, to be negative at the six largest in both 2026 and 2027.
The piece lays the borrowing out in two chapters. Chapter 1 explains twenty-seven ways the sector raises money, from the plainest company bond to a county bond whose main purpose is tax. For each one it gives how it works, who uses it, a real case and a figure. Chapter 2 turns to the companies: who uses which method, how large each one is, what it costs, and which dates come first.
Data as of 2026-11-11 13:04 (Asia/Shanghai). US dollars.
Seven bases, never added together
Every figure carries one of seven bases, such as debt on the balance sheet, a lease not yet started, or a contract commitment. They measure different things, so they are never added into one “AI debt” number. Where we add anything up ourselves, it is marked calc. Figures that only the press reports are marked media report. A non-GAAP figure always sits next to a GAAP figure. Companies report on different dates (6/30, 8/31, 9/30), and the date is given where it matters.
- Debt Interest-bearing debt, principal or carrying amount
- Lease liability Lease liability already on the balance sheet
- Lease not yet started Signed lease that has not commenced, so it is not on the balance sheet yet
- Commitment Purchase, capacity or undrawn commitment, not owed yet
- Backlog or prepayment Owed to the company or prepaid by customers. Not debt
- Guarantee cap Guarantee or backstop, paid only if someone else fails
- Equity-linked Equity or equity-linked raise. Not debt
How to read the figures
- Thick arrow or solid bar: cash that moves today, or an amount that is firm.
- Thin arrow: payments or services over time.
- Dashed: moves only if something happens, such as a draw, a default or a conversion; or an amount that can be cancelled.
- Outline: a limit, such as a credit line or a program, with nothing drawn against it.
- Hatched: a shortfall that is paid for anyway, by a backstop or under a minimum bill.
Chapter 1 · Twenty-seven ways to borrow
The methods run from the closest to an ordinary loan to the furthest from it, in seven families. Each one gets a one-line definition, how it works, who uses it and a real case with numbers. The figure follows how the method works: a timeline where the timing of cash defines it, bars where size or a limit does, a payoff line where the outcome turns on one price, and elsewhere a rating scale, a stack of layers, a branch, a loop or a map of the parties.
Two notes before the list. First, equity also fills the funding gap but is not debt; it appears here because it sits next to debt in the same financing plans, and it is marked equity-linked. Second, every prospectus in the sample says the money is for “general corporate purposes,” so no bond can be called an AI bond as a matter of fact.
Borrowing on the company’s own credit
The closest thing to an ordinary loan. The lender relies on the whole company and its cash flow, not on one building or one batch of GPUs.
01Investment-grade bonds
A company sells bonds to many investors at once, pays interest on a schedule and repays at maturity.
Each bond pays a coupon, interest set as a percentage of the principal, on fixed dates, and the principal is repaid at maturity. Investment grade means a rating of BBB- or higher, the band rating agencies give borrowers with a low risk of default, so the coupon is only a little above what the US Treasury pays. The money is not tied to any data center: every prospectus in the sample says “general corporate purposes,” which can include capital expenditure. Investors rely on the whole company’s cash flow.
- 2026-11-04
- Each coupon date
- 2028 → 2066
- $25B received
- Coupons: fixed 4.850%–6.350%; two floaters at SOFR + 35 and + 53bp
- Principal, repaid tranche by tranche
↑ Above the line: what the borrower receives↓ Below: what it paysArrows are schematic, not to scale.
Case. All five of the largest buyers sold investment-grade bonds in 2026. The newest deal is Microsoft’s: $25B of senior notes priced on 2026-10-29 and closed on 11-04, its first new dollar investment-grade deal since 2017. It came in ten tranches maturing from 2028 to 2066, with fixed coupons from 4.850% to 6.350%, plus two floating-rate tranches at SOFR + 35bp and + 53bp. SOFR is the US dollar overnight benchmark rate, and a basis point (bp) is a hundredth of a percentage point. The notes are rated Aaa/AAA. Earlier, Alphabet sold $45B in dollars plus $35.8B in other currencies, Amazon $37B in 11 tranches in March, Meta $25B in April (after $30B in October 2025), and Oracle $25B in February.
Sources: S-400 · L88, L190, L193; S-401 · L21; S-143 · L2092; S-133; S-144; S-146
02Ultra-long and century bonds
The same bonds, with maturities of 40, 50 or even 100 years.
Buildings last for decades, and a long bond matches that life. The cost falls on investors: a bond’s price falls when interest rates rise, and the longer the bond, the larger the fall. To carry decades of that risk, investors ask for a higher coupon than on a 10-year bond. For the issuer, the benefit is not having to refinance the principal for decades.
Case. The longest runs a full century: Alphabet sold £1B at 6.125%, maturing in 2126. Amazon has a tranche maturing in 2076. Microsoft’s longest new tranche runs to 2066 at 6.350%, the highest fixed coupon in that deal.
Sources: S-139 · fwp-20260210b L96, L132; S-133 · fwp-20260310 L46; S-400 · L193
03Foreign-currency bonds
Bonds sold in other currencies, to reach investors outside the dollar market.
The company sells bonds in euros, yen, sterling, Swiss francs, or Canadian or Australian dollars. Many investors buy only bonds in their home currency, so these deals reach money the dollar market cannot, and the company can raise more without pushing up the rates on its own dollar bonds. The cost is a second currency to manage: the payments are often swapped back into dollars with a currency swap, a contract that exchanges one currency’s payments for another’s.
- Euro€14.5B
- Canadian dollarC$14B
- Swiss francCHF 2.8B
- Sterling£4.25B
Reaches investors who buy only their own currency, without crowding its dollar curve
- UsuallySwapped back into dollars
The cost: one more currency to manage
Case. Alphabet and Amazon each borrowed in at least four currencies besides the dollar. Alphabet sold ¥576.5B, plus deals in euros, Canadian dollars, Australian dollars and Swiss francs. Amazon sold €14.5B, C$14B, CHF 2.8B and £4.25B.
Sources: S-140 · fwp-20260515 L91–L137; S-143 · L2094; S-133 · fwp-20260311 L168–L275; S-30 · L1039–L1160
04Commercial paper and revolving credit
Short-term borrowing that the company draws and repays as its cash needs change.
Commercial paper is a short IOU, lasting from a few days to a few months, sold mainly to money-market funds. A revolving credit line is a bank’s promise to lend up to a limit, which the company can draw and repay at will. Together they are a liquidity cushion for paying bills between larger financings. The unused part of a line is a commitment, not debt, and the amount outstanding can be zero.
The lineDrawn: only this part is debt
Case. A large program can sit unused. Amazon has a $30B commercial-paper program and Alphabet $25B, and neither had anything outstanding. Oracle, by contrast, repaid $1.1B of commercial paper in a single quarter.
Sources: S-30 · L1163; S-143 · L2086; S-11 · L10478; S-60 · L21353; S-288 · L25
05Term loans and delayed-draw term loans
A bank commits a fixed amount; the borrower draws it at once or in stages as it buys equipment.
A bank or a group of lenders commits a fixed amount, and the borrower repays on a schedule at a floating rate: SOFR plus a margin. A delayed-draw term loan is a promise first and debt only as it is drawn. In GPU deals the amount that can be drawn follows a formula. On the funding date it is at most 80% of the capex being funded. On each monthly test date the balance can be at most 85% × (original capex − contract depreciation, six years straight-line). As the GPUs age, the same chips support a smaller loan.
- Signing
- Each GPU purchase
- Every month
- 2030-08-21
- A commitment, not cash yet: $2.6B for DDTL 3.0
- Draw up to 80% of that spend
- Interest over SOFR; prepay if the balance tops 85% of depreciated cost
- Final maturity
↑ Above the line: what the borrower receives↓ Below: what it paysArrows are schematic, not to scale.
Case. The two examples are priced far apart. Amazon signed a $17.5B unsecured term loan in June 2026, fully drawn by September and due September 2029, at SOFR + 0.625%–0.875% depending on its rating. CoreWeave’s DDTL 3.0 committed $2.6B with final maturity on 2030-08-21. Its newer DDTL 5.5, from August 2026, costs SOFR + 5.50% and requires debt-service coverage of at least 1.35×: the cash available for debt payments must be at least 1.35 times the payments due.
Sources: S-403 · L1077; S-388 · L253, L365, L741; S-152 · L120
06Bridge loans
A short loan that carries the borrower until longer-term money arrives.
A bridge loan lasts a year or less. It pays for a purchase now and is repaid from the bonds, shares or asset sales that follow. It can be arranged quickly. The risk is timing: if the long-term money comes late, or costs more than planned, the bridge has to be refinanced.
- Day 0Borrow short
SoftBank: a $40B bridge loan. Nebius: $3.5B from Citi, due in 364 days.
- Right awaySpend it
SoftBank put the money into OpenAI’s equity round.
- Within the yearLong money arrives
Bonds, new shares or an asset sale.
On timeLateRefinance the bridgeThe long money is late, or dearer than planned.
- EndRepay the bridge
Case. The larger of the two is SoftBank’s: a $40B bridge loan to invest in OpenAI. Nebius borrowed $3.5B from Citi for 364 days, just under a year.
Sources: S-270 · L28; S-158 · L27
07High-yield bonds
Bonds from borrowers rated below investment grade, who pay higher interest.
These are bonds from borrowers rated BB+ or lower, the band below investment grade, where the risk of default is judged higher. Investors are paid for that risk with higher coupons and protect themselves with tighter covenants, conditions the borrower must keep, such as limits on further borrowing. The buyers are high-yield funds and private credit.
- AAAMicrosoft
- AA+Alphabet
- AAAmazon
- AA-Meta
- A+Hyperion notesMeta’s campus joint venture
- A
- A-
- BBB+
- BBB
- BBB-OracleSince 2026-07-09
- Investment grade aboveHigh yield below
- BB+SoftBankDollar bonds at 8.625%–9.750%
- BB
- BB-
Case. AI companies are now a large part of new issuance. The Bank of England says 41% of this year’s new (non-refinancing) US high-yield issuance came from AI companies. SoftBank’s dollar bonds are rated BB+, one notch below investment grade, with coupons of 8.625%–9.750%.
Sources: S-306 · L329; S-275 · L41, L55
08Convertible bonds
A bond the holder can exchange for shares, so the interest is low.
The holder can exchange the bond for a set number of shares, which works out to a fixed price per share, the conversion price. That right is worth something, so the holder accepts a low coupon, sometimes zero. If the share price ends above the conversion price, holders convert, and existing shareholders are diluted: each owns a smaller share of the company. If it does not, the bond stays debt and the principal must be repaid in cash.
Value received per bond
- Conversion price
Share price at maturity →
Case. Most coupons in the sample are low, and one is not. CoreWeave sold $4.2B at 2.875% in September 2026. Cipher sold $850M at 0.00%, due 2032-11-01. IREN has six convertibles totaling $6.746B at 0%–3.50%. Anthropic’s converts are the exception: up to $42B at 9.00%, and the cash can only be used for rent.
Sources: S-51 · ex991-pricing L10; S-410 · L959; S-404 · L2786; S-60 · L22081
Equity raised alongside debt
These are not debt. They fill the same funding gap, and some pay a fixed amount that looks like interest, so they are on the map, marked equity-linked.
09Mandatory convertible preferred
Preferred shares that pay a fixed dividend and must turn into common shares on a set date.
Preferred shares rank ahead of common shares for dividends. These pay a fixed dividend for a few years, so for that time they behave much like a bond; on the set date they convert into common shares, and nothing is repaid in cash. Unlike a convertible bond, there is no choice at the end: conversion is mandatory. Rating agencies give them partial equity credit, counting part of the money as equity rather than debt, which protects the issuer’s rating while the cash comes in.
- 2026-10-29
- Each dividend date
- Set date
- Meta receives $12.5B
- Fixed 6.50% dividend
- Converts into common shares; nothing is repaid in cash
↑ Above the line: what the borrower receives↓ Below: what it paysArrows are schematic, not to scale.
Case. Three of the five largest buyers have used them. Meta sold $12.5B at 6.50%, closing on 10-29. Oracle sold $5B at 6.50%. Alphabet raised $19.0B net.
Sources: S-145 · 8k-20261029 L120; S-147 · 424b5-final-20260204 L21; S-143 · L2688–L2694
10At-the-market offerings and forward sales
Selling new shares bit by bit, or at a price fixed now for delivery later.
Under an at-the-market (ATM) program, the company sells new shares through brokers at the market price, in amounts and at times it chooses, up to a program limit. In a forward sale, it agrees a price now and delivers the shares, and receives the cash, later. Nothing is owed. The cost is dilution: existing holders own a smaller share of the company. An unused program limit is room to sell, not money in hand.
Sold: cash receivedUnused or only planned: not cashAgreed: cash comes at settlement
Case. The programs are large, and only part of the money has arrived. Oracle used a $20B program within one quarter. Alphabet has a $40B program, mainly to cover taxes on employee stock; $12.4B was sold and $27.6B remained. Meta agreed a forward sale of 40 million shares to settle by 2027-11-19, and plans about $60B of at-the-market sales; that agreement is not yet signed.
Sources: S-11 · L5555; S-141 · fwp-20260601 L51, L131; S-143 · L2714; S-145 · 8k-20261029 L140; S-04c · L44
11Warrants as payment
The right to buy shares at a fixed, often tiny price, given in exchange for a guarantee, an order or a tenant.
A warrant lets its holder buy shares at a fixed price. When that price is tiny, the warrant is worth almost as much as the shares themselves. Companies hand warrants to a party whose help they need: a stronger company that stands behind them, a customer placing a large order, or a tenant. It costs no cash, but it dilutes existing shareholders. The direction can run either way.
| Gives the warrants | Gets them | Terms | In return |
|---|---|---|---|
| TeraWulf | 73.58M shares at $0.01 | A backstop for its tenant’s rent | |
| AMD | OpenAI | Up to 160M shares at $0.01, vesting as GPUs are deployed; none vested as of 2026-06-27 | A large GPU order |
| Applied Digital | CoreWeave | — | A tenant for its halls |
Case. The warrants go to whichever side the issuer needs: a backer, a large customer or a tenant. Google backstops leases for TeraWulf’s tenant and received warrants for 73.58 million TeraWulf shares at $0.01 each. AMD granted OpenAI warrants for up to 160 million shares. In the other direction, Applied Digital granted warrants to its tenant, CoreWeave.
Sources: S-170 · L1625; S-256 · L123; S-165 · L1877
Rent instead of buy
The largest promises in the sector are leases. A lease that has started is on the balance sheet; many more are signed and have not started yet.
12Finance and operating leases
Renting buildings or machines on a long contract instead of buying them.
Instead of buying a building or machines, the company signs a long lease. Once the lease starts, it records a lease liability, the present value of the rent still to be paid. Leases that work like a purchase, for example by covering most of the asset’s life, are finance leases: part of each payment is booked as interest, as on a loan. The others are operating leases, with a single rent expense. Accounting calls both leases; economically, a long finance lease is close to buying the asset with the landlord’s money.
Finance-lease interest: $769MOther interest: $101M (calc)
Case. At Microsoft, most of the interest now comes from leases. Its CFO: “One financing method is leasing, which we’ve always used.” Of the quarter’s $870M of interest, $769M was lease interest, about 88% (calc). Finance lease liabilities were $70.973B at 2026-09-30.
Sources: S-103 · L534; S-100 · L841, L1815, L1923
13Leases not yet started
Leases signed for buildings not yet handed over; they are not on the balance sheet yet.
The lease is signed but the hall has not been handed over. Under lease accounting, nothing goes on the balance sheet until the lease starts; on that day the whole liability arrives in one step. Until then only a footnote discloses the total rent promised. It is not today’s debt, but the payments are already committed.
Amount shown in the accounts
- Lease starts
Time →
Case. Each of the five largest buyers has more than $130B of them. Microsoft $364.6B, Meta about $355.73B, Oracle $288B (almost all data centers, 15–19 years), Amazon $164.922B (up from $137.214B in June), Alphabet $130.5B. Together about $1.304T (calc; dates differ, 8/31 to 9/30). Moody’s, via the BIS: of $1.2T in lease commitments, about $820B had not started.
Sources: S-100 · L1989; S-01 · L1262; S-11 · L5460; S-403 · L891; S-105 · L1537; S-300 · p3 L76; F-407
14GPU sale-leasebacks and GPU finance leases
A fund buys the GPUs and leases them to the company that uses them.
A fund buys the GPUs, either new or from the company itself, and leases them back. The machines stay in the company’s halls and keep running; the fund holds title, collects rent, and often borrows to pay for the purchase. The company gets the chips without paying for them up front. When the lease runs for most of a chip’s life, accounting treats the deal as a financing, not a rental.
The GPUs stay in its halls
- Title to the GPUs
- $5.4B purchase price
- Rent over the term
A fund; title to the GPUs moves into its name
$3.5B of debt toward the price
The buyer borrows too
Case. In each example the chips go to an AI lab. Valor Compute Infrastructure spent $5.4B buying GPUs to lease to xAI, with $3.5B of debt led by Apollo. SpaceX, which now includes xAI, signed a GPU finance lease of about $40B; it has not closed. Anthropic leases TPU equipment from Broadcom with about $161.2B of rent over five years.
Sources: S-204 · L46, L48; S-267 · L112; S-60 · L22955
Borrowing against GPUs and customer contracts
When the company’s own credit is not strong enough, the lender looks at the GPUs themselves and at who has promised to pay for their use.
15GPU-backed loans
Loans secured on GPUs and the customer contracts they serve.
The GPUs and the customer contracts they serve are the collateral, usually held in a separate subsidiary. The lender sizes the loan to the chips’ depreciated value and to the cash the contracts bring in, so the loan shrinks as the chips age. It adds protections: a cash reserve, a test that customer payments are actually collected, and a coverage ratio. It also takes a guarantee from the parent company and a pledge of the subsidiary’s shares.
Years of straight-line depreciation
Case. GPU loans include the most expensive borrowing in the sample. CoreWeave’s borrowing capacity follows the depreciated cost of its GPUs. IREN borrowed $2.4B from Blue Owl at 9%, repaid in installments over 30 months. The most expensive GPU loan in the sample carries an effective rate of 15%.
Sources: S-150 · L4586; S-162 · L24; S-404 · L2080; S-50 · L1719
16Vendor financing, customer prepayments, bring-your-own chips
Getting hardware without paying for it up front: the vendor finances it, the customer prepays, or the customer brings its own chips.
There are three ways. The vendor, or a lender it arranges, finances the order. The customer pays in advance, so its cash funds the build. Or the customer supplies its own GPUs, and the provider supplies the building, power and operations. Amazon’s newest version is a secured facility it can draw against the technology infrastructure it buys; each draw matures five years later at a floating rate. Until it is drawn, it is a commitment, not debt.
- The vendor’s side lendsAmazon
A secured facility of up to about $80B: draws from 2027 Q1 to 2028 Q3, five years each, floating rate, backed by the equipment bought.
- The customer pays aheadOracle, IREN
Oracle received $11.4B of prepayments in Q1 FY27. IREN’s recent prepayments equal 45%–55% of GPU capex (company figure).
- The customer brings the chipsOracle
In its large contracts, prepayments and customer-supplied GPUs come to $75B together.
Case. In some builds, customers’ money already covers a large share. Oracle: prepayments and customer-supplied GPUs in its large contracts total $75B, and it received $11.4B of prepayments in Q1 FY27. IREN: recent customer prepayments equal 45%–55% of GPU capex (company figure). Amazon: a loan facility of up to about $80B, drawable from Q1 2027 to Q3 2028, secured on purchased technology infrastructure (undrawn).
Sources: S-13 · L114; S-11 · L10478; S-161 · L3; S-403 · L1079
17Take-or-pay contracts
The customer reserves capacity and pays for it whether it uses it or not.
For the buyer, a take-or-pay contract is a commitment; for the provider, it is backlog, revenue already signed. Lenders write the contracted cash into the loan agreement and test whether it actually arrives, a measure called the realization rate. The weak point is the exit: if the customer can end the contract because the capacity was not available as promised, the cash the lender counted on disappears.
Owed whether used or notA cap that can be cancelled
Case. Lenders treat these contracts as something to lend against: CoreWeave borrows against them. Microsoft’s contract with Nebius is worth up to about $17.393B. The largest in the sample are Anthropic’s: its prospectus shows minimum commitments of at least $114.5B to Google, at least $110B to Amazon and at least $44.9B to Microsoft.
Sources: S-150 · L382; S-159 · L11485; S-60 · L22831–L22865
Borrowing through a separate project company
Here the campus borrows, not the company. The debt sits in a separate company that owns the campus; the tech company signs the lease whose rent repays it, and the debt stays off its own balance sheet.
18Joint ventures and special-purpose vehicles
A campus built in a company the tech company only partly owns, so its debt stays off the tech company’s balance sheet.
The campus is built inside a separate company, a joint venture or special-purpose vehicle, co-owned with a fund. If the tech company holds a minority stake and does not control the vehicle, it does not consolidate it, that is, it does not add the vehicle’s accounts to its own, so the vehicle’s debt stays off its balance sheet. The tech company then signs the long lease whose rent repays that debt. Either way the lenders rely on the same tenant; what changes is the accounting, and that has a price.
On Meta’s balance sheet
Signs the lease; does not consolidate the JV
- Rent, up to 20 years from 2029
- Residual value guarantee
- The campus
Off Meta’s balance sheet
Blue Owl funds 80%, Meta 20%. Development cost about $27B.
$27.294B at 6.581%, due 2049
S&P rates the notes A+, a notch below Meta
Case. Hyperion shows the price. Meta and Blue Owl own the campus through a joint venture: the fund holds 80% and Meta 20%, with total development cost of about $27B. The project company sold $27.294B of senior secured notes at 6.581%, due 2049, partly to PIMCO. S&P rated them A+, a notch below Meta. They priced 225bp over Treasuries; two weeks later Meta’s own 20- and 30-year bonds priced at 88 and 98bp. Same tenant, about 1.3–1.4 points dearer through the vehicle (calc; maturities not identical). That gap is, roughly, the cost of keeping the debt off Meta’s balance sheet.
Sources: S-03 · L34, L48; S-200 · L3006; S-206 · L21; S-300 · p6 L259
19Residual value guarantees
The tenant promises the campus will be worth an agreed amount when the lease ends, and pays the gap if it is not.
If the tenant leaves when the lease ends and the campus is worth less than the agreed value, the tenant pays the difference, up to a cap. That puts a floor under the building’s value for its owners and their lenders. While the company judges a payout “not probable,” it records no liability; the footnote shows only the most it could owe.
What Meta pays
- Agreed value
Campus value when the lease ends →
Case. The amount at stake is large and off the balance sheet. Meta’s residual value guarantees on its two joint ventures, Hyperion and El Paso, are capped at $41.493B. Meta judged a payout not probable and recorded no liability.
Sources: S-01 · L1050, L1066
20Construction and syndicated loans
A loan paid out in stages as the building goes up.
The money is released in stages as construction progresses, and the borrower often pays only interest until the building is finished; the principal comes due after. A few banks arrange the loan and then sell pieces of it to other lenders, including loan funds and private credit. That selling-on is called syndication.
- Signing
- Construction
- 2028
- Up to $2.288B committed
- Drawn as the work is done
- Interest only, SOFR + 2.5%
- Principal due
↑ Above the line: what the borrower receives↓ Below: what it paysArrows are schematic, not to scale.
Case. Both loans carry about the same margin, 2.5 points over the benchmark. Abilene’s first two buildings: up to $2.288B, interest-only, SOFR + 2.5%, due 2028; the tenant is Oracle on a 15-year lease. For a campus in Shackelford, Texas, banks committed $23B at about the benchmark + 2.5 points and then sold pieces to loan investors and private-credit funds (media report via Bloomberg; no primary document).
Sources: S-208 · L347, L387; S-214 · L41, L51
21Project bonds
Bonds repaid only from one campus’s rent and assets.
The bonds are issued by the company that owns a single campus and are repaid only from that campus’s lease and assets. If the money runs short, bondholders cannot turn to the parent. So the price depends on the tenant whose rent repays them: the same kind of bond costs about 6% with Google standing behind the rent, and nearly 10% when the tenant is a neocloud, a newer GPU cloud provider with weaker credit.
- River Bend (Hut 8)Google backs the rent6.129–6.192%
- HyperionTenant: Meta6.581%
- Polaris Forge 1 (Applied Digital)Tenants include CoreWeave; three tranches6.75–9.25%
- Helios (Galaxy)Tenant: CoreWeave9.875%
Case. The four examples run from 6.129% to 9.875%, roughly in line with the tenant’s credit. Hyperion 6.581%. Hut 8’s River Bend, with Google backing the tenant’s rent: 6.129%–6.192%. Galaxy’s Helios, leased to CoreWeave: 9.875%. Applied Digital’s three tranches: 6.75%–9.25%, with the parent guaranteeing only that the building gets finished.
Sources: S-200 · L3006; S-183 · L10255, L10275; S-180 · L106; S-165 · L1385–L1410, L1530
22Asset-backed securities (ABS and CMBS)
Many leases or loans pooled and sold as bonds paid from their combined rent.
The leases or loans go into a pool, which issues bonds in slices, or tranches, of different seniority: the senior slices are paid first from the pooled rent, and the junior slices take losses first. CMBS are the version backed by loans on commercial buildings. Most of this money builds shells and power, not the servers inside them. No public prospectus for a GPU or compute ABS has been found.
Cash, top down
- SeniorPaid first
- MezzaninePaid next
- JuniorPaid last
First to lose if the rent falls short
Losses, bottom up
Case. The market is growing quickly, and one rating agency has already marked a building down. US data-center ABS and CMBS issuance: $9.4B in 2021, about $26B in 2025 (preliminary). Sell-side forecasts are $30–40B a year for 2026–27. KBRA valued the building behind one $1.26B single-tenant CMBS 27.2% below the appraisal; on its value, the loan was 108.4% of the building, more than the building was worth.
Sources: S-226 · p12 L509, L517, p4 L115; S-229 · L29; S-306 · L339
23Industrial revenue bonds
Bonds issued in a county’s name for a private project, mainly for the tax treatment; the county does not repay them.
A county issues bonds in its own name for a private project. They are not county debt, and the county does not repay them. Here the buyers are the developer’s own affiliates, so the money leaves the developer group and comes back to it. Instead of property tax, the project pays the county an annual amount called PILOT, short for payment in lieu of taxes. Our reading: the point is tax, not financing. It is the largest number on the map and not new borrowing risk, so it is never compared with other debt.
Signed for series 2025A
Not liable for repayment; receives PILOT in place of property tax
Builds the campus and pays the bonds
Case. The documents themselves call the deal “self-funded”. Doña Ana County, New Mexico, “Project Jupiter”: a cap of $165B in three series ($15B 2025A, $25B 2025B, $125B 2025C for tenant equipment). Series 2025A is signed; the buyer is a developer affiliate.
Sources: S-218 · p5 L170, p6 L212, p7 L236, p8 L284–L285; S-217 · p129 L40
A stronger party stands behind it
A stronger company promises to buy, pay or take over if a weaker one cannot. Nothing moves unless something goes wrong.
24Capacity backstops
A large company promises to buy whatever capacity the provider cannot sell to others.
The backer agrees to buy any capacity the provider cannot sell elsewhere. That turns uncertain demand into a contract the provider can show its lenders. For the backer it is a commitment that costs money only if demand falls short.
Sold to customersBought by the backstop, up to its order
Case. The backers are companies with strong credit, here Nvidia and Meta. Nvidia has a $6.3B order with CoreWeave running to 2032. Meta’s agreement with Nebius is worth up to $15B.
Sources: S-151 · L102; S-159 · L5520
25Payment backstops and guarantees
A stronger company agrees to pay, or step in, if a tenant cannot pay its rent.
When the tenant’s credit is not strong enough for lenders, a bigger company signs that it will pay, or take over the lease, if the tenant fails. The guarantor does not book this as its own debt, though some such agreements are booked as credit derivatives. Economically it is close to a guarantee, and it is what lets the landlord borrow against the lease.
- LeaseAnthropic rents about 401 MW for 20 years
At TeraWulf’s Kentucky campus (Raylan)
- 2026-11-04Google signs credit support
While the rent is paid, nothing moves
- Option 1Google pays the rent
and takes over the lease
- Option 2Google pays a termination amount
and takes equity in the project company
Case. The two largest in the sample are Alphabet’s and Nvidia’s. Alphabet has $75.7B of data-center payment backstops. Nvidia guarantees OpenAI’s data-center leases, capped at $105B. The newest shows how they work. On 2026-11-04 Google signed credit-support agreements for Anthropic’s 20-year, roughly 401 MW lease at TeraWulf’s Kentucky campus: if Anthropic misses rent or goes bankrupt, Google can pay and take over the lease, or pay a termination amount and take equity in the project company. TeraWulf’s press release puts the signed lease commitment at about $19B; the 8-K does not state that total.
Sources: S-105 · L3028; S-254 · L75, L79; S-408 · L73–L79; S-409 · L117
26Utility minimum bills and collateral
Utilities now require data centers to pay for a minimum share of the power they reserve, used or not.
A utility builds grid capacity for the demand a data center reserves. To be sure that cost is recovered, it bills a minimum share of the reserved demand whether it is used or not, and asks customers with weak credit ratings to post collateral. It is another take-or-pay commitment, written by the grid.
Used and billedBilled but not used
Case. AEP Ohio sets the floor close to the full reservation: it bills data centers for at least 85% of contracted demand. Customers rated below A-/A3 must post collateral.
Sources: S-341 · L84, L90
Who does the lending
Every method above needs someone to put up the money. More and more of it comes from private credit funds, whose money in turn comes from insurers, pension funds and individual savers.
27Private credit and BDCs
Funds that lend directly, outside the banking system.
Private credit funds lend directly to companies and projects. They fund GPU loans, project debt and the pieces of loans that banks sell on. Their money comes from insurers, pension funds and, through business development companies (BDCs), individual savers. A BDC is a fund that lends to companies and is sold to individuals; some let investors ask for their money back every quarter, while the loans they hold run for years.
- Insurers
- Pensions
- Individuals, through BDCs
Can ask for their money back
9% of AI investment funding in 2024, 34% in 2025
- GPU loansBlue Owl to IREN
$2.4B at 9%
- Fund debtApollo-led, to Valor
$3.5B toward GPUs leased to xAI
- Pieces of bank loansBought from the arranging banks
Construction and project loans
Case. Private credit’s share has grown quickly, and the strain of that mismatch is already visible. The Bank of England, citing the OECD: private credit’s share of AI investment funding rose from 9% in 2024 to 34% in 2025. Blue Owl lent IREN $2.4B at 9%; Apollo led $3.5B for the fund leasing GPUs to xAI. One BDC received redemption requests for 21.9% of its shares in a single quarter.
Sources: S-306 · L329; S-162 · L24; S-204 · L46; S-300 · p6 L240; S-228
Chapter 2 · The map, company by company
Chapter 1 sorted the borrowing by instrument. This chapter sorts it by borrower, because the same instrument can carry a very different price and risk depending on who stands behind it. It starts with a ladder of rates, then walks down it: the hyperscalers, Oracle, the projects they keep off their balance sheets, the neoclouds, the miners turned hosts, the model labs and the chip vendors. It ends with who holds the debt and which dates arrive first.
In the board below, pick a company type. Each card is one company, with one row per instrument: the bar shows its size and the mark its basis. Hover a row for the rate, maturity and source.
The same data, with interest, payment pressure, a timeline and the money network, lives in Space: Debt Map · Debt Network
The rate ladder
The price of borrowing depends less on what the money buys than on who repays it. The same kinds of data centers and chips are financed at under 5% and at 15%. Line up representative deals from the safest borrower to the riskiest and three patterns show.
First, the same company pays different prices for different structures. Meta pays less than a point over Treasuries on its own 20- and 30-year bonds, but 2.25 points when a fund borrows for it and Meta rents the campus. Second, moving assets under a stronger parent lowers the price. xAI’s 12.5% secured notes were repaid after its merger into SpaceX, and SpaceX then sold $25B of unsecured notes at a weighted 5.855%. Third, project bonds are priced off the tenant: about 6.1%–6.2% with Google behind the rent, 9.875% for a building leased to CoreWeave. At the bottom of the ladder is a GPU loan that costs 15% effective.
- MicrosoftAAA · 2026 notes, 10 tranches4.85–6.35%
- AlphabetAa2/AA+ · 2026-08, $25B, 10 tranches4.5–6.5%
- AmazonA1/AA/AA- · 2026-07, $25B, 8 tranches4.6–6.25%
- MetaAa3/AA- · 2026-04, $25B, 6 tranches4.55–6.45%
- SpaceX (incl. xAI)2026-06, $25B unsecured5.855%
- Hut 8 project (Google backstop)2042 bonds6.129–6.192%
- HyperionA+ · 2049 notes6.581%
- OracleBBB- · 2026-02, $25B, 8 tranches4.55–6.85%
- Applied Digital project3 tranches, 2030–20316.75–9.25%
- SoftBankBB+ · USD bonds8.625–9.75%
- IREN (Blue Owl loan)$2.4B, 30 months9%
- Anthropic → Broadcom partnersConverts up to $42B9%
- Galaxy Helios$3.507B, 20319.875%
- xAI (2025, repaid)$3B secured notes12.5%
- CoreWeave DDTL 1.0GPU loan, effective rate15%
Hyperscalers: three books, side by side
The five largest buyers commit future cash in three ways, which we call three books: debt on the balance sheet; leases, both started and signed but not yet started; and purchase and service contracts. The books are different bases, so the table is read across a row, comparing one company’s books, never down a column into a total.
Read that way, the five look quite different. Microsoft borrows least and leases most. Alphabet’s long-term debt grew 2.69× in nine months, and it gives the largest guarantees. Amazon’s long-term debt grew 2.64×, and it has fully drawn a $17.5B unsecured term loan. Meta put its largest campuses into joint ventures, so their debt is not on its balance sheet.
These companies used to pay for the build out of operating cash. That no longer holds everywhere. In nine months Alphabet spent $131.953B on property and equipment against $127.149B of operating cash flow. Only Microsoft still covers it: $58B of operating cash flow in a quarter against $39.4B of cash capex. As borrowing grows, interest shows up on the income statement: Alphabet’s nine-month interest expense is 6.3 times a year earlier.
| Company | Debt | Lease liabilities | Leases not yet started | Commitments | Guarantees given | Backlog (not debt) |
|---|---|---|---|---|---|---|
| Microsoft | $35.481B2026-09-30 | $70.973B2026-09-30 | $364.6B2026-09-30 | $194.06B2026-06-30 | — | $683B2026-09-30 |
| Amazon | $173.125B2026-09-30 | — | $164.922B2026-09-30 | up to $80B | — | — |
| Alphabet | $125.102B2026-09-30 | — | $130.5B2026-09-30 | $924.4B | $75.7B | — |
| Meta | $80.922B | — | about $355.73B | $454.52B | $68.404B | — |
| Oracle | $125.337B | — | $288B2026-08-31 | — | — | $664B |
- Alphabet9M 2026 · $2.778B×6.3
- Meta9M 2026 · $2.322B×3.3
- CoreWeaveQ3 2026 · $846M×2.7
- Amazon9M 2026 · $3.805B×2.4
- OracleFY27 Q1 · $1.428B×1.5
- MicrosoftFY27 Q1 · $870M—
A year earlierLatest
Oracle: the lowest-rated of the big buyers
What matters at Oracle is less how much it has borrowed than how the two sides line up in time. On one side are customer contracts: $664B of remaining performance obligations, revenue signed but not yet earned, of which only about 13% is due within twelve months. On the other side are its own leases: $288B not yet started, almost all data centers, on 15–19-year terms. The contracts turn into revenue slowly; the rent is fixed for fifteen to nineteen years. Its own 10-K says the long leases and the customer contracts may not match.
The rating agencies already treat those leases as debt. On 2026-07-09 S&P cut Oracle to BBB-, the last investment-grade notch, and counted the unstarted leases as adjusted debt. S&P also estimates that OpenAI accounted for about half of the $638B of obligations Oracle reported at the time; Oracle did not name it.
The financing has also run ahead of the plan. Oracle’s 2026 plan called for one bond deal. It then sold $25B of bonds in eight tranches and $5B of 6.50% mandatory convertible preferred, and used up a $20B at-the-market program within a quarter. At its 2026-10-28 investor day it capped new financing for FY27–FY30 at $20B and said it would lean on vendor financing, customer-supplied hardware and prepayments instead, the three routes described in method 16.
Off the balance sheet: project companies
The biggest buyers have one more way to keep debt off their balance sheets. They put a campus into a separate project company, keep a minority stake, sign a long lease, and let the rent repay the project company’s debt. Hyperion is the template. A fund owns 80% and Meta 20%; the project company borrowed $27.294B at 6.581%, and Meta’s rent services it.
The structure has a price. By our calculation, and with maturities that are not identical, the debt cost about 1.3–1.4 points more than Meta’s own bonds. The other side of it is that the $27.294B is not on Meta’s balance sheet. The pattern repeats with BlackRock in El Paso, with Oracle as the tenant in Abilene, and with xAI’s GPUs at Valor.
One line in the table needs a caveat. Project Jupiter’s $165B cap is the largest number in the group and the least comparable with the rest. The developer’s own affiliate buys the bonds; in our judgment the purpose is mainly tax, not new borrowing risk.
| Vehicle | Instrument | Amount | Rate | Due |
|---|---|---|---|---|
| Hyperion (Meta JV) | Beignet Investor senior secured notes | $27.294B | 6.581% | 2049-05-30 |
| El Paso (Meta JV) | Debt financing | $12.5B | — | — |
| Abilene (Stargate) | Construction loan, first two buildings | up to $2.288B | SOFR+2.5% | 2028 |
| Shackelford, Texas | Bank commitmentsmedia report | $23B | benchmark +~2.5pp | — |
| Valor Compute Infrastructure | Apollo-led debt | $3.5B | — | — |
| Project Jupiter IRB | Industrial revenue bondsnot comparable | cap $165B | — | — |
Neoclouds: sign the contract, then borrow against it
A neocloud signs first and borrows second. It signs a multi-year compute contract with a large customer, borrows against that contract to buy GPUs, and repays the loan from what the customer pays. So its lenders watch two things: what the GPUs are still worth after depreciation, and whether the contract cash actually arrives.
CoreWeave’s DDTL 3.0 is the one contract where the full formula is visible. On the funding date it can borrow up to 80% of the GPU spend. From three months after the commitment period ends, the balance is tested each month and may not exceed 85% of the GPUs’ cost after straight-line depreciation over six years; method 15 charts how the limit falls as the chips age. Two more tests sit on top: realization, at least 0.85× the forecast contract cash; and coverage, at least 1.40×.
The strain is already visible. In the first half CoreWeave paid $6.2B of debt service against $3.7B of operating cash flow. Its carrying debt was $46.873B at 9/30 (calc). That figure is not comparable with the $35.551B of principal at 6/30, so the two are not shown as growth.
The test dates have arrived too. The realization test was due to start on 2026-02-28, and the window in which shareholders could cure a failed test with new equity, as often as needed, closed on 2026-10-28. The company has not said whether it tested or cured. That is not a default; it means we do not know yet. The first coverage test is on 2027-10-31.
- CipherQ2 2026269%Interest $67M ÷ revenue $25M. Some interest is capitalized, so cash interest is higher. Q2 revenue was all mining.
- CoreWeaveH1 2026168%Debt service $6.2B ÷ operating cash flow $3.7B. Principal ~$5.2B plus interest $982M.
- TeraWulfQ2 2026126%Interest $56M ÷ revenue $45M. Some interest is capitalized, so cash interest is higher.
- CoreWeaveQ3 202643%Interest $846M ÷ operating cash flow $1.975B. Deferred revenue rose $2.127B in the quarter, more than the quarter's whole operating cash flow.
- Alphabet9M 20262%Interest $2.778B ÷ operating cash flow $127.149B. PP&E purchases of $131.953B exceeded operating cash flow.
Miners turned hosts: interest larger than revenue
Former bitcoin miners now build halls for AI tenants. They borrow to build first and collect rent only once the leases start, so for a while interest runs ahead of revenue. In Q2 Cipher’s interest was 2.7 times its revenue, and TeraWulf’s interest exceeded its revenue.
Even that understates it. Both capitalize part of their interest, adding it to the cost of the buildings under construction instead of the income statement, so the cash they pay in interest is higher still. IREN’s carrying debt went from $7.593B to $9.897B in one quarter.
Their project bonds are priced off the tenant. Hut 8’s River Bend, with Google behind the rent, pays about 6.1%; Galaxy’s Helios, leased to CoreWeave, pays 9.875%. TeraWulf’s press release cites about $19B of signed lease commitments with Anthropic; the 8-K does not state that total.
Model labs: borrow little, promise a lot
Model labs borrow little themselves and promise a lot. To see their obligations, read the contracts more than the loans: compute contracts, leases, and the money some shareholders borrowed to invest in them.
OpenAI’s only disclosed bank facility is a $4B revolver, undrawn at signing. What it has promised to pay for compute runs to hundreds of billions: over $300B with Oracle over five years (OpenAI’s figure; Oracle has not named it), $250B more of Azure, $38B with AWS plus another $100B over eight years, up to about $22.4B with CoreWeave and over $20B with Cerebras. These are different kinds of commitment and are not added; a total near $1.4T over eight years exists only as a media report.
Some of the money is borrowed by shareholders. In the $110B equity round announced on 2026-02-27, SoftBank took a $40B bridge loan to fund its part and sold BB+ dollar bonds at 8.625%–9.750%.
Anthropic’s prospectus lays a lab’s books open, and they read most clearly basis by basis. Cash: $112.4B at 6/30. Borrowing: a $15B secured revolver, and up to $42B of 9% convertibles whose cash can only pay rent. Minimum purchase commitments: at least $114.5B to Google, $110B to Amazon and $44.9B to Microsoft, plus up to $120.5B with SpaceXAI that can be cancelled on 90 days’ notice. Leases not yet started: $321.4B. Separately, it leases TPU equipment from Broadcom for about $161.2B.
SpaceX, which absorbed xAI, has $38.006B of debt. It has also signed, but not closed, a GPU finance lease of about $40B from an Apollo-managed entity.
Chip vendors as financiers, and money that goes in circles
Chip vendors do more than sell chips. They also finance their customers, each in its own way. Nvidia has only $3.5B of its own notes but $108.5B of guarantees. AMD pays in warrants: up to 160 million shares to OpenAI, none vested as of 2026-06-27. Broadcom backstops the residual value for the funders of its own TPU leases.
Follow the money and some of it goes in a circle: equity goes in, compute contracts come back, and chips get bought. The IMF counted 14 public deals of this kind since 2023: about $300B of equity and $650B of debt in exchange for $1.4T of promised future revenue.
Circular is not the same as fake. Microsoft really did book $10B of revenue from OpenAI in one quarter, and Anthropic’s $16.3B of first-half revenue is real. Two things deserve attention. First, money invested comes back as revenue, so both sides’ numbers grow. Second, a chip vendor that sells and guarantees at the same time moves its customer’s credit risk onto itself. The loops below list disclosed signatories, not proven net round trips of cash.
- Amazon ⇄ OpenAI
$50B of equity goes in; $38B plus $100B of AWS commitments come back.
See it in the network → - Microsoft ⇄ OpenAI
$12B funded; $10B of revenue in a single quarter, plus $250B more Azure.
See it in the network → - Nvidia → OpenAI → clouds
$30B of equity and up to $105B of guarantees; OpenAI signs cloud contracts; the clouds buy chips. The chip-purchase leg is not disclosed.
See it in the network → - Nvidia ⇄ CoreWeave
$2B of equity plus a $6.3B capacity backstop; CoreWeave buys GPUs. The purchase amount is not disclosed.
See it in the network → - Google / Broadcom ⇄ Anthropic
Google is shareholder, lender, guarantor and biggest cloud. Broadcom's partners lend up to $42B that can only pay Broadcom's TPU rent.
See it in the network → - AMD → OpenAI
Warrants only so far. None vested as of 2026-06-27.
See it in the network →
Who ends up holding the debt
Follow the debt down the chain and there are roughly four layers of holders. Banks arrange construction loans and then sell pieces on. Bonds go to insurers, pensions and funds; IBM’s 401(k) plan held about $34.15M of Hyperion notes at the end of 2025, and a non-traded perpetual REIT indirectly owns 63% of Hyperion’s borrower.
Private credit is the third layer, and its share has grown quickly: from 9% to 34% of AI investment funding in a year. Part of its money comes from individuals, the fourth layer, and individuals can ask for their money back. Blue Owl’s OCIC received redemption requests for 21.9% of its shares; Apollo’s ADS met only about a third of the requests it received.
The dates that come first
The first alarm is rarely the final maturity. It comes earlier: a covenant test, the end of a cure window, the day a lease starts, or a peak in principal payments. The table lists the items in the dataset that have firm dates.
| Date | Company | What | Amount |
|---|---|---|---|
| 2026-02-28 | CoreWeave | Realization-rate test starts (≥0.85×). Result not disclosed | — |
| 2026-03-01 – 2028-08-31 | CoreWeave | OEM financing maturities | — |
| 2026-10-28 | CoreWeave | Unlimited equity-cure window closes. Result not disclosed | — |
| 2026-11-04 | Microsoft | $25B notes close | $25B |
| 2026-11-04 | Alphabet | Backstop for the Anthropic lease at TeraWulf signed | — |
| 2026-11-11 – 2026-12-31 | IREN | $1.5B of GPU financing open to draw | $1.5B |
| 2027-01-01 – 2028-09-30 | Amazon | Supplier facility draw window | $80B |
| 2027–2029 | Oracle | Lease commencement window | $288B |
| 2027-07-01 – 2027-12-31 | IREN | Dell deliveries begin | $6.7B |
| 2027-10-01 – 2028-06-30 | TeraWulf | Justified campus delivery | — |
| 2027-10-31 | CoreWeave | First DSCR test (≥1.40×) | — |
| 2027-11-19 | Meta | Forward sale settles by | $24.9B |
| 2028 | Meta | El Paso lease commences | — |
| 2028 | Abilene (Stargate) | Construction loan due | $2.288B |
| 2028 | Nvidia | 4.250% notes due | $3.5B |
| 2029 | Meta | Hyperion lease commences | — |
| 2029-09-30 | Amazon | Term loan matures | $17.5B |
| 2029-12-01 – 2033-12-01 | IREN | Six converts mature | $6.746B |
| 2030–2031 | Applied Digital | Project bonds mature | $6.09B |
| 2030 | Cipher | 7.125% secured notes due | $1.733B |
| 2030-08-21 | CoreWeave | DDTL 3.0 matures | $2.6B |
What to watch, neither predicted nor dismissed
In the eleven days before the data date, nothing in the structure loosened; it gained a few more links. Companies at the top of the ladder keep borrowing long at AAA prices. Those in the middle keep signing 20-year leases with a stronger name behind them. Those at the bottom keep buying undelivered chips with convertibles and 9% GPU loans. The BIS has written that good booms and bad booms are hard to tell apart at the time. The signals below help show whether the strain is building or being absorbed. They are not a reason to buy or sell anything.
- CoreWeave’s 10-Q: realization rate, coverage, any equity cure, whether the monthly GPU test starts to bind, and principal against carrying value and cash interest.
- The next count of leases not yet started at the five largest buyers, and how fast Oracle’s commence.
- Amazon’s supplier facility of up to about $80B: how much is drawn from 2027, against which assets, at what rate.
- IREN: whether later Horizon phases for Microsoft deliver on the revised design, how much of its $1.5B GPU facility is used by 2026-12-31, and how backlog turns into GAAP revenue.
- TeraWulf, Anthropic and Google: whether the project financing closes, whether Google’s support takes effect at handover, and whether the $19B appears anywhere but a press release.
- Nvidia’s guarantee and cloud-commitment balances in filings after 2026-11-17, and any new backstops or warrants.
- Rating actions, redemption limits, defaults or distressed buybacks: the first public signs when pressure shows.
Method. Figures come from the F.F research note v4.2 (data as of 2026-11-11), which cites filings, company releases and calls, rating agencies, central banks and international bodies, labeled media reports and third-party estimates. Source IDs such as S-400 · L88 point to the archived source and line. Amounts are US dollars; billions are written $B.