If AI Demand Disappoints, Who Bears the Data-Center Debt?

Behind the hype of artificial intelligence sits a massive web of debt, long-term leases and equipment loans. The money goes out before the new computing capacity has earned its keep. Once the contracts are signed, rent and loan payments can remain due even if the revenue disappoints.

Nvidia's Ohio deal shows how far those obligations can reach. On August 17, 2026, the chipmaker announced credit support for SB Energy's planned, roughly 8-gigawatt campus. Its initial guarantees cover about 4.25 gigawatts leased to an OpenAI affiliate for 20 years. Those guarantees are capped at $105 billion and become effective as the leases begin and other conditions are met. A company selling the chips is also helping back the buildings that will house them.^1,2

CoreWeave borrows against equipment and customer payments. The specialist cloud provider, known as a neocloud, uses graphics processing units, or GPUs, and the contracts that put them to work to secure financing. It closed a facility with capacity of up to $8.5 billion in March 2026, followed by a $3.1 billion facility in May.^3,4

These are different arrangements, but both connect today's spending to years of expected payments. To see who could absorb a shortfall, it helps to start with how much financing is riding on those expectations.

How much debt is riding on the buildout?

Buildings, chips, cooling and power require financing from several markets:

• Transactions and debt: S&P Global's December 2025 analysis counted about $450 billion of data-center transactions over five years, financed with more than $300 billion of debt. Roughly two-thirds of the spending came in the final two years.⁵
• Equipment spending: S&P estimated about three dollars of equipment, including GPUs and cooling systems, for every dollar spent on the building.⁵
• Corporate bonds: Bank of America raised its forecast for full-year hyperscaler bond issuance to $175 billion in March 2026, Reuters reported.⁶
• Securitized debt: Outstanding data-center asset-backed securities and commercial mortgage-backed securities, or ABS and CMBS, reached $61 billion in 2026, according to Barclays figures cited by the Structured Finance Association in July.⁷
• External funding needs: The same paper cited Morgan Stanley's estimate of a $1.5 trillion external financing gap through 2028. That is an estimate of funding still needed, rather than debt already outstanding.⁷
• Private credit: Commercial Observer reported in October 2026 that some construction lending had shifted from banks toward private credit funds. An April example was a $16 billion PIMCO-led loan for an Oracle-leased Michigan campus.⁸

These figures overlap, so they can't simply be added together. They show how many financing channels depend on the companies expected to pay for the buildout.

Who are the tenants, and how strong is their credit?

A building's finances depend on who owes the rent and where that company gets the cash to pay it. Tenants and computing customers fall into three broad groups:

• Hyperscalers: Microsoft, Alphabet, Amazon and Meta have investment-grade credit and established, cash-generating businesses.
• AI labs: Companies such as OpenAI and Anthropic develop AI models and rely more heavily on outside funding, including venture capital.
• Neoclouds: Providers such as CoreWeave buy computing equipment, often with borrowed money, and rent its capacity to AI labs and technology companies.⁵

CoreWeave's two facilities show why that distinction matters. Its March facility, tied to an investment-grade customer, was rated A3 by Moody's. The floating-rate portion was priced at 2.25 percentage points above the Secured Overnight Financing Rate, or SOFR, a short-term borrowing benchmark. The May facility, serving two non-investment-grade customers, was rated Ba2 and priced at SOFR plus 4.50 percentage points.^3,4

Customer credit helps explain the different terms, alongside other financing conditions. These were facility ratings, rather than CoreWeave's corporate rating. Both facilities were issued through subsidiaries. For the March loan, CoreWeave itself provides a limited guarantee covering specified events, including misconduct and certain bankruptcy actions.^3,4

Some tenants also act as middlemen, renting buildings and selling computing services inside them. Oracle disclosed about $260 billion of leases not yet commenced as of May 31, 2026, substantially all for data centers, generally running 15 to 19 years. It warned that those terms do not typically match its customer contracts. A customer that stops paying or declines to renew can leave Oracle with rent it still owes.⁹

What actually supports repayment?

For a building lender, the expected source of repayment is rent. Blue Owl's Marc Zahr told Commercial Observer that data-center leases commonly run 15 to 20 years, mostly with investment-grade tenants. The publication described construction lending as heavily dependent on preleased projects, with little speculative building.⁸

Guarantees add support, subject to their terms. Under Nvidia's Ohio arrangement, a payment default by the OpenAI affiliate can trigger payment of an agreed shortfall after recoveries from a replacement lease or sale. The tenant has agreed to reimburse Nvidia, though collection still depends on its ability to pay. The guarantees can end early, including if the tenant achieves a satisfactory credit rating or terminates a lease as permitted.¹

Meta's Hyperion campus in Louisiana pairs a four-year initial lease, with renewal options, with a residual value guarantee for the first 16 years of operations. If specified conditions are met after termination or nonrenewal, Meta would make a capped payment linked to the property's value. The guarantee provides support beyond the initial lease, but is not an unconditional 16-year rental commitment.¹⁰

Moody's put five hyperscalers' uncommenced lease commitments at $662 billion at end-2025, Fortune reported. Those commitments were outside balance-sheet lease liabilities at that point. Moody's cautioned that recorded liabilities might not capture some plausible future cash exposures.¹¹

GPU lenders rely on computing contracts and the hardware earning those payments. CoreWeave said its approximately 5.5-year May facility was designed to align funding with equipment deployment and useful life, collecting repayment while the machines remain commercially useful.⁴

Why do buildings and GPUs hold value so differently?

A powered building can serve several generations of chips. Older GPUs may still work while newer models make them less competitive. That changes the time available to repay a loan and what a lender might recover if payments stop.

Companies spread equipment costs over an estimated useful life. A longer estimate generally means a smaller annual depreciation charge and higher reported earnings, all else equal. It does not establish what the hardware could fetch in a sale.

Amazon shortened the estimated life of a subset of servers and networking equipment from six years to five, effective January 2025, citing faster technological development, particularly in AI. Meta moved the other way, extending certain server and network asset lives to 5.5 years from 2025. Alphabet and CoreWeave's 2025 annual reports use six years for the relevant equipment.^12,13,14

Critics argue that the schedules flatter earnings. Forbes reported in June 2026 that investor Michael Burry had estimated roughly $176 billion of understated industry depreciation for 2026 through 2028. That is his assessment, rather than an established loss.¹⁵ Faster obsolescence could mean earlier accounting charges and weaker loan collateral. A depreciation schedule alone does not settle either question.

What does Lyn Alden see in the financing?

In her October 19, 2026 Wealthion interview with Maggie Lake, macro analyst Lyn Alden pointed to rising lease obligations and chip companies helping finance the data centers that buy their products. She described the connections as “circular complexity.”¹⁶ A supplier can depend on the same customer for equipment sales, investment returns and repayment of financing it has supported.

Alden also questioned the true useful life of leading-edge GPUs. She regarded these as valid concerns but thought a reckoning would take more time. Asked whether AI obligations could rival the subprime mortgage problem, she acknowledged longer-term risks while saying the situation had not reached that point.¹⁶

Where would losses land, and in what order?

The International Monetary Fund, or IMF, examined similar connections in October 2026. Its review of 14 transactions since 2023 identified about $300 billion of equity investment and $650 billion of debt associated with $1.4 trillion of future revenue commitments. These are totals for the reviewed deals. The IMF warned that distress at companies receiving investment could weaken expected revenue, investment values and the associated debt.¹⁷

There is no fixed order across separate businesses. Share prices can fall before a payment is missed, while a tenant failure can reach a building lender immediately. The main exposures are:

• Equity holders and venture investors: Lower revenue expectations can reduce the value of AI labs and leveraged neoclouds before a default occurs.
• GPU lenders and private credit funds: Missed customer payments can leave lenders relying on hardware worth less than the loan balance. Stronger customer credit provides support, subject to the contract and the customer's ability to pay.
• Middlemen and vendor financiers: A company can lose customer revenue while remaining bound by a longer lease. Chipmakers that guarantee customer obligations or own customer equity can share in the shortfall.^1,9
• Building owners and their lenders: Tenant failure can interrupt rent during a lease. Renewal risk arrives later: S&P's December 2025 analysis placed a cluster of lease expirations roughly a decade ahead, around the mid-2030s.⁵
• Large hyperscalers: Leases and guarantees remain obligations even when returns disappoint. S&P expected their substantial financial resources to cushion the impact of slower AI adoption.⁵

On a campus dependent on one tenant, failure can quickly reach property lenders and ABS or CMBS investors. The IMF also warned that illiquid private credit held by pensions and insurers could amplify stress.¹⁷

Retail investors can share in losses through shares in the companies involved or through stock and bond funds holding their securities. Their exposure depends on the underlying holdings.^18,19

How do ABS and CMBS protect senior bondholders?

Within a securitized deal, payment priority is clearer. Debt is divided into layers, or tranches. Senior notes have priority for specified payments; junior investors receive what remains. Credit losses generally work in reverse, reaching junior layers before senior bondholders, according to the deal's terms.²⁰

KBRA analysts described common ABS safeguards in March 2026: cash traps tied to how comfortably income covers debt payments, earlier repayment triggers, six-month senior interest reserves and tests that redirect cash when borrowing is high relative to property value. If performance weakens, these mechanisms can send cash toward senior debt repayment instead of distributions to the owner.²¹

These safeguards do not create a replacement tenant. KBRA noted that hyperscale properties often depend on one or a few tenants. Some subordinate classes can pay interest by adding it to the amount owed. The analysts also said there was no workout history yet for assessing hyperscale recoveries under stress, limiting the evidence for a severe downturn.²¹

What's the case that the debt is well protected?

The strongest counterargument begins with how little space is empty. CBRE reported 1.4% vacancy across North America's primary data-center markets in the first half of 2026, with Northern Virginia at 0.2%. More than 80% of capacity under construction was preleased, up from 74.3% a year earlier. Customers had already committed to much of the new supply.²²

Landlords also have long leases with large, cash-generating companies, while guarantees such as Meta's can add support. A permitted site with a grid connection may remain valuable to another tenant after the equipment inside has become less competitive.^8,10,5

S&P contrasted today's reliance on preleasing and limited speculative building with the early 2000s, when data-center vacancy reached 50% to 70% in several major markets.⁵ That difference helps explain lenders' confidence. It does not remove the risk that a particular tenant, contract or location proves weaker than expected.

What happened the last time a network was overbuilt?

The early-2000s telecom bust offers a historical comparison. Congressional testimony in 2002 cited Morgan Stanley figures showing more than $1 trillion raised in equity and debt across the broadband sector since 1998. That excluded bank loans, vendor financing and securitizations.²³

US telecom market value peaked at $2.7 trillion in March 2000 and fell by $1.7 trillion over the following year, according to the same testimony. Global Crossing filed for bankruptcy on January 28, 2002, listing about $22.4 billion in assets and $12.4 billion in debt. Shareholders and creditors absorbed the financial consequences of capacity that had cost more to build than its revenues could support.²³

Today's largest AI tenants bring much stronger cash-generating businesses to their commitments. Leveraged neoclouds and vendors financing their customers share more of the earlier pattern. The comparison illustrates a financing risk: a useful network can survive even when the companies and investors that paid to build it suffer losses.

So who bears the debt?

The borrower remains responsible for repayment, and a guarantor may have to pay when agreed conditions are met. Investors bear losses according to their claims and the cash or assets available to meet them.

For an equipment loan, customer payments and hardware value are central. For a building loan, rent, tenant strength and the site's usefulness matter over a longer period. Strong tenants and long contracts provide support, while weaker customers, fast-changing hardware and obligations that outlast customer revenue can leave less room for disappointment.

Investors financing the same technology can therefore face very different outcomes. The headline numbers show the scale of the commitment. The contracts determine who must keep paying and who could absorb a shortfall, including retail investors in the companies and funds involved.

Frequently asked questions

What is a neocloud? A neocloud is a specialist provider, such as CoreWeave, that rents GPU computing capacity to AI labs and large technology companies, often using debt secured by its equipment and contracts.

Why do GPU loans have shorter terms than building loans? Computing equipment faces a faster competitive cycle. The cited CoreWeave facilities run about 5.5 to 6 years, compared with large-tenant building leases that often run 15 to 20 years. Accounting useful lives do not guarantee resale value over the loan term.^3,4,8

What is a residual value guarantee? It is a promise to cover an agreed shortfall in an asset's value. The contract defines the conditions, calculation and cap. Meta's Hyperion guarantee covers the first 16 years of operations, alongside a four-year initial lease.¹⁰

Are banks heavily exposed? Banks lend for construction, although some lending has shifted toward private credit. That shift does not establish how much risk banks retain. The IMF has warned that connections between banks and nonbank lenders provide more channels through which financial stress could spread.^8,17

Sources

¹ Nvidia, Form 8-K, August 17, 2026, Item 1.01; SB Energy lease guarantees, capacity, cap and payment conditions. ² Nvidia, Second Quarter Fiscal 2027 CFO Commentary, August 2026; PORTS-Pike capacity and 20-year OpenAI leases. ³ CoreWeave, Form 8-K and Exhibit 99.1, March 31, 2026; March financing agreement, facility capacity, rating, pricing and March 2032 maturity. ⁴ CoreWeave, CoreWeave Closes $3.1 Billion Loan Facility, May 18, 2026; customer credit, rating, pricing and approximately 5.5-year maturity. ⁵ S&P Global, Optimism Grows as Capital Flows, but What if AI Promises Fade, December 2025; financing, equipment costs, tenant risk and lease renewals. ⁶ Reuters, Analysts Revise AI Hyperscaler Debt Forecasts After Amazon Bond Sale, March 17, 2026; BofA forecast (secondary reporting). ⁷ Structured Finance Association, How Data Center ABS and CMBS Fit in a Broader Financing Ecosystem, July 23, 2026; Barclays and Morgan Stanley estimates. ⁸ Commercial Observer, With Data Centers, Lenders Are Getting a Lot More Selective, October 27, 2026; lending mix, Michigan financing and industry interviews. ⁹ Oracle, fiscal 2026 Form 10-K; May 31, 2026 lease commitments, Note 9, and customer-contract mismatch risk. ¹⁰ Meta, Hyperion Data Center Joint Venture Announcement, October 21, 2025; initial leases and conditional residual value guarantee. ¹¹ Fortune, Moody's Flags $662 Billion Risk at the Heart of the Data Center Build-Out by Just 5 Companies, February 25, 2026; Moody's analysis of end-2025 lease commitments (secondary reporting). ¹² Amazon, 2025 Form 10-K; server and networking-equipment useful-life changes effective January 1, 2025. ¹³ Meta, 2024 Form 10-K; useful-life changes effective in fiscal 2025. ¹⁴ Alphabet and CoreWeave, 2025 Forms 10-K; accounting useful lives of servers, network assets and computing equipment. ¹⁵ Forbes, GPU Debt Has Gone Investment Grade Here Is Who Holds the Risk, June 9, 2026; Michael Burry's depreciation estimate (secondary reporting). ¹⁶ Wealthion, Lyn Alden interview with Maggie Lake, released October 19, 2026; transcript accompanying Gold Revaluation: Lyn Alden on Debt, Inflation and AI. ¹⁷ International Monetary Fund, Global Financial Markets Face Rising Risks From AI Bond Yields and Cyberattacks, October 13, 2026, and October 2026 Global Financial Stability Report. ¹⁸ Securities and Exchange Commission, Investor.gov, Mutual Funds; fund holdings, investment losses and distributions. ¹⁹ Securities and Exchange Commission, Investor.gov, Bond Funds and Income Funds; holdings and credit risk. ²⁰ National Association of Insurance Commissioners, Asset-Backed Securities, updated August 7, 2025; Federal Reserve, Commercial Bank Examination Manual, securitization section; payment priority and first-loss layers. ²¹ MBA NewsLink, Roundtable Discussing Data Centers Demand and Deal Structures With KBRA, March 2026; structural safeguards, tenant concentration, interest in kind and hyperscale recovery history. ²² CBRE, North American Data Center Demand Continues to Outpace Supply Despite Record Construction Activity, September 8, 2026; first-half 2026 vacancy and preleasing. ²³ US House Financial Services Committee, written testimony on Global Crossing and the telecom collapse, 2002; capital raised, market-value losses and bankruptcy figures.

Wealthion is a dba of Greylock Peak Investments, a registered investment adviser. This article is general information only. It is not investment advice or an offer or solicitation, and it is not tailored to your circumstances. Guests’ views are their own and may not reflect those of Wealthion. Wealthion has not verified guest statements, including any performance figures. Investing involves risk, including loss of principal, and past performance does not guarantee future results. Wealthion refers readers to other registered investment advisers and shares in the fees they receive from referred clients, which creates a conflict of interest. Please see our Form ADV Part 2A and consult your own financial advisor before investing.