The Estimated $1.65T Beyond Big Tech’s Balance Sheets

Five technology giants carry an estimated $1.65T of debt outside their balance sheets, exceeding the estimated $1.35T they report on them. Alphabet, Amazon, Meta, Microsoft, and Oracle are securing long-lived claims on AI infrastructure through leases, joint ventures, special-purpose vehicles, project debt, guarantees, and bonds. The mismatch is not an accounting footnote: the asset itself has outgrown the server.

The investment unit outgrew the server

The early AI buildout looked like a chip procurement race because chips were the visible scarce input. But accelerators became clusters whose performance depended on networking, cooling, power delivery, software, and physical proximity. The useful asset was no longer the GPU but the operational system around it.

In March 2024, Meta detailed two clusters containing 24,576 Nvidia H100 GPUs each, used for workloads including Llama 3 training. At that scale, output depends on whether the entire facility works as one machine.

A company securing accelerators also needed land, substations, cooling equipment, network capacity, construction labor, and a credible path to electricity. Hyperscalers eventually announced 46GW of AI data-center capacity, enough at full utilization to consume as much electricity as roughly 44.2 million U.S. households. They needed more capital, delivered in forms capable of funding an integrated, multiyear asset.

Sponsors and lenders responded with the AI infrastructure finance stack. They divided each campus into separate claims: sponsor equity, project debt, landlord ownership, tenant leases, power contracts, and future compute revenue. One operating asset could appear on several institutions’ books in different forms.

Meta uses capital structure to secure capacity

Meta is the clearest case because it is using nearly every available structure at once. The company moved $30B of AI-data-center debt off its balance sheet through special-purpose vehicles. It also formed a joint venture with Blue Owl for the $27B, 2GW Hyperion campus in Louisiana while retaining roughly 20% equity.

Meta did not use Hyperion as a substitute for direct spending. It projected $115B to $135B of capital expenditure for 2026, up from $72.2B in 2025. Meta is adding external financing to a large on-book program rather than replacing it. Through the joint venture, Meta can secure more capacity without putting every construction dollar and underlying asset on its consolidated balance sheet.

Meta reportedly turned to BlackRock on two fronts. BlackRock was reported to be leading a debt sale exceeding $12B for Meta’s El Paso data center, while Meta was also reported to be leasing capacity from a BlackRock-backed Pennsylvania project. Neither transaction had been confirmed at publication. If completed, BlackRock-backed entities would finance or own the facilities while Meta held long-duration claims on their output.

Meta and its peers can secure a powered site without owning it outright. They can sponsor a project, retain minority equity, sign a lease, provide a guarantee, anchor construction debt, or contract for the resulting compute. The hyperscaler gains control of capacity while owners and lenders take different shares of the financing risk.

Hyperscalers still owe their counterparties. Leases, guarantees, and project vehicles sequence the obligations and assign construction, credit, operating, and residual-value risk to different parties. Reported debt and annual capex show only what a company borrows directly or spends during the period.

Power turns a cluster into a financeable project

Lenders can use project finance when an asset has identifiable costs, counterparties, and future cash flows. AI campuses increasingly qualify because computing capacity and energy supply must arrive together. The financeable unit is the powered campus, not the empty building or the chip by itself.

Meta’s reported Project Walleye financing made that coupling explicit. The Ohio project sought $3B in loans covering both data-center construction and power assets, described in the report as a first-of-its-kind structure. Lenders were funding the system required to make the building operational, rather than assuming electricity would appear later.

Big Tech’s announced plans were estimated to require another 44GW of capacity by 2028, with electricity shortages identified as a constraint. Sponsors, lenders, landlords, and customers can contractually reassign the costs of delay, but they cannot supply a missing grid connection or operating permit.

Developers, lenders, and tenants therefore track the contracts around the physical plant as closely as the headline construction cost. Power-delivery milestones, lease commencement dates, guarantees, completion tests, and capacity reservations determine when payments start and when rent or compute revenue can begin.

Commitments are the comparative balance sheet

A study of Alphabet, Amazon, Meta, Microsoft, and Oracle estimated that their off-balance-sheet debt grew roughly eightfold from 2022 to about $1.65T, exceeding an estimated $1.35T of on-balance-sheet debt.

No analyst should treat every lease, guarantee, project loan, and unsecured bond as legally identical. Some obligations are direct, some contingent, some cancellable, and some tied to assets or counterparties with their own cash flows. Because the terms vary, analysts must examine the full stack rather than rely on reported borrowings.

Microsoft illustrates the timing problem. Its mainly data-center-oriented finance leases rose by nearly $100B in two years to $108.4B, with much of the capacity not yet commenced. A leverage screen focused on current-period rent or debt misses obligations already negotiated but scheduled to begin when facilities become available.

Oracle illustrates the scale problem. It signed about $150B of data-center leases in one quarter, taking its combined data-center and cloud-capacity commitments to $248B. Microsoft and Meta each added nearly $50B of data-center leases in their most recent quarters, pushing major cloud companies’ combined commitments above $700B.

Oracle also moved $66B of AI-data-center debt off its balance sheet, compared with Meta’s $30B, xAI’s $20B, and CoreWeave’s $2.6B. Although the terms differ, each company gains access to scarce infrastructure while investors hold the ownership and financing exposure.

Analysts need a schedule of claims: amount, commencement date, duration, recourse, asset backing, power dependency, counterparty, and expected utilization. Two companies with similar reported debt can have radically different exposure if one owns operating campuses and the other has signed large leases for capacity still under construction.

The players occupy different layers of one stack

Meta can act as sponsor, operator, minority owner, tenant, and compute seller. Blue Owl supplies capital and asset ownership; BlackRock’s reported projects would do the same. Oracle anchors projects through future leases, while Microsoft has accumulated substantial not-yet-commenced finance leases. Alphabet and Amazon sit inside the same sector-wide estimate, although each company mixes ownership, leasing, and direct investment differently.

Nvidia also operates beyond the equipment layer. It disclosed $3.5B in guarantees to companies leasing land, power, and data-center facilities, four times its previous-quarter level. By guaranteeing counterparties’ access to those inputs, Nvidia is helping finance the system required to run its accelerators.

xAI and CoreWeave represent a more concentrated version of the same economics. Because they have less diversified cash flow than hyperscalers, refinancing terms and counterparty quality show up more directly in their credit risk.

Anthropic occupies the demand side. A model developer can sign direct data-center leases, buy cloud capacity, or rent compute from another technology company. Each route turns future workloads into support for someone else’s financing structure. Model developers, facility owners, utilities, lenders, accelerator suppliers, and compute buyers now split the roles required to deliver AI capacity.

These companies can switch positions from one transaction to the next. They can buy infrastructure in one deal and sell capacity in another, trading access to powered GPU capacity without transferring an entire corporate balance sheet.

Wall Street is underwriting utilization

Developers have sought credit ratings while data centers are still under construction, and rating agencies have expanded their coverage. Banks separately marketed more than $56B of investment-grade construction loans tied to Oracle’s future leases, widening the buyer base for exposure that began as project lending.

Lenders do not rely only on Big Tech’s corporate wealth. They underwrite a chain of events: developers must complete the facility, utilities must deliver power, tenants must take occupancy, and contracted payments must begin. A strong tenant can reduce default risk while still earning a poor return on underused capacity. Credit safety and capital productivity are related, not identical.

Hyperscalers have not abandoned conventional borrowing. Their unsecured-bond supply exceeded $155B by May 2026, more than 45% above total 2025 issuance. They use structured financing to supplement debt-market access, not eliminate their dependence on investor confidence.

Meta was reported to be discussing a roughly $10B, two-year compute rental with Anthropic. If signed, Anthropic’s payments would turn available capacity into contracted revenue and help absorb an early or oversized buildout. Meta would use financing to secure the capacity and a compute customer to support its fixed burden.

Deals like Meta’s would deepen long-duration markets for powered compute. Hyperscalers need not use every campus internally if they can lease or sell capacity across the model ecosystem. An external contract makes demand observable and substitutes a renter’s credit for an internal forecast; it supports financing only while the renter makes its scheduled payments.

Restraint and cancellation are real counter-strategies

Developers do not build every announced campus, and companies do not sign every financing package. Microsoft reportedly walked away from U.S. and European data-center projects totaling 2GW. Hyperscalers can preserve optionality until commitments harden because expected demand can change.

Analysts should not give equal weight to announced capacity, a site under evaluation, a signed lease, commenced rent, guaranteed project debt, and an owned operating campus. They would misstate leverage by counting every proposal, just as they would by ignoring future leases entirely.

Apple offers a broader counter-strategy. Its planned 2026 capex was $14B against a reported $650B combined hyperscaler total, and model commoditization could make that restraint rational. If useful models and compute remain purchasable from others, owning the largest infrastructure base is not automatically the highest-return position. Apple can preserve flexibility by buying capacity after suppliers have absorbed construction and residual-value risk.

Local opposition blocked or delayed 17 U.S. data-center projects worth $98B in the second quarter of 2025, while electricity shortages threatened funded schedules. Cancellation rights, commencement triggers, completion guarantees, and power dependencies determine which sponsor, tenant, or lender pays when planned capacity fails to become operating capacity.

Early movers can secure scarce sites and power before competitors, but long commitments create fixed-cost exposure if model economics improve faster than demand grows or equivalent capacity becomes cheaper elsewhere. Companies preserve an option only when the contract allows them to walk away; without that right, the “option” is just a payment schedule with nicer stationery.

Commitment coverage is the metric that survives the structure

An investor comparing hyperscalers must ask not only how much each spent this year, but what each has committed to control over the buildout’s life. Annual capex captures what a company spent during the period, not the multiyear capacity it secured through contracts and separately financed entities.

Analysts can calculate commitment coverage by matching cash flow from AI workloads with the fixed and contingent payments required to obtain their capacity. They need five answers:

1. What capacity is actually committed? Separate announced projects from signed leases, funded construction, energized facilities, and operating clusters.
2. When do obligations commence? Future leases can matter before current-period payments begin; timing determines the burden.
3. Who bears each risk? Identify sponsor equity, lender exposure, guarantees, tenant obligations, construction risk, and residual ownership rather than treating “off balance sheet” as “gone.”
4. Will power and facilities arrive together? Match power-delivery dates to lease commencement and debt-service schedules.
5. What absorbs the capacity? Internal products, cloud customers, and external compute renters must collectively provide enough utilization and cash flow to cover fixed payments.

Analysts should count more than reported interest expense in the denominator. It also includes lease payments, project-support obligations, contracted power, guarantees that may become effective, and the cost of sponsor capital. The numerator should include only cash generated by workloads capable of using the relevant capacity when its payments are due.

The estimated $1.65T outside the five balance sheets is not one liability and should not be compressed into one adjusted-debt figure. It is a schedule of claims on campuses that must be built, powered, and used. Capital markets can divide those claims among owners, lenders, tenants, and sponsors; they cannot divide away the utilization test. An SPV can move a liability, but it cannot move an idle megawatt into revenue.