Five US Tech Giants’ Hidden AI Debt Hit $1.65 Trillion, and Balance Sheets Tell Only Half the Story

Meta, Oracle, Alphabet, Amazon, and Microsoft have accumulated an estimated $1.65 trillion in hidden AI debt and long-term infrastructure commitments.

That total, reported in a hidden-debt study, has increased roughly eightfold in four years. It also exceeds the companies’ combined balance-sheet debt, estimated at about $1.35 trillion.

The phrase “hidden debt” needs care. Much of this amount is not conventional borrowing, and the companies have not violated accounting rules by excluding it from reported debt.

The figure combines future data center leases, cloud capacity contracts, equipment commitments, energy agreements, and other enforceable obligations. Many appear only in financial statement footnotes because the related facilities or services have not started.

That distinction is important, but it does not eliminate the risk. These contracts can require years of payments even when technology, demand, or AI economics change.

The core conflict is therefore not legal debt versus undisclosed debt. It is visible financial leverage versus economic commitments that behave like leverage under stress.

For years, investors treated the largest technology companies as cash-rich businesses funding AI expansion from operating income. The latest disclosures show a more complicated model.

AI infrastructure now depends on long-duration contracts, leased campuses, outside developers, special-purpose financing, and guaranteed access to chips and electricity. Those arrangements shift some construction risk away from technology companies.

They do not necessarily remove the payment burden.

Five US Tech Giants’ Hidden AI Debt Reached $1.65 Trillion

The headline number captures obligations that ordinary debt totals cannot show, but it also groups together contracts with different financial risks.

Nikkei examined recent disclosures from Alphabet, Amazon, Meta, Microsoft, and Oracle. Its analysis estimated that their combined off-balance-sheet obligations reached approximately $1.65 trillion during the latest reporting period.

These obligations expanded as the companies raced to secure data centers, graphics processors, networking equipment, power, and third-party computing capacity.

The five companies also carried about $1.35 trillion of adjusted or reported debt, according to the analysis. That comparison makes the off-balance-sheet total appear especially striking.

However, a purchase commitment is not identical to a bond. A future lease payment is also different from money already borrowed and spent.

A bond creates a direct obligation to repay principal and interest. A lease commitment usually reflects payments for an asset that becomes available later.

A cloud capacity contract gives the buyer access to computing services. A chip supply agreement secures hardware or production capacity.

Each contract delivers something of value. Calling the full amount debt can therefore exaggerate its similarity to bank loans and bonds.

The economic concern appears when those contracts become difficult to cancel. A company can reduce new capital spending quickly, but it cannot always exit signed leases or minimum-purchase agreements.

Meta offers the clearest example. Its March 2026 quarterly filing disclosed $182.88 billion in leases that had not yet commenced.

Those leases primarily covered data centers, colocation facilities, and network infrastructure. They were scheduled to begin between 2026 and 2036, with terms reaching 30 years.

Meta separately disclosed $237.67 billion in non-cancelable contractual commitments. These covered cloud capacity, servers, networks, data centers, and Reality Labs hardware.

Together, those categories exceeded $420 billion. Meta also signed additional infrastructure contracts worth approximately $24 billion after the quarter closed.

These figures appear in the Meta commitments footnote rather than the main debt line. They still describe future claims on Meta’s cash.

Alphabet reported a similar structure. It disclosed $75.6 billion in future lease payments for leases that had not started by March 31, 2026.

Those leases primarily involved data centers. They were expected to commence between 2026 and 2031 and could run for up to 25 years.

Alphabet also reported $232.7 billion in fixed or minimum contractual commitments. These included technical infrastructure supply, energy, and content agreements.

Its combined disclosed commitments exceeded $308 billion before considering every contingent arrangement. The company’s Alphabet obligations illustrate how AI costs extend beyond construction.

Amazon’s filing showed $569.28 billion in total commitments at March 31, 2026. That amount included leases, financing obligations, purchase obligations, and other commitments.

Approximately $106.35 billion related to leases that had not commenced. Another $103.77 billion involved unconditional purchase obligations.

Amazon also reported approximately $364 billion in unrecognized commitments from certain customer contracts lasting longer than one year. Those contracts represent expected revenue, not financial debt.

That is one reason the aggregate estimate requires careful interpretation. A single filing can contain future payments, financing obligations, and expected customer revenue under separate accounting categories.

Amazon’s contractual commitments nevertheless reveal the extraordinary duration of its infrastructure decisions. Many payments stretch far beyond the current AI product cycle.

Oracle disclosed $261 billion in additional lease commitments as of February 28, 2026. Almost all related to data center arrangements not yet shown on its balance sheet.

Those leases were generally expected to begin between late fiscal 2026 and fiscal 2028. Their terms ranged from 15 to 19 years.

Oracle also disclosed $11 billion in unconditional purchase and other obligations, primarily involving data center power. Its Oracle lease disclosure shows how quickly future capacity became a major liability-like commitment.

Microsoft also spends heavily to build, purchase, and lease data centers. Its filings describe continuing investment in cloud capacity and AI training infrastructure.

The accounting categories differ across companies, so direct comparisons remain imperfect. Still, the direction is consistent across all five businesses.

The AI race has moved far beyond annual capital expenditure. It now involves contractual promises lasting through several product generations.

AI Infrastructure Financing Has Moved Into the Footnotes

The financing model lets technology companies reserve scarce capacity before construction finishes, while delaying some balance-sheet recognition.

A data center can take years to plan, permit, connect to the power grid, build, and equip. AI companies cannot wait until demand arrives before securing that capacity.

They must reserve land, electricity, cooling systems, networking equipment, and accelerators in advance. Suppliers also want long-term commitments before accepting construction risk.

This mismatch creates demand for off-balance-sheet arrangements. A developer or special-purpose vehicle can borrow money, build a facility, and lease it to a technology company.

A special-purpose vehicle is a legally separate entity created to own a project or financing arrangement. Its debt generally stays separate from the customer’s corporate debt.

The technology company signs a long-term lease or service contract. Those payments help the project owner repay lenders.

That structure means the technology company might not guarantee the project’s entire borrowing. If the project fails, lenders usually have claims against the facility and its owner.

Yet the tenant’s contract often provides the cash flow supporting the loan. The economic relationship is therefore more connected than the balance-sheet presentation suggests.

The same pattern appears in equipment financing. A partner can purchase GPUs, place them in a facility, and sell computing capacity through a long-term contract.

This arrangement helps the technology company avoid owning every chip directly. It can also transfer residual-value risk to the financing partner.

However, the buyer may still promise minimum payments. That promise matters if newer chips make the contracted hardware less competitive.

Leasing is common across many industries. Airlines lease aircraft, retailers lease stores, and manufacturers sign long-term supply agreements.

The AI buildout differs because of its speed, scale, and technological uncertainty. A 20-year property lease can outlast several generations of servers and accelerators.

The building itself may remain useful, but its power density, cooling design, and network architecture can become less attractive. Retrofitting those systems costs money.

Accounting standards generally record a lease liability when the lessee gains control of the leased asset. A signed lease for an unfinished facility has not commenced.

Until commencement, the obligation normally appears in the commitments footnote. It is not included in the standard lease liability shown on the balance sheet.

That treatment follows accounting rules. It also creates a timing gap between the commercial decision and its most visible financial effect.

The company has already reserved the facility. Investors may not see the corresponding liability among headline balance-sheet totals until years later.

Purchase commitments create another gap. A company may promise to buy chips, power, or cloud capacity without recognizing the full future payment as current debt.

The arrangement becomes economically debt-like when cancellation is costly and demand falls. It remains more flexible when suppliers can resell the capacity or contracts contain adjustment provisions.

Meta disclosed a contingent obligation to purchase up to $14.72 billion of cloud capacity over five years. The amount could fall if the provider sells capacity elsewhere.

That provision clearly makes the obligation different from a bond. Meta’s payment depends partly on the provider’s success in finding another customer.

Other agreements contain fewer escape routes. Investors need contract-level details to distinguish firm commitments from contingent exposure.

Those details are often limited. Companies usually report aggregate amounts without identifying every facility, counterparty, cancellation clause, or pricing mechanism.

This is the opacity at the center of the story. The commitments are disclosed, but their risk cannot be measured from one headline number.

The Real Conflict Is Flexibility Versus Guaranteed Capacity

Big Tech is buying certainty in chips and data centers by giving up financial flexibility for years.

AI infrastructure requires commitments before anyone knows the final level of demand. That forces companies to choose between shortage risk and overcapacity risk.

Underbuilding has immediate consequences. A company can lose customers when it lacks GPUs, power, or available cloud regions.

Model developers also need large computing clusters during concentrated training periods. Delayed infrastructure can postpone a model, product launch, or enterprise contract.

Overbuilding produces a slower financial problem. The company pays for unused capacity while depreciation, lease expense, and energy costs pressure margins.

The five technology giants have chosen to protect capacity. Their contracts reserve resources before those resources become operational.

This choice makes strategic sense during shortages. It becomes harder to defend if AI revenue grows more slowly than infrastructure commitments.

Meta demonstrates that tradeoff. It operates consumer services with billions of users, but AI infrastructure spending now reaches beyond its traditional advertising model.

The company needs compute for recommendation systems, advertising tools, generative products, and model development. It also wants enough capacity to compete with cloud providers and AI laboratories.

Yet infrastructure arrives in large blocks. A campus designed for gigawatt-scale power cannot expand with the precision of a software subscription.

Meta must estimate future demand years before opening the buildings. Forecasting errors can therefore create unused compute or expensive capacity shortages.

Oracle faces a different pressure. Its cloud growth gives it a reason to secure data centers, but its financial position differs from richer hyperscalers.

Alphabet, Amazon, Meta, and Microsoft generate large cash flows from established businesses. Oracle has relied more visibly on external financing for cloud expansion.

In February 2026, Oracle announced plans to raise between $45 billion and $50 billion through debt and equity during the calendar year.

The company said the plan would support Oracle Cloud Infrastructure while preserving an investment-grade rating. That statement recognizes the balance-sheet tension directly.

Oracle’s backlog can support investment when customers consume contracted capacity. It can also concentrate risk when a small number of AI customers drive expected demand.

The relevant question is not whether those customers signed contracts. It is whether their future revenue and financing can support the full contracted period.

Alphabet has greater financial room, but it faces the same technology cycle. Its infrastructure must serve Google Cloud, Gemini products, search, advertising, and internal research.

Amazon can spread infrastructure across AWS and its broader operations. Still, its $569.28 billion commitment schedule shows how much future activity has already been contracted.

Microsoft combines Azure demand with its own AI services and external model partnerships. That approach can raise utilization, although it also links infrastructure planning to partner demand.

The five companies are not equally exposed. Treating the $1.65 trillion as one uniform debt pile would hide those differences.

Cash generation, contract duration, customer concentration, cancellation rights, and facility reuse all matter. So does the timing of when each obligation begins.

The central reversal remains clear. Companies pursued off-balance-sheet financing to preserve flexibility, yet the contracts themselves reduce flexibility.

They transferred construction and ownership risk to outside partners. In exchange, they promised enough future payments to make those projects financeable.

This is not free capital. It is a different allocation of risk.

Why the $1.65 Trillion Figure Can Mislead Investors

The total is useful as a warning signal, but it should not be treated as $1.65 trillion of conventional corporate borrowing.

A skeptical reading starts with the definition. “Hidden debt” is an analytical label, not a standard accounting category.

Some included obligations represent lease payments for assets that companies expect to use. Others represent purchases of equipment, energy, or computing services.

Those payments generate operating capacity. Debt principal, by contrast, usually reflects money already received.

The aggregate also uses undiscounted amounts. A payment due two decades from now does not have the same present value as a payment due this year.

Balance-sheet lease liabilities typically discount future payments. Commitment tables often present undiscounted contractual totals.

Comparing those numbers without adjustment can overstate the difference. Investors should separate nominal future payments from present-value liabilities.

The analysis can also mix gross and net economic exposure. A technology company may owe lease payments while expecting customer revenue from the same facility.

Amazon’s unrecognized customer commitments demonstrate this issue. Future revenue obligations are commercially important, but they are not debt owed by Amazon.

Likewise, some facilities can serve multiple products or customers. Capacity reserved for one AI workload might support another service if demand changes.

Not every contract has that flexibility. Specialized campuses and tightly configured clusters can be harder to repurpose.

The risk also depends on recourse. Debt held by a special-purpose vehicle may not become the technology company’s legal liability.

If a project defaults, creditors might seize the facility rather than pursue the tenant’s wider assets. The tenant could lose capacity without assuming the project loan.

However, losing a critical data center can damage operations. A company might support a troubled project voluntarily because replacement capacity is scarce.

This creates an implicit commitment without a formal guarantee. Credit analysts often examine such strategic support when assessing off-balance-sheet structures.

The opposite error is also possible. Dismissing every commitment because it is not recorded debt understates the burden of non-cancelable payments.

A company cannot preserve cash by simply refusing to pay a valid lease. Renegotiation can involve penalties, litigation, or loss of essential infrastructure.

Moody’s previously analyzed $969 billion of undiscounted future lease commitments across the same five companies. It identified $662 billion that remained unrecorded.

The lease-risk analysis estimated that the unrecorded amount equaled 113 percent of adjusted debt. Its narrower scope still showed substantial exposure.

The difference between $662 billion and $1.65 trillion does not necessarily indicate conflicting facts. The larger estimate appears to include broader contractual obligations beyond leases.

That methodological distinction should be prominent whenever the headline figure appears. Without it, readers can mistake an economic exposure estimate for audited corporate debt.

The best interpretation lies between alarm and dismissal.

The $1.65 trillion figure does not prove that five technology giants face an immediate solvency problem. Their businesses generate revenue, cash, and valuable infrastructure.

It does show that ordinary debt ratios capture less of the AI buildout than investors may assume. Important commitments sit across multiple footnotes and reporting categories.

Those disclosures provide totals, but often omit the details needed to model downside scenarios. Investors rarely receive full information about utilization guarantees, contract repricing, or exit rights.

The resulting uncertainty deserves attention. A small change in demand can have a large effect when fixed commitments are enormous.

Who Comes Under Pressure If AI Demand Misses Expectations

The first pressure will appear in margins, cash flow, and contract renegotiations, not necessarily through an immediate wave of defaults.

Data center leases usually begin when facilities become available. A company’s expense burden can therefore rise after the original contract was signed.

That delay matters during rapid expansion. Today’s financial statements reflect only part of the infrastructure already ordered.

As more facilities commence, companies recognize assets and lease liabilities. They also record rent, depreciation, interest, energy, and operating expenses.

Reported leverage can rise even without another financing announcement. Free cash flow can weaken as payments accelerate.

Meta’s lease schedule illustrates this pipeline. Its $182.88 billion of uncommenced leases was due to start across a ten-year window.

Oracle’s $261 billion of uncommenced data center leases was expected to begin much sooner, generally by fiscal 2028.

The timing makes Oracle particularly important to watch. Its infrastructure obligations can become visible while its cloud business is still scaling.

Pressure can spread beyond technology companies. Data center developers borrowed and raised private capital based on anticipated tenant payments.

Utilities expanded generation and transmission plans around projected electricity demand. Equipment suppliers reserved manufacturing capacity for large orders.

Banks, bondholders, and private-credit funds financed projects using contracts with major technology companies as support.

If tenants reduce expansion, project developers can lose expected growth. If tenants challenge contracts, creditors may question the value of unfinished facilities.

GPU residual values create another concern. Accelerators can retain value when demand exceeds supply, but new generations can change performance economics quickly.

A lender financing servers must estimate future resale value. That estimate becomes fragile when hardware ages faster than the loan amortizes.

The largest technology companies can absorb some forecasting mistakes. Their suppliers and financing partners may have less room.

This produces a layered risk structure. The strongest company sits at the top, while developers, equipment owners, and creditors carry project-specific exposure below.

Off-balance-sheet financing can distribute risk across that structure. It can also make the ultimate concentration harder to identify.

Investors should therefore avoid focusing only on corporate default. Several earlier warning signs can emerge first.

Cloud gross margins can decline as new capacity opens before demand arrives. Depreciation and lease expenses can grow faster than related revenue.

Capital expenditure can stay high because existing projects cannot stop midway. Management can then reduce spending elsewhere, including hiring and product development.

Companies may attempt to sublease unused capacity. They can also sell compute externally, restructure contracts, or delay facilities that have not reached construction milestones.

Those responses do not automatically signal failure. They show that management’s original demand forecast changed.

Enterprise customers also face consequences. Cloud providers might use longer contracts or minimum-spend commitments to transfer utilization risk downstream.

AI services could become more expensive if providers need to recover fixed infrastructure costs. Competition could limit that pricing response.

The result can be lower margins rather than higher customer prices. That outcome would directly challenge valuations built around profitable AI growth.

Knowledge workers and enterprise buyers should care because infrastructure financing affects product stability. Capacity shortages, cost controls, or canceled projects can change service availability.

Companies building long-term AI workflows should track provider concentration and contractual dependence. Teams can also maintain a searchable knowledge base for filings, vendor terms, and infrastructure changes.

The financial story eventually reaches product decisions. A cloud provider under margin pressure can revise quotas, discounts, regional availability, or contract terms.

Three Signals Will Test the Hidden AI Debt Thesis

The next test is whether AI revenue, operating capacity, and financial disclosure improve before more long-term commitments become active.

The first signal is utilization as new data centers open.

Companies rarely disclose a simple utilization rate for AI clusters. Investors must infer it from cloud growth, depreciation, lease expense, and management comments.

Strong cloud revenue alongside stable margins would support the infrastructure strategy. It would suggest that demand is arriving fast enough to absorb contracted capacity.

Rising depreciation with slowing cloud growth would weaken that case. It would indicate that infrastructure is entering service faster than customers consume it.

The timing difference deserves close attention. A company can report strong demand today while carrying even faster capacity growth scheduled for future quarters.

Oracle’s fiscal results will offer an especially useful test because many leases begin within a concentrated period. Backlog conversion must keep pace with infrastructure expense.

Meta presents a different indicator. Its advertising and consumer AI products must generate enough incremental value to justify infrastructure not sold through a traditional cloud model.

The second signal is the movement of uncommenced leases into reported liabilities.

Investors should record each company’s uncommenced lease total every quarter. A falling total is not automatically positive because the related leases may have started.

The better measure combines newly commenced liabilities, new commitments, and cash payments. That reveals whether the infrastructure pipeline is shrinking or merely changing accounting categories.

New commitments also matter. Meta added approximately $24 billion of infrastructure contracts shortly after its March quarter ended.

Continued additions would strengthen the argument that the $1.65 trillion estimate understates future exposure. Slower additions would show that companies are regaining flexibility.

Disclosure quality should improve as the totals grow. Investors need clearer separation between leases, purchase obligations, customer-backed capacity, and contingent commitments.

They also need more information about cancellation rights and counterparty concentration. Aggregate totals alone cannot support reliable stress testing.

The third signal is whether companies begin selling, delaying, or renegotiating capacity.

Selling spare compute can be a rational business expansion. It becomes a warning when management originally built that capacity for internal demand.

Project delays can reflect permitting or electricity constraints. Repeated delays following weaker demand would carry a different meaning.

Contract renegotiations would provide the clearest stress signal. They would show that expected AI economics no longer support the original payment schedule.

The same applies to impairment charges. An impairment means an asset’s expected economic value has fallen below its recorded value.

Large impairments involving AI equipment or data centers would weaken the claim that these commitments remain fully productive.

No single quarter will settle the debate. The contracts extend across years, while AI adoption and hardware performance change quickly.

Readers should resist two simple conclusions. The first is that $1.65 trillion of hidden AI debt guarantees a financial crisis.

The second is that wealthy technology companies can absorb every commitment without consequence. Both claims ignore contract structure, timing, and operating performance.

The more defensible conclusion is narrower. Five US technology giants have promised far more future spending than headline debt figures communicate.

Those promises helped them secure scarce AI infrastructure. They also tied future cash flows to demand forecasts made during an exceptional investment race.

Watch utilization, newly recognized lease liabilities, and capacity renegotiations in that order. Together, those signals will show whether the commitments support durable growth or expensive excess.

The hidden AI debt estimate has already changed the question investors should ask. The issue is no longer only how much Big Tech spends on AI.

The issue is how much spending has already become difficult to reverse, and whether future AI revenue arrives before those promises come due.