Big Tech Hidden Debt: What the $1.65 Trillion Means

America’s largest technology companies have accumulated an extraordinary volume of obligations connected to artificial-intelligence infrastructure. Some are conventional bonds. Others are leases for data centers that have not yet opened, contracts to buy chips and power, construction commitments, guarantees for financing partners, and promises to support customers whose spending ultimately flows back to the same technology ecosystem.

That is the substance behind the widely circulated estimate that five U.S. technology giants—Alphabet, Microsoft, Amazon, Meta Platforms and Oracle—carry roughly $1.65 trillion of “hidden debt.” The number is alarming, but the label is imprecise. Much of the total is not debt in the ordinary sense, and much of it is not hidden. It is disclosed in securities filings, often in footnotes describing future lease payments, purchase commitments and contingent obligations. The disclosures can be dense, scattered and difficult to compare, but they are not the secret special-purpose entities and falsified accounts that defined Enron.

The more important issue is economic rather than criminal. Big Tech is committing capital at a pace that has begun to outrun even its prodigious cash generation. The industry is shifting from a period in which most infrastructure could be funded from current operating cash flow to one in which leases, bonds, equity issuance, customer prepayments, joint ventures and vendor financing all matter. Those structures do not automatically indicate misconduct. They do, however, transfer more of the AI boom’s risk into long-dated fixed payments and interconnected financing arrangements.

The central question is therefore not whether the companies have concealed a second balance sheet. It is whether future AI demand will be large and profitable enough to support the obligations that are already being signed. If utilization remains high, customers honor their contracts and AI services produce durable margins, the commitments may look like rational advance purchases of scarce capacity. If demand disappoints, the same contracts can become expensive reminders that a legally disclosed obligation can still be economically dangerous.

Last updated: August 4, 2026, 11:30 a.m. ET.

Key Takeaways

• The $1.65 trillion figure is not a standardized debt measure. It combines uncommenced leases, purchase and construction commitments and other contractual obligations reported on different dates and under different definitions.
• Most of the obligations are disclosed. U.S. accounting rules generally recognize a lease liability when the lease commences, while requiring material commitments for future leases to be described in the notes.
• The Enron comparison is misleading. Enron used deceptive entities and fraudulent accounting to misstate its condition. Big Tech’s principal AI commitments are generally visible in public filings, even when they are difficult to assemble.
• The scale is nevertheless unprecedented. Alphabet disclosed $811 billion of material purchase and other contractual obligations at June 30, 2026; Meta disclosed $349.31 billion of non-cancelable contractual commitments and $278.99 billion of uncommenced lease obligations.
• Cash-flow pressure is becoming visible. Amazon reported negative trailing free cash flow after sharply increasing capital expenditure, while Alphabet’s second-quarter capital expenditure exceeded operating cash flow.
• The real risk is interconnected financing. Chip suppliers, cloud providers, AI laboratories, infrastructure funds and data-center developers increasingly finance one another, which can amplify losses if a major customer or project fails.
• The decisive test is monetization. AI adoption is real, but current evidence shows a wide gap between capital spending and measurable productivity or revenue at many customers.

What the $1.65 Trillion “Hidden Debt” Estimate Actually Measures

Nikkei Asia’s estimate brought together future obligations disclosed by Alphabet, Microsoft, Amazon, Meta and Oracle and described the combined amount as off-balance-sheet debt. Secondary reconstructions of the calculation indicate that the total included approximately $821.4 billion of leases that had been signed but had not yet commenced and about $829.2 billion of purchase, construction and related commitments.

That aggregation is useful as an attention-grabbing measure of future claims on cash. It is not equivalent to adding five companies’ bank loans and bonds. A bond is a funded financing transaction: the company receives cash today and promises principal and interest in return. An uncommenced lease is a contract for future access to an asset that may still be under construction. A purchase commitment is a promise to buy goods or services, often in exchange for chips, electricity, network capacity or completed infrastructure. A guarantee is contingent: the company pays only if another party fails to meet an obligation. These categories differ in timing, seniority, accounting treatment and economic risk.

The inputs also do not necessarily share the same measurement date. One widely circulated reconstruction noted that some components were drawn from annual filings while others came from more recent quarterly reports. Microsoft’s purchase and construction commitments in that reconstruction, for example, were based on an older reporting period than some lease figures. That matters because the commitments have been rising so quickly that a six- or twelve-month difference can materially distort comparisons.

There is another important mismatch. Future lease commitments are commonly disclosed as undiscounted cash payments, while the lease liability eventually recognized on the balance sheet is measured at present value. A dollar payable twenty years from now is not economically identical to a dollar due next month. Adding gross future payments across long contracts produces a larger number than the liability that would be recorded when those leases begin.

Renewal options can create further confusion. Nvidia’s agreement connected to Hut 8’s Beacon Point data-center project in Texas was reported as being worth as much as $50.2 billion. Yet Reuters, citing Financial Times reporting, said the base 15-year contract value was approximately $19.6 billion, with the larger figure including renewal options. Describing the full amount as a present, unavoidable debt obligation would exaggerate what has actually been contracted.

None of these qualifications makes the total irrelevant. On the contrary, the exercise reveals how much capital the AI buildout may require before projects begin generating revenue. The better description is future contractual commitments and contingent exposures. That phrase is less dramatic than hidden debt, but it is more faithful to the underlying economics.

Why the Commitments Are Not Yet on the Balance Sheet

The timing of recognition is central to the controversy. Under U.S. generally accepted accounting principles, a company that leases an operating data center generally records a right-of-use asset and a corresponding lease liability when the lease commences—meaning the asset has been made available for the company’s use. Signing a contract years before construction is completed does not necessarily create an immediately recognizable lease liability.

That rule reflects an exchange that has not yet begun. The tenant has not received control of the facility, and the landlord may still need to obtain permits, connect power, install cooling systems, complete construction or meet technical specifications. Recognition at commencement aligns the liability with the point at which the company gains the economic benefit of the leased asset.

Material agreements are not supposed to disappear before that date. Accounting guidance requires companies to disclose significant leases that have been executed but have not commenced, including enough information for investors to understand the nature and expected timing of the obligation. That is why the figures appear in footnotes rather than in the face of the balance sheet.

Purchase commitments follow a related but distinct logic. If Alphabet signs a multi-year agreement to buy processors, energy or cloud infrastructure, it has not necessarily incurred a present liability for the full contract amount on day one. The supplier has not yet delivered all the goods or services. As performance occurs, the purchaser records assets or expenses and the related payable. Before delivery, the commitment may be binding and economically important without meeting the accounting definition of a recognized liability.

This distinction can feel artificial because a contract may be difficult or expensive to cancel. Investors care about whether cash will leave the company, not only whether an accounting standard permits recognition today. That is why commitments deserve analysis even when they are not liabilities under GAAP. But calling every future payment debt obscures what the company receives in return. Many purchase commitments will become servers, buildings, electricity and computing capacity that support future revenue. A bondholder provides financing; a supplier provides an asset or service.

Guarantees add another layer. A guarantee can create a disclosed maximum exposure without producing an immediate cash outflow. Alphabet, for example, reported billions of dollars of financial guarantees and much larger notional exposure through credit derivatives and other backstop arrangements at June 30, 2026. The economic cost depends on whether the underlying counterparties perform. A guarantee that never pays is not equivalent to debt outstanding, but it still exposes the guarantor to losses precisely when the underlying project or customer is under stress.

Accounting recognition therefore answers a narrow question: when does a present obligation meet the standard for appearing as a liability, and at what amount? Investors need to ask a broader one: how much future cash is committed, how certain are those payments, what assets or revenues stand behind them, and what happens if the expected demand never arrives?

Gross Future Payments Are Not Present-Value Liabilities

Lease disclosures often present undiscounted payments over the life of a contract. The recognized lease liability, once the facility becomes available, is generally the present value of those payments. Discounting matters most for the very long data-center leases now being signed. Meta disclosed terms extending as long as thirty years; Alphabet disclosed uncommenced leases with terms as long as twenty-six years. A simple sum of all nominal payments can overstate the current economic burden, although the appropriate discount rate and the risk of future payments remain important.

Inflation can also change the interpretation. If a contract fixes nominal rent for decades, the real burden may decline over time. If payments escalate with inflation, power prices or capacity use, the opposite may be true. Public summaries rarely preserve this contract-level detail, which is another reason a single trillion-dollar headline cannot substitute for security-by-security analysis.

Commitments Can Still Restrict Strategic Flexibility

Even when a commitment is matched by future assets, it reduces flexibility. A company that has reserved vast quantities of capacity cannot redirect all its capital when technology changes, customers migrate to a different architecture or regulators intervene. The useful life of an AI accelerator may be far shorter than the lease on the building that houses it. Power connections and cooling systems can have long economic lives, but the profitability of a specific data-center location depends on demand, electricity costs and network design.

The accounting is therefore not the scandal. The investment decision is. The footnotes show how much of management’s future freedom has already been exchanged for access to scarce infrastructure.

Why the Enron Comparison Fails

Enron was not merely a company with complicated footnotes. It was a fraud built around false financial performance, concealed losses, abusive related-party structures and misleading consolidation decisions. The FBI’s account of the investigation describes a conspiracy that overstated Enron’s financial condition and led to convictions across senior management. Investors were not simply inattentive to a properly disclosed future lease schedule; they were given financial statements that did not fairly represent the business.

That distinction matters because “Enron” is an allegation of deliberate deception, not a synonym for leverage. Alphabet’s hundreds of billions of purchase commitments are disclosed in an SEC filing. Meta’s uncommenced leases are disclosed in an SEC filing. Microsoft has repeatedly published the value of leases not yet begun, allowing readers to see the number rise from $92.7 billion at June 30, 2025 to $329.1 billion a year later. Oracle’s expanding lease commitments have been discussed in credit research and its own filings. Investors may dislike the structures, disagree with the estimates or conclude that management is taking too much risk. Those judgments are different from alleging fictitious accounts.

Post-Enron regulation was explicitly designed to make material off-balance-sheet arrangements more visible. Section 401 of the Sarbanes-Oxley Act directed the SEC to require disclosure of arrangements that may have a material current or future effect on financial condition, liquidity or capital resources. The resulting framework does not guarantee simple reporting, but it reflects the lesson that economic obligations can matter even when they do not appear as conventional debt.

The analogy is not entirely useless. Enron remains a reminder that formal compliance does not eliminate the need for skepticism, especially when transactions are complex, related parties are involved and management benefits from favorable presentation. The AI ecosystem contains special-purpose vehicles, infrastructure funds, customer financing and reciprocal commercial relationships. These deserve close scrutiny. Yet the relevant warning is about complexity and incentives, not evidence of an Enron-scale accounting conspiracy.

A fair conclusion requires three separate statements:

• The future obligations are large enough to affect valuation and credit analysis.
• The accounting treatment can delay balance-sheet recognition and make comparisons difficult.
• The available public evidence does not support describing the principal commitments as concealed criminal fraud.

Conflating those statements produces a sensational headline but a weak analysis.

Alphabet: The Largest Commitment Number, and the Clearest Warning About Aggregation

Alphabet’s June 30, 2026 quarterly filing provides the strongest example of why “hidden debt” is simultaneously provocative and misleading. The company disclosed $811 billion of material purchase commitments and other contractual obligations, including $200.7 billion classified as short term. That figure is larger than the annual economic output of many countries, and it dwarfs Alphabet’s recognized debt. Yet the filing does not say Alphabet borrowed $811 billion. It says the company entered agreements under which it expects to purchase infrastructure, services and other inputs.

The same filing reported $85.2 billion of leases that had not yet commenced, mainly for data centers expected to begin between 2026 and 2031, with terms ranging from one to twenty-six years. Alphabet also entered a separate short-term lease with an estimated $5.8 billion commitment beginning in the third quarter of 2026. Once the facilities become available, qualifying lease obligations move onto the balance sheet at present value.

Alphabet’s physical investment is already visible. Technical infrastructure had a gross carrying amount of $247.18 billion at June 30, while assets not yet placed in service totaled $122.81 billion. Net property and equipment reached $321.21 billion. The uncommenced commitments therefore sit alongside an enormous installed base, not in place of it.

The company’s financing mix also changed materially in 2026. Alphabet issued $20 billion of U.S.-dollar notes and approximately $31.8 billion of foreign-currency notes. It also raised equity through several transactions. The details are important because public descriptions of an “$85 billion equity raise” can imply that the full amount was completed immediately. According to Alphabet’s second-quarter SEC filing, completed net proceeds by June 30 included approximately $20.5 billion from Class A and Class C common stock, $10 billion from a private placement to Berkshire Hathaway and $19 billion from mandatory convertible preferred stock. That is about $49.5 billion of completed net proceeds. A further $40 billion at-the-market program had been authorized, but no shares had been sold under it by quarter-end.

The distinction does not make the capital program small. Reuters reported the announced package as an $80 billion equity raise, alongside Alphabet’s plan to spend approximately $180 billion to $190 billion on capital expenditure in 2026. The company was using debt and equity simultaneously because the investment opportunity was too large to frame as a routine annual capital budget.

Alphabet also disclosed contingent exposures. At June 30, financial guarantees had a maximum exposure of about $7.6 billion. Credit derivatives and related arrangements added $43.8 billion of notional backstop exposure, while another estimated $24.1 billion of potential support remained subject to final terms. The filing also described $20 billion of future capital funding commitments to a private company, contingent on milestones.

These numbers should not simply be added to the $811 billion. Some categories can overlap economically or relate to the same broader transactions. A maximum guarantee exposure is not the same as a purchase commitment, and a notional amount is not necessarily the expected loss. The correct reading is that Alphabet has become a financier, guarantor, buyer and builder across the AI supply chain. Its risk cannot be captured by a single debt ratio.

Cash Flow Shows the Immediate Cost

Alphabet generated $39.1 billion of operating cash flow in the second quarter of 2026 and spent $44.9 billion on capital expenditure. A simple operating-cash-flow-minus-capex calculation was therefore negative by approximately $5.8 billion before considering the company’s preferred free-cash-flow adjustments. Alphabet remained exceptionally liquid, with approximately $242.5 billion of cash, cash equivalents and short-term marketable securities. The issue was not near-term solvency. It was the direction of travel: investment had reached a scale at which even one of the world’s most cash-rich businesses needed external capital and contractual financing.

The bullish interpretation is straightforward. Alphabet sees demand for cloud computing and AI services that exceeds available capacity, and it is securing scarce power, land, chips and construction years in advance. The skeptical interpretation is equally clear. A company that commits $811 billion before the economics of many AI services are mature has less room for error than a company that can reduce spending quarter by quarter.

Meta: A Quarter-Trillion-Dollar Lease Pipeline Before July’s Additional Contracts

Meta’s disclosures illustrate how rapidly the balance-sheet picture can change even when the economics have already been contracted. At June 30, 2026, the company reported $278.99 billion of uncommenced operating and finance lease obligations. The agreements covered data centers, colocation facilities and network infrastructure expected to commence from the remainder of 2026 through 2036, with terms extending beyond one year and as long as thirty years.

Meta then signed additional data-center leases in July with an estimated value of approximately $68 billion. Those facilities were expected to commence in 2027 and 2028, with terms of eighteen to twenty years. The July agreements were not part of the June 30 uncommenced-lease total. Together, the disclosures demonstrate how a balance-sheet snapshot can understate the economic commitments already made by the time investors read the filing.

The company also disclosed $349.31 billion of non-cancelable contractual commitments, primarily for third-party cloud capacity and investment in servers, network infrastructure, data centers and Reality Labs hardware. Approximately $53.52 billion was due during the remainder of 2026 and $81.65 billion in 2027. Meta separately reported contingent obligations of as much as $14.72 billion for cloud capacity over five years and $10.8 billion of restricted cash in escrow related to infrastructure purchases.

Those figures come from Meta’s second-quarter 2026 SEC filing. They are not allegations from an anonymous critic. The filing makes the scale clear, although a reader must navigate multiple notes to see the full picture.

Meta had approximately $84 billion of notes outstanding, with future interest payments of roughly $89.4 billion split between short- and long-term periods. That recognized debt is analytically different from the gross future lease and purchase payments. The former represents funded borrowing. The latter represents infrastructure and services that Meta expects to receive over time.

Meta’s strategy is built around the belief that AI will improve advertising, content recommendations, creator tools and consumer products while supporting entirely new computing platforms. The company has already survived one major investor revolt over capital intensity. Spending on the metaverse triggered sharp criticism earlier in the decade, followed by a “year of efficiency” that restored margins and confidence. The AI cycle is larger and more directly connected to Meta’s core advertising engine, but it also locks in costs for longer periods.

The economic question is utilization. A leased data center operating near capacity can generate attractive returns if Meta’s models improve ad conversion, engagement and monetization. The same building becomes a fixed-cost burden if model efficiency improves faster than demand, if customers move toward smaller local models, or if regulators constrain the data and distribution advantages that support Meta’s AI economics.

Microsoft: Uncommenced Leases Rose From $92.7 Billion to $329.1 Billion in One Year

Microsoft offers the cleanest time series for the acceleration in AI infrastructure commitments. The company reported $92.7 billion of leases not yet commenced at June 30, 2025. The figure rose to $106.2 billion by September 30, $155.1 billion by December 31 and $196.6 billion by March 31, 2026. At June 30, 2026, Microsoft’s annual filing disclosed $329.1 billion of data-center leases not yet commenced, expected to begin between fiscal 2027 and fiscal 2033. Some remained conditional on construction and other milestones.

The increase was not an accounting trick. It was evidence of a company reserving future capacity faster than completed facilities could enter service. As each qualifying lease commences, Microsoft records a right-of-use asset and lease liability. The off-balance-sheet amount should therefore be understood as a pipeline of future balance-sheet additions, subject to contract conditions and present-value measurement.

Microsoft’s existing cloud economics provide a stronger foundation than many AI startups possess. Azure annual revenue exceeded $100 billion, and the company’s cloud contracted backlog reached $678 billion. In the June quarter, total revenue was approximately $90 billion, up 18% from the previous year. Demand for Azure and AI services remained strong enough that capacity constraints, rather than a lack of customers, were a recurring concern.

Yet the cash burden was visible. Reuters reported that quarterly capital expenditure reached approximately $41 billion, up 70% year over year, while free cash flow fell 23% to $19.6 billion. That is not a liquidity crisis; Microsoft remained highly profitable and generated substantial cash. It is evidence that revenue growth and cash conversion can diverge when infrastructure spending accelerates.

Microsoft also extended the estimated useful life of certain data centers and buildings to twenty-five years from fifteen. Longer useful lives reduce annual depreciation expense because the cost is spread over more periods. The change may be reasonable if modern facilities remain productive for longer, especially when shells, electrical systems and cooling equipment outlast several generations of servers. It also raises the quality-of-earnings question that accompanies every estimate change: are reported margins improving because the economics improved, or because the accounting expense arrives more slowly?

The answer is not necessarily one or the other. A data-center campus can remain useful for decades while the accelerators inside it are replaced every few years. Investors should separate building life, power infrastructure, networking equipment and chips rather than treating “AI capex” as one asset class.

Microsoft’s risk is less about finding any demand than earning adequate returns on capacity contracted at today’s prices. Its relationship with OpenAI, the expansion of competing models, and the bargaining power of large enterprise customers all influence whether Azure’s backlog converts into high-margin revenue or expensive pass-through computing.

Amazon: Negative Free Cash Flow Returns as AWS Capacity Expands

Amazon’s AI exposure is embedded in a broader organization that includes cloud computing, retail, logistics, advertising and subscriptions. That diversity makes its obligations harder to isolate, because fulfillment centers, transportation equipment and warehouses also generate large leases and capital expenditure. The AI story is nevertheless concentrated in Amazon Web Services, where demand has pushed the company to accelerate data-center and chip investment.

For the second quarter of 2026, Amazon reported net sales of $200.6 billion. AWS revenue reached $42.2 billion, an increase of 37% from the prior year. Reuters reported that the company raised its expected annual capital expenditure by approximately 10% to $220 billion. AWS backlog rose to about $496 billion from $364 billion in the previous quarter.

The market initially welcomed the growth, sending Amazon shares up nearly 9% in after-hours trading following the results. The operational data supported a bullish argument: customers were signing enough cloud contracts that most AWS capacity for 2027 was already reserved, according to management comments reported after the quarter.

The cash-flow result was more sobering. Amazon reported trailing free cash flow of negative $7.6 billion, compared with positive $18.2 billion a year earlier. The reversal reflected the scale of capital investment rather than a collapse in operating demand. It nevertheless demonstrates why external financing and long-term commitments are becoming more important. An enterprise can report accelerating cloud revenue and still consume cash when new capacity must be paid for before customer revenue is collected.

Amazon’s first-quarter filing also disclosed financing obligations connected to data centers and fulfillment facilities, with approximately $340 million current and $8.2 billion long term at March 31, 2026. Those recognized obligations are only one part of its infrastructure economics. Its annual report said cash capital expenditure reached $128.3 billion in 2025 and was expected to increase in 2026.

AWS has several advantages that reduce risk. It serves a large and diversified customer base, offers mature cloud products beyond generative AI, and can design its own Trainium and Inferentia chips to reduce dependence on Nvidia. Amazon can also use AI infrastructure internally across retail, advertising and logistics. The risk is that capacity expansion becomes a race in which every hyperscaler builds for the same future customers and prices fall faster than unit costs.

Backlog helps, but it is not the same as recognized revenue or cash. Contract terms, cancellation rights, customer concentration and the timing of deployment determine how valuable the backlog ultimately proves to be. The bigger the buildout, the more those details matter.

Oracle: The Most Leveraged Hyperscaler Faces the Sharpest Credit Questions

Oracle occupies a different financial position from Alphabet, Microsoft, Amazon and Meta. It has a large installed software business and growing cloud infrastructure demand, but it entered the AI investment cycle with more leverage and less free cash flow than the largest consumer-internet platforms. That makes its commitments especially relevant to bondholders.

Oracle’s remaining performance obligations reached approximately $638 billion at May 31, 2026, up from $138 billion a year earlier. The backlog reflected a surge in large cloud and AI contracts. In its fiscal third quarter, Oracle had already reported RPO of $553 billion, up 325% year over year, alongside $17.2 billion of quarterly revenue and 84% growth in infrastructure-as-a-service revenue.

Management argued that many large AI contracts were structured to reduce financing risk through customer prepayments or customer-supplied graphics processors. That distinction matters. A customer that prepays part of a data-center build or provides its own expensive chips reduces Oracle’s initial capital burden. It does not eliminate construction, power, lease or execution risk.

CreditSights reported that Oracle disclosed approximately $248 billion of additional lease commitments expected to commence through fiscal 2028, up from roughly $100 billion in the prior quarter, plus approximately $10 billion of new purchase obligations for cloud capacity. These commitments were a major reason investors began to treat Oracle less like a mature software company and more like a capital-intensive infrastructure operator.

The credit market has reason to be more sensitive. Oracle carried substantial long-term debt and increased its spending plan as it pursued contracts linked to OpenAI and other major AI customers. Financial Times reporting described a sharp market reaction after Oracle increased its capital-expenditure outlook, with shares falling 10.8% and credit concerns rising.

Oracle’s bull case is anchored in contracted demand. A $638 billion backlog gives the company unusual visibility if customers perform and capacity opens on schedule. Its database franchise creates relationships with enterprises that may prefer to run AI workloads near existing data. The skeptical case is based on concentration and funding. Large contracts with a small number of AI buyers can create impressive backlog while exposing Oracle to a few counterparties whose own business models are not yet proven.

That is the point at which “off-balance-sheet” analysis becomes more useful than a simple debt total. Oracle’s recognized bonds show funded leverage. Its future leases show capacity it has committed to use. Its backlog shows promised customer demand. Credit analysis must test whether the timing and quality of the backlog match the timing and rigidity of the obligations.

Nvidia: Supplier, Investor, Tenant and Potential Guarantor

Nvidia is not one of the five companies in the Nikkei aggregate, but it sits at the center of the financing network that makes the aggregate important. The company sells the accelerators that power the buildout, invests in customers and infrastructure funds, commits to cloud services, leases facilities and supports projects that create demand for its own products.

At April 26, 2026, Nvidia’s quarterly filing reported $119 billion of manufacturing, supply and capacity commitments, with $95 billion due during the remainder of fiscal 2027. It also disclosed $30 billion of multi-year cloud-service commitments and $6 billion of other vendor commitments. Investment commitments totaled $27 billion.

The company had invested heavily in private companies and infrastructure funds, including $18.6 billion during the quarter. A $1 billion equity-method investment in infrastructure funds carried maximum exposure of approximately $2.3 billion when future commitments were included. Nvidia also disclosed facility lease guarantees with maximum gross exposure of $3.5 billion, partly supported by escrow funded by partners.

These obligations were large, but Nvidia’s cash generation was larger than that of almost any hardware supplier in history. First-quarter revenue reached $81.62 billion, net income was $58.32 billion and operating cash flow was $50.34 billion. Cash and marketable debt securities totaled approximately $50.34 billion, and the company had $25 billion of commercial-paper capacity with no outstanding borrowings at quarter-end.

The Texas data-center lease demonstrates both the strategy and the danger of headline numbers. The agreement associated with Hut 8’s Beacon Point project was reported at up to $50.2 billion, but the base 15-year value was approximately $19.6 billion, with renewal options accounting for the difference. Nvidia’s role helps a developer finance construction and ensures that a major facility uses Nvidia systems. If the project succeeds, Nvidia can benefit as chip supplier, ecosystem owner and potentially investor. If demand fails, Nvidia may face losses through several channels at once.

That pattern resembles vendor financing, a long-established practice in capital-intensive industries. Aircraft manufacturers help airlines obtain financing. Telecommunications-equipment companies extend credit to carriers. Industrial suppliers support dealers and customers. Vendor finance can expand a market that would otherwise be constrained by customers’ balance sheets.

It can also manufacture demand temporarily. A supplier that funds a customer’s purchase records revenue today while retaining exposure to the customer’s ability to pay tomorrow. The accounting can be entirely legal and still weaken the quality of revenue. The correct questions are whether transactions occur at market terms, whether collection is probable, whether the customer has independent sources of cash, and whether the supplier’s downside is limited.

Financial Times reporting has described Nvidia considering or arranging much larger support packages across the AI ecosystem, including potential backing connected to OpenAI. Those reported discussions should not be treated as completed obligations unless final terms appear in public filings. They nevertheless indicate how the chipmaker’s role is evolving. Nvidia is no longer only selling scarce components. It is helping organize the capital structure of the industry that buys them.

The AI Buildout Is Changing Big Tech’s Capital Structure

For much of the past decade, the largest technology companies were criticized for holding too much cash. Their core businesses required comparatively little physical capital relative to the revenue they produced. Search advertising, social networks, operating systems and software subscriptions generated cash faster than management could reinvest it, leading to vast share-repurchase programs and large securities portfolios.

Generative AI reverses that model. Training and serving advanced models require accelerators, networking equipment, power, cooling systems, buildings and land. The leading platforms must purchase hardware before they know exactly how customers will use it or what price competition will permit. They also need to reserve electric capacity and construction slots years before a facility opens. Scarcity pushes them toward long contracts; long contracts create fixed obligations.

The Bank for International Settlements observed in its 2026 Annual Economic Report that the five largest hyperscalers were set to spend more than $1 trillion on AI capital expenditure across 2025 and 2026. The BIS noted that commitments were growing faster than earnings and free cash flow, prompting some companies to issue debt. Goldman Sachs estimated a baseline of roughly $765 billion of annual AI capital expenditure in 2026, potentially rising to $1.6 trillion by 2031 under its broader scenario.

Those estimates use different definitions, but both describe the same structural shift. The investment is too large to be financed solely through the traditional annual budgeting process. Companies are combining several sources of capital:

• Operating cash flow from advertising, software, retail and cloud businesses.
• Public bonds and commercial paper.
• Common equity and convertible securities.
• Long-term leases that spread payments over the useful period of a facility.
• Customer prepayments and minimum-spend commitments.
• Joint ventures and infrastructure funds that bring in outside capital.
• Supplier credit, guarantees and other vendor-financing arrangements.

Each source allocates risk differently. Equity absorbs losses but dilutes existing owners. Debt preserves ownership but requires fixed payments and can threaten solvency. Leases match payments to asset use but reduce flexibility. Customer prepayments reduce funding needs but can create refund or service obligations. Joint ventures share capital requirements but may obscure leverage when readers focus only on the sponsor’s consolidated balance sheet. Guarantees preserve a project’s access to funding while leaving the guarantor exposed in a downturn.

Management’s willingness to borrow can signal confidence, but confidence is not evidence that a project will earn its cost of capital. Corporate history contains many examples of executives using debt to pursue investments they believed were transformative. Some were right. Others locked shareholders into years of restructuring. Financing choice reveals who bears the risk; it does not prove the return.

Why Equity Issuance Is More Informative Than the Debt-versus-Equity Debate

Alphabet’s 2026 capital raise is especially revealing because the company used common stock, a private placement, mandatory convertible preferred securities and bonds at the same time. It was not choosing debt instead of equity. It was expanding every available channel.

Berkshire Hathaway’s $10 billion participation attracted attention because Berkshire is associated with disciplined valuation and businesses that generate cash rather than consume it. The investment is relevant as evidence that a sophisticated buyer accepted Alphabet’s terms. It is not a guarantee that AI infrastructure will deliver an attractive industry-wide return. Berkshire could have negotiated protections, discounts or a security structure unavailable to ordinary shareholders, and its assessment relates to Alphabet as a diversified company rather than to every data-center project.

The signal in the financing is scale. A company with more than $200 billion of liquid assets would not ordinarily need such a broad capital program for routine maintenance. Alphabet was telling the market that the opportunity—and the cash requirement—had moved beyond incremental spending.

Free Cash Flow Is the First Place the Strain Appears

Income statements recognize infrastructure cost over time through depreciation. Cash-flow statements show when the money is actually spent. During a rapid buildout, cash expenditure arrives before much of the associated accounting expense and before the facilities generate their full revenue. That is why free cash flow can deteriorate even while reported revenue and operating income remain strong.

Alphabet’s second-quarter capital expenditure exceeded operating cash flow. Amazon moved from positive to negative trailing free cash flow. Microsoft’s quarterly free cash flow fell despite strong cloud growth. Oracle’s capital needs increased while its debt burden was already substantial. These are not signs that the companies are unprofitable in the ordinary sense. They are signs that the AI investment cycle is consuming a greater share of internally generated cash.

Free cash flow is not a standardized GAAP measure. Companies can define it differently, most commonly as operating cash flow minus purchases of property and equipment. Lease-financed construction complicates the calculation because a company can obtain use of an asset without reporting the full amount as current capital expenditure. Customer-financed equipment and joint ventures can also shift cash timing. Investors should therefore reconcile company-defined free cash flow with capital commitments and changes in lease liabilities.

A useful framework separates four stages:

1. Contracting: The company signs a lease, purchase agreement or guarantee. The cash effect may be small, but strategic flexibility declines.
2. Construction and prepayment: Cash begins to leave through deposits, progress payments, escrow and direct capex.
3. Commencement: The facility becomes available, lease liabilities are recognized and depreciation begins.
4. Utilization: Revenue and operating cash flow must grow enough to cover ongoing payments and replacement investment.

Headline capex captures only part of this sequence. A company can reduce reported cash capital expenditure by leasing a facility, yet still assume a long stream of fixed payments. Another company can buy the same facility outright and report a large immediate cash outflow but lower future rent. Comparing the two requires capitalizing leases and examining total cash commitments, not simply ranking reported capex.

Liquidity Is Strong, but Liquidity Is Not the Same as Return

The largest platforms remain far from a conventional liquidity crisis. Alphabet held more than $240 billion of cash and short-term marketable securities. Microsoft, Amazon, Meta and Nvidia generate substantial operating cash. They can access global bond markets on favorable terms, and their equity valuations provide another source of capital.

That financial strength limits near-term default risk. It does not protect shareholders from low returns on investment. A company can pay every bill and still destroy value if the assets it builds earn less than the cost of capital. In fact, strong balance sheets can permit poor projects to continue longer because management does not face immediate financing discipline.

Credit investors focus on whether fixed payments can be met. Equity investors should focus on whether residual cash flow per share improves after those payments, depreciation, stock compensation and replacement capex. The same infrastructure program can be safe for bondholders and disappointing for shareholders.

Depreciation: The Expense That Arrives After the Cash

AI infrastructure highlights a basic accounting asymmetry. Cash is paid when equipment and facilities are acquired. The expense appears gradually as depreciation over the estimated useful life. During the growth phase, current capital expenditure can greatly exceed current depreciation. Reported earnings therefore reflect only a portion of the cash investment being made today.

This is not aggressive accounting by itself. Matching the cost of a long-lived asset to the periods that benefit from it is a foundation of accrual accounting. The judgment lies in estimating useful lives, residual values and impairment. If a server remains productive for six years, depreciating it over six years is sensible. If rapid model and chip advances make it obsolete after three years, a six-year schedule overstates current profit and delays recognition of the economic loss.

Data centers contain assets with very different lives. Land can remain valuable indefinitely. Buildings and electrical infrastructure may operate for decades. Cooling equipment, networking hardware and generators require periodic replacement. AI accelerators can become economically inferior well before they stop functioning. A single blended depreciation policy can conceal these differences.

Microsoft’s extension of certain data-center and building useful lives to twenty-five years illustrates the judgment involved. The change may accurately reflect a shift toward durable campus infrastructure that can host successive generations of equipment. It also lowers annual depreciation relative to a fifteen-year life. Investors should examine whether maintenance and replacement capex support the longer estimate and whether shorter-lived technology is accounted for separately.

Adjusted earnings measures that exclude depreciation are especially unhelpful for a capital-intensive AI business. Earnings before interest, taxes, depreciation and amortization can help compare operating performance before financing and historical investment choices. It cannot answer whether a company earns an adequate return after replacing the assets required to produce its revenue. For hyperscalers, depreciation is no longer a peripheral accounting charge. It is the delayed income-statement expression of the largest investment program in corporate history.

Stock-Based Compensation and the Buyback Treadmill

Stock-based compensation creates a separate presentation issue. Technology companies often emphasize non-GAAP earnings that add back the expense because issuing shares does not require an immediate cash payment. Economically, employees still receive value from shareholders. Existing owners are diluted unless the company uses cash to repurchase enough shares to offset new issuance.

The cash cost may therefore appear in the financing section of the cash-flow statement rather than in payroll. A company grants stock, excludes the expense from adjusted profit, and then buys shares in the market to prevent the share count from rising. The repurchase is presented as capital returned to shareholders even when a significant portion merely neutralizes employee dilution.

This matters more during an AI spending boom because buybacks compete with infrastructure for cash. A company can preserve adjusted earnings by excluding stock compensation, preserve the share count by repurchasing stock and preserve capex growth by borrowing. None of the steps is necessarily improper, but the combination can make operating performance look more cash-generative than the consolidated capital allocation really is.

Investors can avoid the confusion by tracking three figures together:

• GAAP stock-based compensation expense.
• Gross share repurchases.
• The change in diluted shares outstanding.

If repurchases are large but the share count barely falls, much of the cash is compensating for dilution. A true return of capital should reduce the claim of other shareholders on future earnings or distribute cash directly. Repurchases that merely keep the denominator stable are closer to an employee-compensation funding mechanism.

Vendor Financing: Ordinary Practice, Unusual Scale

Vendor financing is not new. Industrial companies have long helped customers purchase expensive products. Boeing and Airbus support aircraft financing. Equipment manufacturers provide leases and credit. Telecommunications suppliers funded network operators during earlier technology cycles. The logic is commercially rational: a supplier with a strong balance sheet can unlock demand from a customer whose project is viable but whose financing capacity is limited.

The AI version is unusually interconnected. Nvidia invests in model developers and infrastructure funds that buy Nvidia chips. Cloud providers extend credits or guarantees to AI companies that purchase cloud capacity. Data-center developers depend on long-term leases from the same hyperscalers that may finance or guarantee the project. AI laboratories raise capital from strategic investors and then spend much of it on those investors’ computing platforms.

This network can accelerate useful investment. An AI developer may have compelling demand but lack the collateral and history required for conventional debt. A supplier that understands the technology may be better placed than a bank to assess the project. Strategic investors can coordinate chips, power and construction faster than arm’s-length financing markets.

The risk is circularity. Revenue quality weakens when a customer’s ability to buy depends on financing from the seller. The seller may book revenue, gain an equity interest and retain a guarantee exposure to the same counterparty. If the customer succeeds, all three positions can appreciate. If it fails, the supplier can lose the customer, the investment and the guarantee payment at the same time.

Three conditions help distinguish productive vendor financing from demand that is being pulled forward:

• Independent end demand: The customer should have users or contracts that generate cash outside the financing loop.
• Market terms: Pricing, credit support and revenue recognition should resemble transactions with independent counterparties.
• Limited recourse: The supplier’s maximum loss should be identifiable and manageable relative to its cash flow.

The last condition is particularly important for guarantees. An equity investment has a defined maximum loss: the capital invested can fall to zero. A guarantee can transform a customer’s default into a direct claim on the guarantor’s cash. If guarantees are written across correlated AI projects, losses can arrive together rather than independently.

Nvidia’s current disclosed guarantee exposure remains modest relative to its quarterly cash generation. The concern is trajectory. A company whose products dominate an industry has an incentive to support more projects, especially when capital markets cannot independently fund the desired buildout. Each individual transaction may appear affordable. The portfolio can become material before investors receive a simple consolidated view.

The Revenue Test: AI Demand Is Strong, but the Required Return Is Stronger

The investment case for the AI buildout rests on two propositions. First, demand for computing will continue to grow rapidly. Second, the owners of the infrastructure will capture enough revenue and margin to justify the capital. Evidence for the first proposition is already substantial. Microsoft, Amazon, Oracle and Google Cloud have reported strong growth and large contracted backlogs. Enterprises are experimenting with coding assistants, customer-service tools, search, document analysis, marketing and scientific applications. Consumer use has expanded quickly.

The second proposition is harder. A customer can use more computing while the provider earns a poor return if prices decline, model efficiency improves, competition intensifies or expensive capacity sits idle between peaks. Cloud revenue is not the same as AI profit, and AI spending is not the same as AI revenue.

This distinction is essential when evaluating large numerical claims. Gartner forecast that worldwide AI spending would reach approximately $2.52 trillion in 2026, including infrastructure, devices, software and services. That is an expenditure forecast across the ecosystem, not the annual revenue that hyperscalers need to earn from generative AI. JPMorgan Asset Management offered a different framework, estimating that current investment might require roughly $650 billion of annual AI revenue to produce a 10% return under its assumptions. Other models reach higher or lower figures depending on margins, asset lives, utilization and the share of spending attributed specifically to AI.

No single revenue hurdle should be treated as definitive. A data center also runs conventional cloud workloads. AI can improve advertising and retail margins without appearing as a separate subscription. Internal productivity gains may create value without generating external revenue. Some infrastructure supports national-security, research or strategic goals that management may accept at lower direct returns.

Still, the magnitude of the capital program requires a large and persistent economic payoff. A temporary surge in experimentation is not enough. Customers must continue paying after free trials and pilot budgets end. Model developers must charge enough to cover inference costs. Enterprises must find applications that improve revenue, reduce labor cost or manage risk. Consumer products must convert engagement into advertising, subscriptions or transactions.

Backlog Is Encouraging, but It Is Not Cash

Microsoft’s $678 billion cloud backlog, Amazon’s $496 billion AWS backlog and Oracle’s $638 billion remaining performance obligations provide evidence that customers intend to spend. They also reduce uncertainty about capacity utilization. Yet backlog quality varies.

Investors should examine contract duration, cancellation provisions, customer concentration, minimum-spend requirements, pricing adjustments and whether the counterparty has financing independent of the provider. A five-year commitment from a profitable global enterprise differs from a long-dated promise by an AI startup whose largest investor is also its cloud supplier. Both may be legally enforceable, but the latter concentrates ecosystem risk.

Backlog also creates an accounting timing issue. Revenue is recognized as services are provided, not when a contract is announced. The provider may spend heavily to build capacity years before the revenue appears. If construction costs rise or a customer delays deployment, the return can deteriorate even when the nominal contract value remains unchanged.

Price Competition Can Transfer AI’s Value to Customers

Transformative technologies often create enormous social value without guaranteeing extraordinary profits for every infrastructure provider. Railroads changed commerce, but many railroad investors lost money. Fiber-optic networks enabled the modern internet, while excess capacity and leverage destroyed numerous telecom companies. Airlines transformed travel but spent decades struggling to earn their cost of capital.

AI may follow a similar pattern. Competition among cloud providers can push inference prices down. Open-source models can reduce the premium available to proprietary developers. More efficient chips and software can lower the computing required for a given task. Those changes benefit users and expand adoption, but they can reduce returns on facilities financed under earlier assumptions.

The bull case is that lower prices stimulate so much additional usage that total revenue continues to rise. The bear case is that efficiency and competition commoditize the infrastructure faster than demand expands. The outcome will differ across the stack. Scarce power and well-connected data-center sites may retain value even if model providers face price pressure. Specialized chips may earn high margins while older generations depreciate rapidly. Distribution platforms may capture more value than the laboratories that train models.

The Productivity Paradox: Real Adoption, Limited Measured Impact

The most persuasive evidence against dismissing AI as a fad is that businesses are using it. The most persuasive evidence against assuming immediate economic transformation is that many businesses cannot yet identify a measurable effect.

A 2026 digest of global business research published by the National Bureau of Economic Research reported that more than 90% of executives saw no employment effect from AI during the previous three years, while 89% reported no effect on labor productivity. The findings do not show that AI is useless. They show that adoption, organizational change and measured output do not arrive simultaneously.

Productivity statistics often lag general-purpose technologies. Companies first experiment, redesign workflows, clean data, train employees and integrate systems. Some early gains are offset by implementation cost or by employees using time saved to improve quality rather than output. Aggregate data can miss benefits concentrated in particular occupations or young firms.

Recent U.S. data already contain signs of investment impact. A Federal Reserve analysis of the AI buildout found that AI-related components contributed meaningfully to economic growth during parts of 2025 and early 2026. That contribution primarily reflected investment in equipment, software and structures. It does not yet prove that the installed capital has raised economy-wide productivity enough to justify its cost.

The Kansas City Federal Reserve similarly found that recent productivity improvements were not yet broad enough across industries to attribute the shift mainly to AI. The evidence is consistent with an investment boom that may precede a productivity boom rather than confirming one.

Small Businesses May Be Capturing Value Faster

Young companies face fewer integration barriers. They can build workflows around AI from the beginning rather than inserting new tools into decades-old systems. A Gusto survey of new business formation found that 60% of new business owners used AI to help launch in 2025, double the share two years earlier. Half said the tools helped them start faster or at lower cost, although only 3% said they would not have started without AI.

The results suggest useful but incremental benefits: drafting a website, preparing marketing copy, handling administrative tasks, analyzing data or producing basic designs. Those applications can be economically meaningful for a founder paying a small monthly subscription. They do not automatically create enough industry revenue to support trillions of dollars of infrastructure.

Company spending data from Ramp reinforce the gap between broad adoption and concentrated expenditure. Ramp found a median AI spend of approximately $11.38 per employee per month among firms in its data, while the top 1% spent thousands of dollars per employee. The average can therefore be driven by a small group of aggressive adopters. For infrastructure providers, the distribution matters: a few enormous buyers can support rapid growth while making revenue vulnerable to changes in those customers’ funding.

The Big-Market Delusion and the Mathematics of Everyone Winning

Aswath Damodaran and Bradford Cornell describe a recurring investment pattern in “The Big Market Delusion”. A new technology is associated with a vast potential market. Many companies enter. Investors value each company as though it will capture a large share, even though the combined assumptions exceed the size or profitability of the market.

AI has the features that make the pattern powerful. The technology is real. The potential applications span nearly every industry. Market-size estimates are enormous and difficult to falsify in the near term. Companies can point to genuine growth while disagreeing about who will ultimately earn the profits.

The delusion does not require a fake product or dishonest executive. It can emerge from individually plausible forecasts. An investor assumes Microsoft retains cloud leadership. Another assumes Amazon does. Another expects Oracle to capture a major share of AI infrastructure. Nvidia is valued for continued chip dominance, while competing semiconductor companies are also valued for taking share. Model developers, software vendors and data-center operators are each assigned attractive margins. The assumptions may be reasonable one company at a time and impossible in aggregate.

The financing boom intensifies the effect because capital spending becomes evidence for the size of the opportunity. One company’s capex is another company’s revenue. Nvidia’s sales validate data-center demand; the resulting cloud capacity supports AI-laboratory valuations; those valuations enable laboratories to sign larger cloud contracts; the contracts justify more data-center construction. The loop can reflect real growth while still running ahead of end-user cash flow.

The practical defense is not to deny the market’s potential. It is to reconcile company forecasts with a consistent industry total. Investors should ask how much revenue the entire ecosystem can generate, what portion remains after power and hardware costs, and how many firms can earn premium margins simultaneously.

Why Disclosure in a Footnote Can Still Affect Prices

Public information does not become useful merely because it is available. A filing may contain every material number and still require considerable effort to assemble. Commitments can appear in lease notes, contractual-obligation tables, liquidity sections, variable-interest-entity disclosures, guarantee notes and subsequent-event sections. Definitions can change between companies and periods.

Robert Bloomfield’s incomplete revelation hypothesis explains why difficult information may be incorporated into prices more slowly or less completely. Information processing has a cost. A fact buried in a long filing is public, but using it requires time, expertise and attention. Investors rationally devote less effort to facts whose expected trading value appears small.

Richard Sloan’s influential research on earnings quality reached a related conclusion from accounting data. His 1996 study separated the cash and accrual components of earnings and found that the market did not fully reflect their different persistence. Companies with earnings supported more heavily by accruals subsequently underperformed those with stronger cash components. The result became known as the accrual anomaly.

The relevance to AI infrastructure is not that every lease commitment predicts poor returns. It is that cash, accruals and commitments convey different information. A company can report strong earnings while cash is absorbed by construction. A customer contract can support backlog while requiring large prepayments. A longer depreciation life can improve current profit without changing current cash expenditure. Investors who rely on one adjusted metric can miss the interaction.

Why Professional Investors Do Not Immediately Eliminate the Mispricing

Classical finance assumes that informed traders correct prices. In practice, the process is risky and expensive. Andrei Shleifer and Robert Vishny’s work on the limits of arbitrage showed that professional investors depend on capital providers who may withdraw money after short-term losses. An arbitrageur can be fundamentally right and financially unable to hold the position until the market agrees.

Mark Mitchell, Todd Pulvino and Erik Stafford documented similar frictions in situations with apparently obvious valuation discrepancies. Costs, short-sale constraints, timing uncertainty and capital risk allowed mispricing to persist.

Those limits are especially relevant to narrative-driven technology stocks. A skeptical investor may conclude that a company’s obligations are underappreciated but cannot know when the market will care. Revenue can beat expectations for several quarters while the long-term return on new capacity deteriorates. Shorting the stock exposes the investor to unlimited upside risk and potential redemptions before the thesis resolves.

Complex disclosure can therefore “work” without deceiving every sophisticated reader. The marginal buyer may focus on headline growth, adjusted earnings and backlog. The specialist who reads the footnotes may lack the capital, mandate or timing confidence to move the price. Disclosure satisfies the legal requirement while attention determines the market effect.

Regulation After Enron: Better Disclosure, Not Simple Disclosure

Post-Enron reforms improved visibility but did not create a single standardized measure of future obligations. SEC rules require management to discuss material commitments, liquidity and off-balance-sheet arrangements. Lease accounting brings most commenced leases onto balance sheets. Consolidation rules address many special-purpose entities. Auditors and audit committees face stronger oversight.

The remaining difficulty is comparability. Companies use different labels and table structures. One includes certain supplier commitments in a broad total; another discloses them by category. One reports an undiscounted maximum exposure; another reports carrying value or expected loss. One lease includes renewal options; another excludes periods not considered reasonably certain.

Regulators could make the data more useful by requiring a reconciled table with common categories: funded debt, commenced lease liabilities, uncommenced lease payments, unconditional purchase obligations, conditional commitments, guarantees, customer prepayments and joint-venture exposures. Investors could then see which obligations are recognized, which are contingent, what assets are expected in return and how the totals changed during the quarter.

The debate over research-analyst conflicts also deserves precision. The court-supervised Global Research Analyst Settlement imposed after the dot-com era has been modified and its remaining undertakings terminated for major banks. That does not mean all research safeguards disappeared. FINRA has emphasized that Rule 2241 continues to govern conflicts between investment banking and equity research. Investors should still examine whether analysts’ employers underwrite securities for the companies they cover, particularly when those companies require repeated capital raises.

The Bull Case for Big Tech’s AI Commitments

The strongest argument in favor of the spending begins with scarcity. Power, grid connections, land, transformers, cooling equipment and advanced accelerators cannot be obtained instantly. A platform that waits for demand to become certain may discover that competitors have reserved the capacity. Long-term contracts are therefore a rational response to long construction lead times.

Second, the largest platforms possess existing businesses that can monetize AI in multiple ways. Alphabet can improve search, advertising, cloud services, productivity software and autonomous systems. Microsoft can embed AI across Azure, Office, developer tools, security and enterprise applications. Amazon can monetize AWS while improving retail logistics and advertising. Meta can use models to increase engagement and ad conversion. These internal benefits are not captured by a narrow calculation of standalone AI subscription revenue.

Third, many obligations are matched by contracted customer demand. Backlog has risen dramatically. If customers are creditworthy and contracts are enforceable, future revenue can support the lease and purchase schedule. The platforms also retain pricing, product and distribution advantages that startups lack.

Fourth, financial capacity remains exceptional. The companies can absorb temporary negative free cash flow without immediate distress. They can issue long-term debt, sell equity and reduce buybacks. Unlike speculative infrastructure booms financed by weak borrowers, much of the AI buildout is sponsored by companies with some of the strongest cash-generating franchises in the world.

Finally, historical comparisons warn against underestimating demand for computing. Internet traffic, cloud adoption, video streaming and mobile use all grew beyond early forecasts. Cheaper inference can unlock applications that are uneconomic today. If AI becomes embedded in every software interaction, current capacity estimates may prove conservative.

The Bear Case: Fixed Commitments Meet Uncertain Economics

The skeptical case begins with duration mismatch. Companies are signing leases for fifteen, twenty or thirty years to house technology that may change dramatically within three. The building may remain useful, but its location, power design and economics can become less attractive if computing shifts toward smaller models, edge devices, new chip architectures or regions with cheaper electricity.

Second, the industry may be counting the same demand more than once. Every hyperscaler cites AI as a growth driver. Every model developer plans to scale. Every chip supplier expects a large market. The aggregate revenue assumptions can exceed what end users are willing to pay, even when each individual forecast appears plausible.

Third, financing connections can conceal customer weakness. A startup can sign a huge cloud contract because it has raised capital from the cloud provider or chip supplier. The contract becomes backlog for one party and a cash obligation for another, but the ultimate source of repayment remains future product revenue. If that revenue fails to appear, several companies can recognize losses simultaneously.

Fourth, accounting estimates may delay recognition of weak returns. Longer useful lives reduce depreciation. Adjusted metrics exclude stock compensation and other expenses. Lease financing can reduce current capex compared with ownership. None of those choices creates cash, and impairment charges can arrive abruptly when expectations change.

Fifth, strong balance sheets can encourage overbuilding. Management teams are competing for technological leadership and may rationally prefer excess capacity to the risk of shortage. What is rational for each company can be destructive for the industry if everyone follows the same strategy. Capacity gluts usually emerge from individually sensible investment decisions.

The bear case does not require AI to fail. The internet succeeded spectacularly while many internet-era investments failed. AI can transform productivity and still produce poor returns for owners who paid too much for capacity or financed the wrong layer of the stack.

A Practical Stress Test for Big Tech Hidden Debt

The most useful analysis does not ask whether the $1.65 trillion headline is true or false. It breaks the exposures into cash-flow scenarios. Investors can monitor the following indicators each quarter.

1. Operating Cash Flow Versus Cash Capital Expenditure

When capex persistently exceeds operating cash flow, the company must use cash balances, borrow, issue equity or rely on leases and partner financing. One quarter can reflect project timing. A sustained gap indicates a structural funding requirement.

2. Uncommenced Leases and the Commencement Schedule

The total shows future fixed capacity. The start dates show when liabilities and cash payments become more visible. Rapid growth in leases beginning within two years creates a different risk from distant, conditional agreements.

3. Purchase Commitments Due Within Twelve to Twenty-Four Months

Short-term commitments place immediate pressure on liquidity. Alphabet’s $200.7 billion short-term component and Meta’s specified 2026 and 2027 payments deserve more attention than the headline life-of-contract total.

4. Backlog Conversion

Investors should compare growth in remaining performance obligations with recognized revenue and collections. Backlog that grows faster than revenue can be positive if capacity is constrained, but concerning if deployments are delayed or customers depend on fresh financing.

5. Customer Concentration

A diversified enterprise customer base reduces correlated default risk. A few large AI laboratories can create concentration even when the nominal backlog is enormous.

6. Guarantees, Credit Derivatives and Maximum Exposure

These disclosures reveal risks that do not appear in ordinary debt tables. The key is not only the maximum amount but the probability of payment, collateral, recourse and whether counterparties are exposed to the same AI cycle.

7. Depreciation Lives and Impairments

Changes in useful lives affect margins without changing current cash. Rising asset write-downs would indicate that technology or demand is changing faster than accounting assumptions.

8. Stock Compensation and Net Share Count

High repurchases with little decline in diluted shares suggest that cash labeled as shareholder returns is mainly offsetting employee compensation.

9. Power and Utilization

Data centers create value only when electricity, chips and customers arrive together. Delayed grid connections can strand completed buildings; low utilization can destroy unit economics even when demand exists elsewhere.

10. Credit Spreads and Rating Commentary

Bond markets often react to leverage and contingent exposure before equity narratives change. Rising credit-default-swap spreads or negative rating outlooks do not prove distress, but they show that the cost of protecting against default has increased.

How to Build a More Useful Adjusted-Leverage Measure

There is no single accounting ratio that converts every AI commitment into debt without creating new distortions. Investors can nevertheless build a more informative framework by separating funded obligations, fixed operating claims and contingent exposures instead of forcing them into one total.

Start With Recognized Net Debt

The first layer is conventional debt minus cash and highly liquid securities. This measures funded borrowing and the immediate resources available to repay it. Net debt is most useful for Oracle and other companies where bonds are already material. It is less informative for Alphabet or Meta when enormous cash balances coexist with even larger future purchase obligations.

Add Commenced Lease Liabilities

Operating and finance lease liabilities represent facilities and equipment already available for use. They are fixed claims on future cash and should generally be included when comparing companies that lease infrastructure with companies that own it. The corresponding right-of-use asset must also be considered; adding the liability without recognizing the productive asset would overstate leverage.

Show Uncommenced Leases Separately

Future leases should not automatically be added at their undiscounted maximum value. A better approach is to group them by expected commencement year, identify conditional projects and estimate present value where sufficient data are available. Near-term, non-cancelable leases deserve more weight than distant agreements dependent on construction milestones. Renewal options should be separated from base terms unless exercise is reasonably certain.

Distinguish Purchase Commitments by What the Company Receives

A commitment to buy electricity for an operating data center resembles a future operating cost. A commitment to buy servers creates an asset that will later be depreciated. A construction commitment may become owned property. A cloud-capacity commitment may support revenue but leave no residual asset. Treating all four as debt ignores important differences in recovery value and flexibility.

The short-term portion deserves special attention because it competes directly with operating cash flow. Long-term amounts should be assessed against expected revenue, inflation terms and cancellation rights. Analysts should also avoid double-counting commitments already reflected in capital-expenditure guidance or financed through customer prepayments.

Apply Probability Weights to Guarantees

Maximum guarantee exposure is a stress-case number, not an expected cash payment. A probability-weighted estimate can be more useful, although it requires judgment about counterparties, collateral and correlations. The analysis should become more conservative when the guaranteed parties depend on the same technology cycle or financing source. Ten individually low-risk guarantees can behave like one large risk if all customers fail under the same market shock.

Compare Fixed Claims With Recurring Cash Generation

After classifying the obligations, investors can compare annual fixed payments with operating cash flow before growth capex. The ratio should be calculated under several scenarios rather than one forecast. A base case might assume continued cloud growth and high utilization. A downside case could assume lower prices, delayed facility openings and customer defaults. A severe case could add asset impairments and reduced access to equity capital.

This approach produces a range rather than a dramatic single number. That is a strength. The economic risk of AI commitments depends on timing and performance, and false precision can be as misleading as ignoring the footnotes.

Measure Returns on Incremental Capital

The final step is to connect spending with outcomes. Investors should track incremental operating profit and cash flow relative to the growth in invested capital. Company-wide return on invested capital can remain high for years because mature advertising or software businesses subsidize new infrastructure. Segment disclosures, cloud margins, depreciation growth and utilization indicators help isolate whether recent AI investment is earning an adequate return.

A useful warning sign would be capital employed rising rapidly while incremental revenue slows and depreciation accelerates. A reassuring pattern would be new capacity entering service, backlog converting to cash and free cash flow recovering even as absolute investment remains high.

The purpose of adjusted leverage is not to manufacture a larger debt figure. It is to understand how much future cash is fixed, how much remains discretionary, what assets stand behind the commitments and which assumptions must hold for shareholders to earn a return.

What Happens Next

The next phase of the AI capital cycle will be defined less by announced spending than by conversion. Facilities contracted in 2024 through 2026 will begin operating, moving lease liabilities onto balance sheets and increasing depreciation, interest-like lease costs, power expense and maintenance requirements. At the same time, purchase commitments will become delivered chips, servers and network equipment.

That transition makes quarterly reporting more informative. Investors will be able to compare the growth in installed assets with growth in cloud revenue, advertising performance and customer utilization. If revenue expands faster than the new cost base, concerns about hidden leverage should fade. If capex remains elevated while free cash flow weakens and backlog conversion slows, the footnote commitments will become harder to dismiss as harmless timing differences.

Financing structure will also evolve. The largest companies can reduce buybacks, issue additional bonds or sell more equity. Infrastructure funds and private credit are likely to finance more data-center development. Customers may be asked for larger prepayments or minimum-spend guarantees. Suppliers may continue to invest in the companies that buy their products.

The most important scheduled information will come from ordinary filings rather than dramatic announcements: additions to uncommenced leases, short-term purchase obligations, asset useful-life estimates, guarantee exposure, cash capital expenditure, and changes in remaining performance obligations. Those tables will show whether the industry is moving toward self-funded growth or becoming increasingly dependent on capital-market support.

Frequently Asked Questions

What is Big Tech hidden debt?

Big Tech hidden debt is a popular label for obligations that do not appear as conventional funded debt on the face of the balance sheet. The category commonly includes leases that have been signed but not commenced, long-term purchase and construction commitments, guarantees, credit backstops and exposures through joint ventures. The label is imprecise because many of those items are disclosed and are not debt under accounting rules.

Is the $1.65 trillion figure accurate?

It is best treated as an estimate of combined future commitments rather than an exact debt balance. The calculation aggregates different categories reported under different definitions and, in some reconstructions, at different dates. It is useful for showing scale but should not be compared directly with a standardized balance-sheet debt figure.

Which companies are included in the $1.65 trillion estimate?

The estimate covers Alphabet, Microsoft, Amazon, Meta Platforms and Oracle. Nvidia is central to the broader AI financing story but was not one of the five companies in the reported aggregate.

Why are future data-center leases not recorded immediately?

Under U.S. lease accounting, a lessee generally recognizes a right-of-use asset and lease liability when the lease commences and the asset is available for use. A signed lease for a data center still under construction is disclosed if material, but it may not be recognized on the balance sheet until the facility is delivered.

Are purchase commitments the same as borrowing?

No. Borrowing provides cash today in exchange for repayment and interest. A purchase commitment is an agreement to buy goods or services, such as chips, energy or cloud capacity. The commitment can be economically binding and reduce future flexibility, but the company expects to receive an asset or service in return.

Does this mean Big Tech is committing accounting fraud?

The available public evidence does not support that conclusion. The principal obligations are described in SEC filings and related disclosures. Investors can reasonably criticize complexity, aggressive non-GAAP presentation or excessive risk without equating those practices with Enron’s fraudulent financial statements and concealed related-party losses.

Which company has the largest disclosed commitment?

Alphabet disclosed $811 billion of material purchase commitments and other contractual obligations at June 30, 2026, the largest single company figure discussed in this analysis. It should not be interpreted as $811 billion of funded debt, and it should not be added mechanically to every guarantee and lease number because categories can differ or overlap.

Why is Meta’s lease number important?

Meta reported $278.99 billion of uncommenced lease obligations at June 30 and then signed approximately $68 billion of additional data-center leases in July. The sequence shows how quickly obligations can grow between reporting dates and how future balance-sheet liabilities can be economically committed before recognition.

Why is Oracle considered riskier than the other hyperscalers?

Oracle entered the AI buildout with more recognized leverage and less cash generation than Alphabet, Microsoft, Amazon or Meta. Its backlog is large, but so are its future lease commitments. The credit question is whether contracted customer payments arrive in time and at sufficient margin to support construction, leases and debt service.

Is Nvidia financing demand for its own chips?

Nvidia has invested in AI companies and infrastructure funds, committed to cloud services and guaranteed certain facility leases. These activities can help customers and developers finance projects that use Nvidia systems. That resembles conventional vendor financing. The risk is that Nvidia may lose revenue, investment value and guarantee payments together if a supported customer fails.

What is the difference between capex and free cash flow?

Capital expenditure is cash spent to acquire long-lived assets. Free cash flow is commonly calculated as operating cash flow minus capital expenditure, although companies may use different definitions. During an infrastructure boom, capex can rise faster than operating cash flow, reducing free cash flow even when revenue and accounting profit continue to grow.

What should investors watch most closely?

The most useful indicators are operating cash flow relative to capex, the start dates for uncommenced leases, short-term purchase commitments, backlog conversion, customer concentration, guarantees, changes in depreciation lives, data-center utilization and credit spreads. Together they show whether capacity is becoming productive or merely more expensive.

Final Assessment

Big Tech’s hidden-debt controversy begins with a legitimate discovery and ends with an exaggerated analogy. The legitimate discovery is that the AI buildout has created future claims on cash far beyond the debt visible in ordinary balance-sheet summaries. The exaggeration is to treat every claim as borrowed money or evidence of an Enron-style concealment.

The filings are explicit enough to reject the fraud narrative. Alphabet disclosed $811 billion of contractual obligations. Meta disclosed hundreds of billions of leases and purchase commitments. Microsoft provided a quarterly trail showing uncommenced leases more than tripling in a year. Nvidia reported supply, cloud, investment and guarantee commitments. These are not invisible transactions.

They are difficult transactions. The numbers are spread across footnotes, measured on different bases and attached to assets that may not operate for years. That complexity reduces the usefulness of a simple debt ratio and creates room for optimistic presentation. Adjusted earnings can exclude real compensation costs. Longer asset lives can postpone depreciation. Leasing can reduce current capex while preserving long-term fixed payments. Backlog can look secure even when the customer depends on financing from the same ecosystem.

The strongest supporting interpretation is that the leading platforms are reserving scarce infrastructure for demand they can already see. Their cloud backlogs are large, their core businesses are profitable, and their access to capital is exceptional. The strongest concern is that the industry has committed to a scale of construction that requires not merely AI adoption, but sustained high-margin monetization across many companies whose forecasts cannot all be right at once.

The next evidence will not come from deciding whether $1.65 trillion is the perfect number. It will come from watching the obligations move through the financial statements: commitments becoming assets, assets becoming depreciation, backlog becoming revenue, and revenue becoming—or failing to become—free cash flow.

AI does not need to be fraudulent to produce losses. It only needs to earn less than investors and management teams assumed when they signed the contracts. The debt is mostly not hidden. The return remains uncertain.

This article is provided for general informational purposes and does not constitute financial, investment, tax, or legal advice.

Sources

• Nikkei Asia: Five US tech giants’ hidden debts soar to $1.65tn on opaque AI funding
• Alphabet Form 10-Q for the quarter ended June 30, 2026
• Reuters: Alphabet equity capital raise and AI spending plan
• Meta Platforms Form 10-Q for the quarter ended June 30, 2026
• Microsoft Form 10-K for the year ended June 30, 2026
• Reuters: Microsoft cloud growth, capital expenditure and free cash flow
• Amazon second-quarter 2026 earnings release filed with the SEC
• Reuters: Amazon second-quarter cloud growth and capital spending
• Oracle Form 10-K for the year ended May 31, 2026
• Oracle fiscal 2026 third-quarter financial results
• CreditSights analysis of Oracle lease commitments
• Financial Times: Oracle spending, debt and market reaction
• Nvidia Form 10-Q for the quarter ended April 26, 2026
• Reuters: Nvidia and the Hut 8 Texas data-center lease
• SEC rulemaking on disclosure of off-balance-sheet arrangements
• Federal Bureau of Investigation: Enron case history
• Bank for International Settlements Annual Economic Report 2026
• Goldman Sachs: Assumptions shaping the scale of the AI buildout
• JPMorgan Asset Management: How AI is being monetized
• Gartner worldwide AI spending forecast for 2026
• National Bureau of Economic Research: Global evidence on business use of AI
• Federal Reserve: The AI buildout and the economy
• Federal Reserve Bank of Kansas City: Industry evidence on AI and productivity
• Gusto: New business formation and AI use in 2026
• Ramp Economic Lab: AI spending per employee
• Bradford Cornell and Aswath Damodaran: The Big Market Delusion
• Richard Sloan: Do Stock Prices Fully Reflect Information in Accruals and Cash Flows About Future Earnings?
• Robert Bloomfield: The Incomplete Revelation Hypothesis and Financial Reporting
• Andrei Shleifer and Robert Vishny: The Limits of Arbitrage
• Mark Mitchell, Todd Pulvino and Erik Stafford: Limited Arbitrage in Equity Markets
• FINRA: Global Research Analyst Settlement retirement and Rule 2241

Affiliate disclosure: Businessfinance.news may earn compensation from qualifying actions completed through selected links on this website, at no additional cost to the reader. Affiliate relationships do not influence our editorial reporting, analysis, or conclusions.