AI's hidden debt load: The hyperscalers' off-balance-sheet trillions

One headline has been circulating in the financial press for weeks: the world's largest technology firms have piled up more than a trillion dollars in hidden debt that never appears on their balance sheets. The video examines the five hyperscalers building the data centers behind the boom: Alphabet, Amazon, Meta, Microsoft, and Oracle. Confusion follows naturally, since one headline cites one trillion while another talks about three. The author's task here is to collect every off-balance-sheet line item from the companies' official filings into a single spreadsheet and review the company and category breakdown together.

The setup comes first: the AI spending spree had already attracted plenty of doubt. OpenAI alone has taken on more than 600 billion dollars in spending commitments through 2030, and its circular financing arrangements with Nvidia and others are under scrutiny. Former money printers have turned into money furnaces. The newer worry is different: to fund all of this, the companies have started borrowing heavily, and some have moved toward complex ownership structures that appear designed to shift borrowings off the balance sheet.

Then comes the video's key qualifier: these obligations are only supposedly hidden. Their absence from the balance sheet is deliberate in places, yet they are not hard to locate, because every company discloses them in a dedicated commitments section of its filings. The author takes on the legwork of assembling the figures in one place. There is also a disclosure of personal bias: he holds positions across all five names, including direct ownership of Meta, Google, and Microsoft shares. The stated goal is not investment advice but making visible a risk that matters to nearly every investor through index funds.

The first stop is the debt that does sit on the balance sheet. Outside Oracle, large American tech firms were historically famous for spotless balance sheets: minimal borrowing and surplus cash. AI expenditure has pushed debt upward, but the picture stays calm. Alphabet's 25-billion-dollar bond sale after its latest filing is added conservatively to the table. The striking result: Alphabet and Microsoft show negative net debt, meaning they could wipe out borrowings with cash on hand if they chose. Where net debt is positive, ratios against profit and equity stay low; four of the five sit at or below the US information-technology sector average debt-to-equity ratio of 0.51. Interest coverage looks comfortable, and two firms even earn more interest than they pay. The single caveat: aggressive reinvestment leaves little spare cash for actually retiring debt.

Turning off the balance sheet: the trillion-plus figure in the headlines feeds on broadly four buckets, namely non-cancellable contract commitments, funding and construction commitments, leases not yet commenced, and the most disputed item, financial support and guarantees. The rest of the video opens up each bucket in turn.

The first bucket covers legal obligations to purchase goods or services down the road. Details differ by company, but the contents cluster under three headings: data-center components such as chips, long-term take-or-pay power purchase agreements that bill whether the electricity is used or not, and cloud capacity rented from third parties. The shared logic is supply assurance: locking in chip deliveries amid surging demand and power supply amid enormous consumption. These stay off the balance sheet for a plain accounting reason: the supplier has not delivered yet, so neither liability nor expense has materialized; they sit in the filing footnotes until the contracts are executed.

The sums are staggering: the five companies carry more than 1.5 trillion dollars in such commitments, overwhelmingly tied to AI data-center construction. Against the level of five years ago the stack has multiplied, so nearly all of the increase traces back to AI. Distribution is uneven: Alphabet leads by far at more than 800 billion, followed by Meta at roughly 350 billion. Measured against total assets, these two have effectively pledged the majority of their asset base to future spending.

The second bucket, funding and construction commitments, concerns spending tied to joint ventures and investment pledges rather than goods and services. Microsoft's funding obligations around its OpenAI stake are one example, some of them conditional on milestones. Amounts run smaller than the first bucket, near 100 billion dollars across the five names, yet still material.

The third bucket draws the most headline attention: leases signed but not yet commenced, kept in the footnotes under standard practice because their start dates lie ahead. Leases come in two forms, mortgage-like finance leases and operating leases that merely rent usage rights; most of these look like operating leases over data-center space. Analysts still treat operating leases as debt-like, since skipping a payment carries consequences close to default. Across the five firms the total passes one trillion dollars, a sum that would more than double the roughly 700 billion in debt and leases already on the books. In dollar terms Meta and Microsoft stand out with more than 300 billion each in future lease load.

The fourth bucket is where debt can fairly be called hidden: financial guarantees and backstops. Some companies have created variable-interest entities and special-purpose vehicles kept under 50 percent ownership, which then qualify for equity-method accounting. The vehicle's assets and liabilities never consolidate; the whole business appears as a single asset line. Meta appears to have used this structure to raise debt inside the vehicle while keeping it off its own books.

The concrete case is Hyperion in Louisiana, among the largest data-center projects on earth at an estimated cost above 50 billion dollars. The project sits with a vehicle named Beignet Investor, 20 percent owned by Meta. In October 2025 that vehicle borrowed 27.3 billion dollars in one of the biggest corporate bond sales ever recorded. The campus is being built for Meta alone, which becomes its sole tenant from 2029. Up to this point everything can look strategic: Meta avoids the landlord business, forms a separate operator, leases from it, and lets it carry the debt. The catch is that Meta guaranteed the full issue amount through a residual value guarantee to get the bonds sold. The debt belongs to the vehicle on paper, yet Meta stands behind every missed payment, and Meta's rent is the vehicle's only income for servicing the bonds.

The El Paso project repeats the pattern as a BlackRock joint venture: a vehicle called Sopaipilla Investor has sold 12.3 billion dollars in secured notes, again with Meta's support behind them. The video's verdict is blunt: although each detail sits in the filings, the arrangement reads as an elaborate device for making debt harder to find. Lease payments fund the vehicles' borrowings, and combined with the guarantees, the companies end up borrowing in their own names by a roundabout route.

So how bad is it really? The video's answer is calmer than the headlines: only two companies carry sizeable guarantees, Meta and Alphabet, in the tens of billions and below 100 billion. Oracle shows 3.3 billion of similar guarantees, and Meta's 14.7-billion contingent purchase commitment is counted alongside. Material against revenue and assets, yet payable out of less than a year of earnings if ever called. The trillion-dollar hidden-debt impression suggests all five names concealed hundreds of billions in guarantees each; the reality is narrower.

Two technical objections cover the larger buckets. First, these totals are not equivalent to debt: just as a household budget would not book future phone, rent, and subscription payments as money owed, these are contracted sums paid later in exchange for goods and services. Second, the figures are nominal rather than discounted to present value the way balance-sheet debt is. Meta's roughly 350-billion lease stack, for instance, aggregates contracts starting through 2036 with terms from one to thirty years; 350 billion due tomorrow and 350 billion spread over decades are different burdens. Maturity detail exists only in fragments, so the headline totals overstate the load in present-value terms.

A rough experiment still gets run: adding every bucket except guarantees onto reported debt visibly worsens the ratios against historical norms. The one mitigating factor is that elevated leverage reads as normal for classic data-center operators: pre-AI Equinix posts ratios comparable to these adjusted figures. Oracle screens as the weakest credit throughout. The closing argument holds regardless: AI data centers differ from classic ones, today's demand premium may fade, and the group is migrating from an asset-light to an asset-heavy model. Fixed payment promises raise risk; where flops like the Metaverse once allowed a pivot, these commitments look like a point of no return. Open questions remain: whether the pledges equal future capex or add to it, and whether revenue catches up in time. No answers yet; the trajectory bears watching.