The $2.6 Trillion Hidden in Hyperscaler Footnotes

Alphabet, Amazon, Meta, Microsoft, and Oracle report roughly $700 billion in combined on-balance sheet debt. Their legally binding, non-cancellable obligations, the ones buried in the footnotes of their filings, exceed $2.6 trillion. That gap is the story.

It matters right now because the reckoning is no longer hypothetical. Moody’s has escalated its warnings through 2026, S&P Global Ratings has projected that all six major hyperscalers will generate negative free operating cash flow in 2026 and 2027, and these off-balance sheet obligations are actively migrating onto reported balance sheets as data centres open their doors.

This piece gives you a working framework for locating, categorising, and interpreting those obligations in company filings. By the time you finish, you will know where to look in a 10-K, what the three categories of commitment actually mean, and how to assess real financial exposure instead of relying on the headline debt figure that most coverage stops at.

Why $700 billion in reported debt tells only a fraction of the story

By any conventional measure, these five companies look financially healthy. Combined reported debt of roughly $700 billion, inclusive of recognised leases, is modest against their revenues, their cash holdings, and their interest coverage. Interest costs consume only a minor slice of total revenues, and two of the five actually earn more on their cash than they pay on their borrowings, producing negative net interest expense.

The screening metrics reinforce that impression:

• Alphabet and Microsoft both hold negative net debt, meaning their cash exceeds total borrowings, a position that survives even Alphabet’s recent $25 billion bond issuance.
• Oracle is the outlier, carrying materially higher debt ratios relative to equity and income than the other four.
• The wider US information technology sector runs a debt-to-equity ratio of around 0.51 times, and companies in this cohort with positive net debt sit at or below that mark.

On these numbers, you would conclude the group is well-capitalised and comfortably able to service what it owes. That conclusion is correct, and it is exactly why the structure underneath it is so easy to miss.

Hyperscaler capital expenditure projections of $610 billion to $650 billion collectively in 2026 nearly double 2024 levels, and because these four firms account for roughly 17% of the S&P 500, the fixed-cost commitments behind that spending function as a macro-level variable rather than a company-specific one.

When “well-capitalised” obscures the real exposure

The danger is not that these companies are secretly fragile. They are genuinely well-capitalised, and that is what allows the off-balance sheet structure to pass every standard screen an analyst runs.

The reported debt figures do a good job of communicating creditworthiness to a general audience. But if you are trying to assess the true fixed-cost burden of AI infrastructure spending, they are the wrong starting point. The most significant commitments are accumulating in channels that GAAP accounting excludes from the balance sheet by design, which means the debt metrics analysts most often cite are pointing you at only a fraction of what has actually been committed.

The accounting rules that keep $2.6 trillion out of the ledger

Before treating this as a scandal, it helps to understand that the rules keeping these obligations off the balance sheet are internally consistent. They were not written to hide anything. Two standards do most of the work:

1. ASC 842 (leases): A right-of-use asset and a matching lease liability appear on the balance sheet only when the lease commences, meaning when the landlord actually makes the property available for use. Before that handover, even a signed, binding lease stays out of the reported figures.
2. ASC 440/470 (purchase commitments): These are treated as executory contracts, where neither side has yet performed. The buyer does not control the goods and the seller has not earned the payment, so no liability is recognised until the contract is fulfilled or becomes explicitly loss-making.

That logic holds up fine for a company signing a normal supply contract or a standard office lease. It strains badly when applied to a cohort signing trillion-dollar commitments years before the underlying data centres even exist.

This is why credible institutions are now treating these footnotes as debt equivalents. Forensic accounting firm New Constructs argues ASC 842 creates a loophole and explicitly adds uncommenced leases back into its standardised operating-lease debt calculations. Researchers at the Bank for International Settlements go further, characterising the combination of special purpose vehicles, long-term leases, and offtake agreements in stark terms.

Researchers at the Bank for International Settlements describe these arrangements as “shadow borrowing”: multi-year operating expenses that are economically identical to debt, channelling private credit into data-centre projects while bypassing the balance sheet.

Moody’s has reached a similar practical conclusion, treating long-term data centre leases as debt-equivalent liabilities in its own analysis.

The accounting distortions compound across categories: AI infrastructure accounting at the major hyperscalers also stretches GPU depreciation lives to five or six years against a two-to-three-year economic reality, flattering current-period margins while deferring infrastructure costs into future income statements that have not yet had to absorb them.

The takeaway for you is procedural. Because the leverage lives in the footnotes rather than the balance sheet, you have to read the commitments and contingencies footnote and the leases footnote to see it. Anyone relying on GAAP-reported debt alone is missing the exposure that is actually building.

Three categories, five companies, $2.6 trillion: the numbers broken down

The off-balance sheet total splits into three categories. Taken one at a time, each is large. Stacked together, they reframe what these companies have committed to.

The first and largest is non-cancellable contractual commitments, which total over $1.49 trillion across the five firms. These are legally enforceable purchase agreements, drawn from mid-2026 regulatory filings.

The non-cancellable commitments break into three sub-types:

• Data centre hardware: AI semiconductors, servers, and related components.
• Take-or-pay power contracts: Long-term electricity supply deals, structured so the buyer pays for a minimum volume regardless of actual use, with terms running from 2 to 26 years.
• Rented compute: Third-party computing capacity contracts.

One figure deserves a flag. Meta’s Q2 2026 Form 10-Q reports $278.99 billion in uncommenced leases, though some earlier tallies put the figure above $300 billion. The filing number is the one to use. And Amazon’s own filings state that many of its purchase obligations are “generally cancellable in full or in part,” which means its headline total may overstate genuinely irrevocable exposure.

The interpretive point sits with the largest number. Alphabet’s $811 billion in non-cancellable commitments, on its own, exceeds the entire combined on-balance sheet debt of all five companies. So when you read a headline about “Alphabet’s debt load,” understand that the figure being described is a small fraction of what the company is actually contractually locked into, which changes any downside stress test you might run.

Uncommenced leases: the obligations that will move onto the balance sheet

The second category, uncommenced lease obligations, totals over $1.09 trillion. These are not permanent footnote residents. Under ASC 842, each one migrates onto the reported balance sheet the moment its facility opens, and some of those start dates run all the way out to 2036.

Adding these amounts to existing borrowings would more than double the group’s current $700 billion in on-balance sheet debt. The third and smallest category, roughly $100 billion in funding and construction commitments, covers investment relationships and infrastructure partnerships, including contingent arrangements such as Microsoft’s investment relationship with OpenAI.

Moody’s has been tracking the migration closely. By July 2026 it counted roughly $1.2 trillion in total lease commitments, with over $820 billion not yet started, and warned that the trend risks “material deterioration” in credit profiles as those leases commence.

What rating agencies and forensic analysts are saying in 2026

What makes the institutional response worth taking seriously is not its volume but its convergence. Moody’s, S&P, and Fitch have arrived at overlapping concerns from three different analytical starting points.

• Moody’s approaches it through lease accounting. It found $969 billion in undiscounted future lease commitments at the end of 2025, with uncommenced leases alone equal to 113% of the group’s adjusted debt, then updated the total to around $1.2 trillion by July 2026 and flagged the risk of “material deterioration” in credit profiles.
• S&P approaches it through cash flow. It projects that AI infrastructure spending exceeding $1.3 trillion by 2027 will push all six large hyperscalers into negative free operating cash flow across 2026 and 2027, with credit quality “gradually weakening.”
• Fitch approaches it through construction and utilisation risk on the bonds funding these data centres.

Fitch has warned that investors funding AI data centre bonds have placed “blind faith” in guarantees from marquee hyperscaler tenants, noting that construction delays or under-utilisation could disrupt cash flows despite take-or-pay structures.

S&P’s timing projection is the one worth pinning to the wall.

S&P Global Ratings expects all six major hyperscalers to generate negative free operating cash flow in 2026 and 2027, with recovery not anticipated until 2029.

Oracle’s credit downgrade to BBB-minus in July 2026 provides the clearest live example of the rating agency concern materialising: S&P cited $55.66 billion in fiscal 2026 capex, negative free cash flow, and leverage heading toward the mid-4x range, translating the abstract risk of commitment overextension into a concrete balance sheet outcome.

When three major agencies independently reach the same credit concern from lease accounting, cash flow modelling, and bond structure, the off-balance sheet obligations stop being an accounting curiosity. They become a factor that is already shaping how institutional debt capital will be priced for these companies. Because rating assessments feed directly into borrowing costs and portfolio allocation decisions, the footnote obligations have already begun to influence the financial conditions these firms operate under.

Bear case, bull case, and the historical episodes that frame the stakes

Both sides of this debate have serious institutional backing, and neither collapses under scrutiny.

Two historical episodes give the debate its weight.

• The late-1990s telecom and fibre boom: Public telecom capex surged from $47 billion in 1995 to $121 billion in 2000 (roughly $213 billion in today’s money), with the industry devoting over $444 billion to capital spending and accumulating $306 billion in debt. Capacity swaps and Indefeasible Rights of Use contracts inflated reported earnings until demand fell short and financing dried up, leaving a decade of unused “dark fibre” and widespread bankruptcies.
• The early 2010s cloud commitment trap: Enterprises locked into multi-year, fixed cloud-spend deals that became restrictive liabilities when their business models shifted.

The role reversal is the striking detail. During the cloud era, the hyperscalers were the vendors collecting those locked-in commitments. Today, regarding data centre real estate and power, they are the locked-in tenants.

The telecom precedent is not a forecast of collapse. It is a precise map of how fixed-capacity commitments built on optimistic demand forecasts can turn from strategic assets into structural liabilities, and if you are evaluating positions in this cohort, you should know that map exists.

The structural funding gap runs deeper than the commitment totals alone suggest: hyperscaler cash flows among major players are approaching or exceeding 90% of operating cash flow according to Vital Knowledge research, with Barclays models pointing to more than $200 billion in debt issuance required to close the shortfall between 2026 and 2028.

What to watch in upcoming filings, and what the numbers need to do

You do not need to feel anxious about a risk you cannot quantify. Each of the three obligation categories has a specific home in the filings, and once you know where to look, every new quarter becomes a datapoint rather than a headline.

Where to find each category:

1. Non-cancellable purchase obligations: The commitments and contingencies footnote in the 10-K or 10-Q.
2. Uncommenced leases: The leases footnote, specifically the “not yet commenced” subsection.
3. Funding and construction commitments: The investing activities section of the cash flow statement and the related footnotes.

Then track three variables that will tell you whether the bull or bear case is winning:

1. Revenue growth against committed capacity costs: Is the AI revenue ramp keeping pace with the fixed obligations?
2. The rate at which uncommenced leases commence: Each one that starts moves obligations onto the reported balance sheet.
3. Any disclosed cancellations or restructurings: Particularly relevant for Amazon-style cancellable obligations.

If you monitor the lease commencement rate across quarterly filings, you will have earlier visibility into balance sheet inflation than anyone waiting for rating agency downgrades or equity analyst revisions. One caveat worth holding onto: Moody’s has criticised disclosure limitations in this cohort, so a standard read may not surface everything relevant.

S&P Global Ratings expects negative free operating cash flow through 2027, with recovery not anticipated until 2029. That is the horizon that frames the intervening period.

The overall calibration is simple. The $2.6 trillion in obligations is a worst-case fixed commitment if demand disappoints. The degree to which it becomes a real problem is a function of the AI revenue ramp, not the accounting rules.

This article is for informational purposes only and should not be considered financial advice. Investors should conduct their own research and consult with financial professionals before making investment decisions. Past performance does not guarantee future results. Financial projections are subject to market conditions and various risk factors.

Reading the footnotes is now part of the investment thesis

The gap between roughly $700 billion in reported debt and $2.6 trillion in total obligations, a ratio of about 3.7 to 1, is not a scandal. It is what happens when GAAP accounting, designed for ordinary contracts, meets a category of spending signed at a scale and speed the rules never anticipated.

That gap does not make these obligations uniformly dangerous, nor uniformly manageable. Their risk profile is contingent on AI revenue realisation, and you now have the tools to assess that contingency as new data lands each quarter.

The practical takeaway is the filing framework: the commitments and contingencies footnote, the “not yet commenced” leases subsection, and the cash flow investing activities. S&P’s 2029 cash flow recovery timeline gives you the horizon to benchmark against in the meantime.

In a market where five companies have committed sums equivalent to a meaningful fraction of US GDP, reading the footnotes is not specialist analysis. For any informed position in this cohort, it is the baseline, and it is where the investment thesis will either hold or break.