Capital

1.65 Trillion In AI Debt Never Reached A Balance Sheet

Five US tech giants carry 1.65 trillion dollars of AI-related contractual obligations that never appear as debt. The disclosure sits in a footnote, and it tells you what your next cloud renewal will cost.

CapitalBy Servola Tech Desk2026-07-216 min read

AI-assisted, edited by humans. | Editor-in-Chief: Leon Soliman | Editorial standards

1.65 Trillion In AI Debt Never Reached A Balance Sheet

Key takeaways

Five US vendors carry 1.65 trillion dollars, roughly 1.5 trillion euros, in AI-related contractual obligations that exceed the 1.35 trillion dollars of debt actually printed on their balance sheets, so the vendor you are negotiating with is more committed than its accounts show.
Oracle's off-balance-sheet obligation reached 273.3 billion dollars at the end of May 2026, more than thirty times its level four years earlier, and Meta's is around 420 billion dollars, close to triple its recorded debt.
Moody's puts the five hyperscalers' unrecorded data-centre lease commitments at 662 billion dollars, equal to 113 percent of their adjusted debt, and warns of a material deterioration in credit profiles if AI revenue does not keep pace.
The obligations stay off the books because a lease is only recognised once the building operates, and because the vendors judge renewal to be not reasonably certain, which is itself a statement about which capacity they expect to give back.
Before signing your next cloud or GPU renewal, read the lease-commitment note in the vendor's latest quarterly report rather than its balance sheet, and keep your own term shorter than the vendor's committed recovery window.

The number that was not in the debt line

Oracle's most recent quarterly report contains a figure that does not appear anywhere in its debt line. At the end of May 2026 the company had 273.3 billion dollars, roughly 253 billion euros, of contractual obligations tied to AI infrastructure. Four years earlier the same figure was less than a thirtieth of that. It is disclosed, it is real, and under the accounting standards it is not debt.

Nihon Keizai Shimbun went through the recent financial statements and related documents of Alphabet, Microsoft, Amazon, Meta and Oracle and added those obligations up. The total came to 1.65 trillion dollars, around 1.5 trillion euros, an eightfold rise in about four years. The five companies' actual recorded debt is 1.35 trillion dollars. The commitments that sit outside the balance sheet are now larger than the ones inside it.

Meta alone accounts for about 420 billion dollars of the total, close to three times its transparent debt. These are not projections or analyst estimates of future spending. They are signed contracts: data-centre leases, server orders, and graphics processing units bought but not yet delivered.

How a gigawatt of compute stays off the books

The mechanism is dull and that is why it works. A lease liability is recognised when the asset is handed over or begins operating, not when the contract is signed. A data centre under construction in 2026 that opens in 2028 produces two years of obligation with no corresponding line in the accounts. The same applies to GPU purchase contracts for units the vendor has ordered and not received. Until delivery, the commitment lives in the footnotes of the quarterly report.

There is a second judgement doing quiet work. Lease accounting only captures extension periods and residual value guarantees when renewal is reasonably certain. AI hardware cycles are short, so the vendors argue that renewal is not reasonably certain, and the extensions drop out of the recognised liability. That argument is legitimate. It is also an admission worth reading twice.

Moody's has put a number on the narrower slice. Future data-centre lease commitments across the same five companies come to 662 billion dollars, roughly 610 billion euros, and that unrecorded amount equals 113 percent of their most recent adjusted debt. The agency expects data-centre capital expenditure across the sector to reach 500 billion dollars in 2026 and warns that as these leases move onto balance sheets, adjusted debt rises and financial flexibility falls. Morgan Stanley flagged the growth in data-centre lease contracts as a major risk factor in an investor report. Two institutions with no shared incentive arrived at the same concern.

El Paso, and the structure that repeats

The mechanism is easiest to see in a single deal. BlackRock is selling more than 12 billion dollars, around 11 billion euros, of bonds to finance a Meta data-centre campus in El Paso, Texas. The debt is issued by a holding company for BlackRock's 80 percent stake in the project vehicle, Project Sopaipilla Holdings. Meta holds the other 20 percent and is the anchor tenant. JPMorgan and Morgan Stanley were mandated to arrange investor calls, with pricing expected within days.

Read the ownership split carefully. Meta gets the capacity and does not carry the construction cost. The bondholders carry the asset. Meta carries a lease. And that lease, until the campus operates, is a footnote.

This is not a loophole anyone is hiding. It is disclosed, audited and increasingly standard, which is precisely the problem for a buyer: the structure is invisible in the metric most procurement teams actually look at. A vendor comparison built on reported leverage will rank these companies in the wrong order.

Why a pre-committed cost becomes your price floor

The solvency question is not the interesting one for a European operator. These are large companies with real cash generation and the ratings agencies are describing risk, not distress. The consequence that reaches your budget is narrower and more certain: a contractual obligation that must be recovered sets a floor under the price it is recovered through.

Capacity that has been ordered and paid for gets sold whether or not demand arrives on schedule. That sounds like it should mean discounts, and in the short run it can. Over the length of a three-year enterprise agreement it means the opposite. A vendor with 1.65 trillion dollars of pre-committed spend has very little room to absorb a soft quarter through pricing, because the obligation does not soften with it. The flexibility that used to show up as aggressive renewal terms is already spent.

The second inference is more useful and almost nobody is drawing it. The vendors have formally judged that renewal of these leases is not reasonably certain. That judgement is made site by site. It is, in accounting language, a statement about which capacity they are prepared to walk away from. If you are choosing a region for a workload with a five-year life, the vendor's own lease-recognition treatment is a better signal of that site's expected longevity than any roadmap slide you will be shown.

Read the note, not the balance sheet

The practical change to your process is small and it takes about twenty minutes per vendor. Open the most recent quarterly report, go to the commitments and contingencies note, and find the undiscounted lease payments not yet commenced. That figure, not the debt line, is the vendor's real forward commitment. Compare it against the same vendor's disclosure a year earlier. The direction of travel matters more than the level.

Then set your own term against it. If a vendor is recovering obligations over a period longer than your contract, you are the flexible party in the relationship and should be paid for it in the terms. If your contract runs longer than the vendor's committed recovery window on the capacity serving you, you are exposed to a repricing you cannot see coming. Under IFRS 16 the same disclosure logic applies to the European entities you contract with, so the note is available on both sides of the Atlantic.

None of this requires a view on whether the AI build-out pays off. It requires only that you stop reading the wrong number. The debt line describes what these companies borrowed. The footnote describes what they promised. Your renewal is priced off the second one.

Frequently asked questions

Is off-balance-sheet debt a sign that these companies are hiding something?

No. The obligations are disclosed in the footnotes of quarterly reports, audited, and treated in line with accounting standards that recognise a lease only when the asset begins operating. The issue is not concealment but placement: the figure sits where most vendor comparisons never look, so buyers who screen on reported leverage rank vendors in the wrong order.

How large is the gap between recorded and unrecorded obligations?

For Alphabet, Microsoft, Amazon, Meta and Oracle combined, AI-related contractual obligations reach 1.65 trillion dollars against 1.35 trillion dollars of recorded debt. Moody's measures the narrower data-centre lease slice at 662 billion dollars, equal to 113 percent of the five companies' most recent adjusted debt.

What should appear in our next cloud negotiation because of this?

Two things. Ask the vendor for its undiscounted lease payments not yet commenced, from the latest quarterly report, and compare it with the prior year. And align contract length so that your term is not longer than the recovery window on the capacity serving you, because a pre-committed cost is recovered through price and leaves the vendor little room to hold terms flat.

The AI build-out will be judged on whether the revenue arrives. That verdict is years away and no procurement team needs it. What is available today is a disclosed number, sitting in a note, describing exactly how much a vendor has already promised to pay before it quotes you anything. Buyers who read it will negotiate against the real constraint. Buyers who read the balance sheet will negotiate against a summary of the past.

Off-balance-sheet debtData centresOracleMetaMoody'sCloud contracts

Even a Trillion-Dollar Lab Must Rent Its Compute

Anthropic, valued near a trillion dollars, is in talks to rent about 10 billion dollars of compute from Meta over two years. Why the shortage is structural and what owners should budget for.

4 min read

Infrastructure

The Nvidia Tax Just Found a Second Source

Meta's MTIA chip Iris enters production in September after clearing testing in six weeks. Why a hyperscaler's second source caps the price you pay for compute.

3 min read

Law & Regulation

The EA Buyout Now Hinges on Brussels

The largest take-private in history has cleared EA shareholders and US antitrust. The last real gate is the EU - and not classic merger review, but the Foreign Subsidies Regulation, facing its first blockbuster test.