NEW YORK, 16 AUG 2026 — The five largest cloud companies have disclosed lease obligations of roughly US$1.2 trillion. Of that, US$725 billion covers leases that have not yet commenced, which means they do not appear in the financial statements most people read.
That figure is the story. Not the size of the AI build-out, which is well covered, but how much of the obligation to pay for it currently sits outside the accounts.
The numbers
The companies are Meta, Microsoft, Alphabet, Amazon and Oracle. Goldman Sachs credit strategists led by Amanda Lynam put collective capital expenditure at US$405 billion in 2025, around US$750 billion for 2026, and roughly US$1.2 trillion in 2027.
Capital spending tripling in two years is remarkable on its own. The financing question is what changes its character.
What an uncommenced lease is
An uncommenced lease is a precise accounting term.
A company signs a lease for a data centre that does not exist yet, or has not been handed over. Under the accounting rules, the liability and the corresponding asset are recognised when the lease commences — when the company gets control of the space. Before that, the commitment is real, contractual and enforceable, and it is disclosed in the notes rather than carried on the balance sheet.
None of that is irregular. It is how lease accounting works, the amounts are disclosed, and any analyst who reads the notes can find them.
The problem is that many investors and automated screens do not read the notes. Standard leverage ratios and debt-to-equity calculations are based on the headline balance sheet, which does not yet include this US$725 billion obligation. Goldman's own framing is that this "can understate leverage and future liquidity needs as these obligations are eventually recognised and contractual payments come due."
Why the financing mix is the real shift
Until recently the AI build-out was funded out of operating cash flow. These are among the most profitable companies in history, and they were spending money they had already earned.
Projecting that roughly 35 per cent of 2027 capital expenditure will be debt-funded marks a fundamental change. Debt has to be serviced on a schedule that does not care whether the demand forecast was right, and it converts a discretionary spending decision into a fixed obligation.
Goldman's strategists were careful about the channel, noting that they do not expect all of it in traditional investment-grade corporate bonds, and that there is scope for the hyperscalers "to lean more into broader debt financing channels". Broader channels means private credit, structured facilities and vehicles that are less visible than a bond issue.
The figures do not fully reconcile
The numbers in circulation differ, so a note of caution is in order.
The Goldman analysis reported this week puts disclosed lease obligations at about US$1.2 trillion with US$725 billion uncommenced, across five hyperscalers. A separate analysis attributed to the Financial Times puts purchase commitments at close to US$1.5 trillion across six companies, adding Nvidia, and other coverage cites lease commitments nearer US$1.5 trillion with about US$1 trillion uncommenced.
Those are not the same measure. Purchase commitments and lease commitments are different obligations, the company sets differ, and the reporting dates differ. We have used the Goldman figures because we could read them directly, and we are flagging the variance rather than reconciling estimates that were never meant to agree.
The direction is not in dispute in any version.
What this means from here
The first consequence is immediate for anyone buying cloud capacity in this region.
Debt-funded capacity has to earn its service cost. A data centre built with retained earnings can be priced to win market share for years; the same building financed at scale on a schedule cannot. That pressure shows up as firmer pricing, shorter discounting windows and more insistence on committed-use contracts rather than on-demand rates.
Regional buyers signing multi-year cloud or GPU commitments in the next eighteen months are signing into that environment. The question worth asking a vendor is not what the rate is, but what the rate is contingent on, and what happens at renewal if their cost of capital moves.
The second consequence is about where the capacity lands. Southeast Asia has attracted a large volume of announced data centre investment, much of it from companies now financing through debt rather than cash. Announced capacity is a plan, and plans funded by borrowing are more sensitive to interest rates and demand revisions than plans funded from profits. A regional government counting on an announced project should treat the financing structure as part of the risk, not just the headline figure.
Reasons not to panic
This is not a hidden crisis, for two reasons.
The obligations are disclosed. They sit in the notes, Goldman found them by reading filings, and nothing here involves concealment. This is a question about which readers look where, not about what companies have told the market.
And these are firms with enormous cash generation and low existing leverage. A US$725 billion future obligation spread across Meta, Microsoft, Alphabet, Amazon and Oracle is large in absolute terms and modest against their combined operating cash flow. The risk is not that they cannot pay. It is that the flexibility they have enjoyed — spending when it suits, pausing when it does not — is being converted into commitments that run on a schedule.
What we could not establish
The split of the US$725 billion by company, and the commencement dates. Both would show which firms are most committed and when the recognition lands, and neither was in the material we could read.
Also unestablished: how much of the projected debt is already arranged, the terms, whether any of it carries covenants tied to utilisation, how much sits in private credit rather than public markets, and what proportion of the leased capacity is already contracted to customers rather than built speculatively. That last point — how much capacity is pre-sold versus speculative — is the most important and the most opaque.
What to watch
The first signal is the next round of quarterly filings, specifically whether uncommenced lease balances keep growing at this rate. A sharp deceleration would say the companies themselves have started pricing in demand risk.
Then there is where the borrowing actually happens. Investment-grade bonds are visible and priced in public; private credit is neither. A shift toward the latter would make this harder to track precisely as the amounts grow.
Finally, watch for cloud providers in this region pushing for firmer committed-use terms. That is the mechanism by which a financing decision in Seattle or Redmond reaches a procurement conversation in Singapore, and it will arrive as contract language before it arrives as a headline.
View author profile → · Editorial policy
