The balance sheets of tech giants may be far heavier than what their reported numbers suggest.
Goldman Sachs' latest research reveals that off-balance-sheet leasing commitments of AI hyperscale cloud providers such as Google, Microsoft, Amazon, and Meta have reached as much as $1.5 trillion, of which about $1 trillion has not yet commenced and thus is not reflected on their financial statements.
Meanwhile, another set of data compiled by Morgan Stanley shows that procurement commitments of these companies, as well as Nvidia and Oracle, are approaching a combined $982 billion.
Together, the two amount to nearly $2 trillion in potential undisclosed financial obligations. The core risk of these hidden liabilities is that, should data center investments fail to generate the expected returns, these contractual obligations will gradually surface, posing substantial pressure on the liquidity and leverage levels of the relevant companies.
Currently, the bond prices of hyperscale cloud providers have already seen a modest pullback, with the market increasing its focus on their long-term repayment capacity.
From traditional credit indicators, the financial condition of hyperscale cloud providers remains solid.
According to Morgan Stanley, the average net leverage of this cohort is only 0.5x, lower than the overall tech industry’s 0.8x, and even further below the US non-financial corporates’ average of 1.8x. More striking, these companies' total cash holdings still exceed their reported debt levels.
This forms the basis of the optimists' logic: core businesses continue to generate significant cash flow, debt size is relatively limited, and credit ratings are generally high. Even if data center returns fall short of expectations, equity investors will be hit first rather than creditors. This judgement is reasonably sound within the framework of traditional financial analysis.
However, as modes of financing have diversified, the picture of prudence painted by these indicators is being eroded by a series of commitments outside the balance sheets.
The precedent for this off-balance-sheet financing originated from the structured finance solution Meta designed for the "Hyperion" data center in Louisiana.
Meta and Blue Owl set up a joint venture called Beignet, with Blue Owl developing and owning the data center asset. Meta only holds a 20% equity stake in Beignet but commits to leasing Hyperion for at least 20 years.
It is this lease guarantee that enabled Beignet to issue amortizing bonds totaling $27 billion—a record-setting bond issuance.
Although Meta essentially assumes the repayment obligation, this debt does not appear on Meta’s own balance sheet.
This “have-it-both-ways” structure quickly sparked imitation among other hyperscale cloud providers, and a variety of distinct leasing arrangements have since emerged.
The reason these commitments can remain “invisible” lies in the specific stipulations of US Generally Accepted Accounting Principles (US GAAP).
Under GAAP ASC 842, lease payment obligations must only be recognized on the balance sheet once the lease actually commences, with a right-of-use asset recognized at the same time. Lease commitments yet to begin are only disclosed in the financial reporting footnotes.
Goldman Sachs analysts combed through these footnotes to tally around $1 trillion in “not yet commenced” lease commitments. This figure has surged from around $200 billion in total lease commitments five years ago and has further increased from Goldman’s own estimate of $750 billion last month.
Notably, Fitch and Moody’s do not include uncommenced leases in their rating metrics, while only S&P adopts a more conservative approach by taking into account those with “substantive debt characteristics.”
Goldman’s analysts are candid: from a credit perspective, the current treatment may understate corporate leverage and future liquidity needs, since these obligations will eventually be recognized and the contractual payments will gradually come due.
In addition to leasing, hyperscale cloud providers have undertaken substantial procurement commitments for computing power, chips, equipment, and power supply.
These commitments also do not present themselves as traditional debt on the balance sheet, yet they constitute substantial financial obligations that must be fulfilled in the future.
Morgan Stanley’s credit analysts show that as of the end of Q1 this year, Google, Microsoft, Amazon, Nvidia, and Oracle together have procurement commitments totaling $982 billion.
Added to the $1 trillion in leasing commitments not on financial reports per Goldman’s statistics, the total potential hidden financial obligations approach $2 trillion.
These concerns have already begun to show up in the bond markets.
Take Beignet bonds, for example: their yield has risen from the 5.65% low at issuance last autumn to roughly 6.95% now, with prices falling noticeably.
However, analysts point out that part of the pressure is a passive transmission from the general rise in US Treasury yields, and it does not yet amount to a strong signal of credit deterioration.
But the core logic of the market is increasingly clear: whether these hidden burdens can be digested ultimately depends on whether data center capital expenditures deliver the expected returns.
Currently, hyperscale cloud providers remain in a phase of large-scale expansion, and the visibility of investment returns is limited. For debt investors, how to incorporate these nearly $2 trillion in potential obligations into their valuation frameworks, outside of traditional metrics, has become an unavoidable and pressing question.