The Hidden Debt of Tech Giants in AI: $1.65 Trillion Off-Balance Sheet
Behind the official accounts of Alphabet, Microsoft, Amazon, Meta, and Oracle lies a second debt mass, larger than the declared one. A study by the Japanese economic daily Nikkei quantified this "invisible" exposure at about $1.65 trillion, a figure that has grown nearly eightfold in four years and exceeds the $1.35 trillion in debt that the five companies regularly report in their balance sheets.
The technique behind this accounting is far from new. Companies transfer the cost of chips, servers, and energy infrastructure to separate legal entities, often joint ventures, so that the burden does not appear directly in their accounts. According to current accounting rules, GPUs and servers ordered with long-term contracts but not yet delivered, as well as leasing agreements for data centers that are not yet operational, can be treated as off-balance sheet items.
The most cited case involves Meta's Hyperion data center in Louisiana. The Menlo Park company and investment firm Blue Owl Capital contributed equity into a separate structure that has incurred $27 billion in debt. Meta is the sole tenant of the facility but claims it does not have to account for that debt because it is not the entity responsible for finding new tenants in case of issues. A week ago, the company announced that the total investment in the site will exceed $50 billion, having also signed a guarantee to cover investor losses if the lease agreement is terminated.
Meta's off-balance sheet debt amounts to about $420 billion, nearly three times that which is actually recorded. Oracle, engaged in the Stargate project with OpenAI through leasing agreements with external operators, saw its hidden exposure rise to $273.3 billion by the end of May, an increase of over thirty times in four years; this is in addition to $260 billion in future leasing commitments that will eventually flow into the official balance sheets. NVIDIA, for its part, declares $119 billion in purchase obligations. Alphabet and Microsoft also maintain their own off-balance sheet structures, although with less publicized details.
The overall picture is linked to a common rush: the sector is expected to invest over $3 trillion by 2028 in the construction and equipping of data centers for artificial intelligence, much of this spending financed through leveraging the chips installed within the very plants.
The parallel with the collapse of Enron, the U.S. energy company that failed in 2001 due to hidden debts through a network of shell companies, is almost inevitable, although the contexts are profoundly different. As analyst Gil Luria noted in a statement to Bloomberg Law, the crime of Enron was not in using special purpose vehicles but in keeping them hidden. Today, thanks to stricter regulations and strengthened transparency obligations, the use of these instruments is fully legal: information is disclosed, albeit in the footnotes of quarterly balance sheets, accessible to anyone who wants to look deeper.
The fact remains that an increase in low-transparency joint ventures can still generate market concerns, even with proper accounting practices. Some observers have already begun to express caution. Morgan Stanley has dedicated an in-depth analysis to the phenomenon in a report to investors, while the rating agency Moody's, already in a report from February, had flagged the growth of commitments related to uninitiated leasing contracts. S&P has also cut Oracle's credit rating due to excessive leverage. In March, economists from the Bank for International Settlements defined this mechanism of raising funds from institutional investors without increasing on-balance sheet debt as "shadow borrowing," expressing fears regarding potential delays in data center projects and the possible contagion of AI-related anxiety on the entire market.
A Japanese auditing firm executive noted that there is growing concern that the actual financial burden on tech companies is significantly higher than what is visible in the official balance sheets. Accounting expert Tom Selling posed a direct question: what would happen if one of these companies were indeed a house of cards supported by this kind of accounting treatment?
The timing of the study is not coincidental. Four of the five companies analyzed will present quarterly results in the next two weeks, and the data could therefore further increase. In communications, official debt will likely appear under control, but the hidden figure in the footnotes may not be.
The critical knot regards what happens afterward. When a data center becomes operational, the related leasing contract is immediately recorded on the balance sheet. If the demand for artificial intelligence turns out to be lower than expected, the value of the facility would be revised downward, with losses falling on the involved financiers and insurers.
The companies involved assert that future revenues will more than cover these commitments. The backlogs publicly declared by Microsoft, Alphabet, and Amazon for cloud and other services amounted to about $1.45 trillion at the end of March, and the CEO of Amazon Web Services, Matt Garman, called the investments by his company "non-speculative."
An additional element of uncertainty concerns the nature of the demand sustaining today's AI industry, partly fueled by a circular investment mechanism: NVIDIA and large tech companies invest in data center operators and artificial intelligence companies, whose money then transforms into GPU purchases and revenues from cloud services. This intertwining makes it more difficult to assess the true size of the underlying demand, increasing the risk of overspending on data centers relative to the actual needs of the market.