The report, which analyzed the financial disclosures of Amazon, Meta, Alphabet, Microsoft, and Oracle, highlights how the unprecedented build-out of AI data centers is straining traditional accounting metrics. As of the end of 2025, these five tech giants had amassed a staggering $969 billion in total undiscounted future lease commitments, or data centers that have yet to be built. However, more than two-thirds of this total, that $662 billion figure, is for leases that have yet to commence, meaning that under generally accepted accounting principles, or GAAP, these companies are not required to recognize these massive obligations on their current balance sheets.

To put the scale of this hidden obligation into perspective, Moody’s accounting analysts David Gonzales and Alastair Drake calculated that the unrecorded $662 billion is equivalent to 113% of these five hyperscalers’ most recent adjusted debt.

Gonzales told Fortune in a statement that it’s “not as if [these hyperscalers] have have avoided a liability through structuring,” characterizing the $662 billion at issue as “yet to be on the balance sheet,” rather than missing. “More accurately,” he added, “they have not yet received the services to trigger this liability as of this time, but they will.”

For instance, Alphabet disclosed in a footnote in the third quarter of 2025 that it had “entered into leases primarily related to data centers that have not yet commenced with future lease payments of $42.6 billion, including a purchase option considered reasonably certain to be exercised, that are not yet recorded on our Consolidated Balance Sheet. These leases will commence between 2025 and 2031 with noncancelable lease terms between one and 25 years.”

That came after a Q2 disclosure stating Alphabet had “entered into leases primarily related to data centers that have not yet commenced with future lease payments of $23.9 billion, that are not yet recorded on our Consolidated Balance Sheets. These leases will commence between 2025 and 2031 with noncancelable lease terms between one and 25 years.”

From the second to the third quarter, future lease payments, an unrecorded future obligation, jumped from $23.9 billion to $42.6 billion.

Meta, Amazon, and Microsoft declined to comment. Alphabet and Oracle did not respond to requests for comment.

What is going on with these leases?

The root of this accounting phenomenon lies in the unique nature of AI hardware and the rules governing corporate leases. Historically, U.S. data center leases spanned 10 to 15 years. But because the cutting-edge semiconductor and technology equipment required for AI typically has a useful life of just four to six years, hyperscalers are demanding shorter initial lease terms with options to renew. And “to make the investment case for landlords,” the note explains, “these structures are often backstopped by a significant off-balance-sheet guarantee from the lessee.”

Under GAAP, a lease renewal period is only included in a company’s lease liability if the renewal is deemed “reasonably certain”—a high threshold requiring greater than 70% certainty. (GAAP emerged in the 1930s as a response to the 1929 stock market crash that coincided with, even caused, the onset of the Great Depression.) Because future AI strategies and technology upgrades are highly unpredictable, tech giants can justifiably argue that they are not reasonably certain to renew, thereby keeping the potential renewal costs off their books.

But landlords still require financial security to construct these specialized, multibillion-dollar facilities. To bridge this gap, hyperscalers are utilizing significant off-balance-sheet guarantees, most notably “residual value guarantees” (RVGs). An RVG acts as a backstop; if a tech company cancels or does not renew a lease, it must pay the landlord the difference if the data center’s market value falls below a pre-agreed threshold.

Crucially, U.S. GAAP allow companies to defer reporting these expected RVG obligations. Unless it is deemed “probable” that the lease will end without renewal, the contingent liability of the RVG does not need to be recorded on the balance sheet. Moody’s notes that “if a company concludes a lease renewal is likely to be exercised, but not reasonably certain, it can avoid classifying both the lease renewal periods and the residual value guarantee as liabilities.”

Meta Platforms provides a striking example of this practice. In its recent SEC filings, Meta disclosed entering into data center leases commencing in 2029 with an initial commitment of approximately $12.3 billion. Alongside this, Meta provided an RVG with a massive aggregate threshold of $28 billion. Because Meta deemed the RVG payouts “not probable,” no liability was recorded for this $28 billion guarantee.

In a chart book published nearly simultaneously with Moody’s report, Apollo Global Management chief economist Torsten Slok worked to put the enormity of data center spending into perspective. With total capital expenditure on data centers estimated at roughly $646 billion, or about 2% of U.S. GDP, Slok noted that is roughly equivalent to the GDP for Singapore, Sweden, and Argentina. Defense spending in 2025, meanwhile, was around $917 billion.

Moody’s warned that these opaque accounting practices mask the true economic risk facing the tech industry. While leasing reduces upfront capital investments, carrying such massive future commitments severely limits a company’s financial and operating flexibility, especially if AI industry conditions change rapidly. Because these liabilities are hidden, Moody’s concluded, in its own jargony way, that it is considering new ways to look at this issue.

“The accounting liability is unlikely to reflect certain plausible future scenarios … With this in mind, we will continue to assess cash exposures and debt-like adjustments as time progresses and the dates of new leases draw nearer. We may make a nonstandard adjustment to Moody’s adjusted debt based on our expectation of likely cash outflows.”