Moody’s Flags an Accounting Gap in Big Tech’s AI Build Out

The AI arms race looks exciting from the outside. Inside the financial statements, it looks more like a stack of footnotes that could keep an audit team busy for weeks. Right now, five companies, Amazon, Meta, Alphabet, Microsoft, and Oracle, sit at the center of a $662 billion question. That number, flagged by Moody’s Ratings, represents future data center lease commitments that have not yet started and therefore do not appear as liabilities on their balance sheets. It is not illegal. It is not hidden in the sense of fraud. It is simply how GAAP works. But let’s not kid ourselves, this is real economic exposure. For accounting and finance professionals, this is where it gets interesting. 

Source: Fortune.com 

Are $662B in lease commitments really “off the books”?

Moody’s analyzed disclosures as of year-end 2025 and found that the top five hyperscalers had $969 billion in total undiscounted future data center lease commitments. Of that, $662 billion relates to leases that have not yet commenced. Under U.S. GAAP, a lease liability is recognized when the lease begins and the company gains control of the underlying asset. So, until the concrete is poured and the lease term starts, no liability hits the balance sheet. Moody’s calculated that the unrecorded $662 billion equals roughly 113% of these companies’ most recent adjusted debt. That comparison makes people sit up straight. 

Source: Fortune.com 

To be clear, as Moody’s analyst David Gonzales noted, these companies have not avoided a liability through clever structuring. They simply have not yet received the services that trigger recognition. The obligation will show up as the projects come online, many between 2025 and 2031. Still, from a credit and risk perspective, you would be kidding yourself if you ignored it. 

Why are AI data center leases structured this way?

Historically, U.S. data center leases ran 10 to 15 years. AI changed the math. The core hardware, GPUs and advanced semiconductor systems, often becomes obsolete within four to six years. Hyperscalers do not want to lock themselves into 15-year commitments on facilities built around rapidly evolving technology. So, they negotiate shorter initial terms, often four years, with multiple renewal options that can extend the lease 20 years or more. Here is where GAAP steps in. A renewal period only gets included in the lease liability if renewal is “reasonably certain,” generally interpreted as more than 70% likely. Given the pace of AI development, companies can reasonably argue that renewal is not reasonably certain. At the same time, landlords need comfort before investing billions in highly specialized facilities. That is where residual value guarantees, or RVGs, enter the picture. 

If the tenant does not renew and the data center’s market value falls below a set threshold, the tenant compensates the landlord for the shortfall. Under GAAP, that contingent obligation only gets recorded if payment is “probable,” typically over 50% likely. So, if management concludes renewal is likely but not reasonably certain, and termination is not probable, both the renewal term and the RVG can stay off the balance sheet. That is perfectly compliant. It also creates a gap between accounting presentation and economic reality. 

How big are these guarantees in real life?

Meta provides a clean example. In recent SEC filings, Meta disclosed leases commencing in 2029 with an initial commitment of about $12.3 billion. Alongside that, it provided a residual value guarantee with an aggregate threshold of $28 billion. Because Meta deemed RVG payments not probable, no liability was recorded for that guarantee. Alphabet’s disclosures tell a similar story. In Q2 2025, it disclosed $23.9 billion in data center leases not yet commenced. By Q3, that figure had jumped to $42.6 billion. These leases will begin between 2025 and 2031, with noncancelable terms ranging from one to 25 years. This is not small change. Apollo’s chief economist Torsten Slok estimated total capital expenditure on data centers at roughly $646 billion, around 2%of U.S. GDP. For context, U.S. defense spending in 2025 was about $917 billion. When AI infrastructure approaches defense-scale dollars, you pay attention. 

Now imagine you are a controller at a mid-cap tech firm watching this. Your audit committee already grills you on lease disclosures under ASC 842. You have to explain weighted average lease terms and discount rates. Now the largest companies in the world carry future commitments equal to more than their adjusted debt, largely in footnotes. You can bet credit analysts read those footnotes line by line. 

Why are rating agencies stepping in early?

The scale of AI infrastructure has overwhelmed traditional project finance channels. Data center developers now seek ratings even while facilities are still under construction. S&P, Moody’s, Fitch, and KBRA have expanded coverage to projects that are not yet complete. Fitch rated more than 35 data center projects in the past nine months, most tied to hyperscaler-backed facilities under construction, with average deal sizes around $3 billion. Two-thirds of the ratings issued so far sit at investment grade. S&P assigned an A+ rating to $27 billion of debt tied to Meta’s Hyperion data center in Louisiana. That rating relies heavily on long-term leases and financial commitments from Meta, including agreements to start paying rent even if construction faces delays. 

In many cases, the rating on the project is capped by the tenant’s credit rating. As one rating director put it, lenders are effectively taking risk on the hyperscaler, not the bricks and mortar. This dynamic matters. When rating agencies say they may apply “nonstandard adjustments” to adjusted debt to reflect likely cash outflows from lease renewals or RVGs, they signal that reported GAAP numbers may not drive credit decisions on their own. As Benjamin Graham wrote in The Intelligent Investor, accounting figures are the beginning of analysis, not the end. That line feels relevant right now. 

What does this mean for the rest of us?

• First, this is not Enron 2.0. The disclosures sit in the footnotes. The rules operate as written. The issue lies in how users interpret them. 
• Second, leverage looks different depending on which lens you use. A traditional debt-to-EBITDA metric might appear comfortable. Add hundreds of billions in future lease commitments, and the picture changes. 
• Third, flexibility matters. If AI demand stalls, or a new architecture renders today’s facilities less useful, companies still face enormous fixed commitments. Leasing reduces upfront capex but does not eliminate long-term exposure.

For practitioners, this raises practical questions. When you evaluate a client’s covenant compliance, do you factor in off-balance-sheet lease commitments? When advising boards on capital allocation, do you stress-test scenarios where renewal becomes likely? If you work in valuation, how do you treat contingent guarantees that are not recorded? This is where the rubber meets the road. 

Takeaway 

The AI build-out sits at the center of capital markets, credit analysis, and financial reporting. $662 billion in not-yet-commenced lease commitments will eventually migrate onto balance sheets as projects come online. In the meantime, rating agencies adjust their own metrics to capture what GAAP leaves in disclosure. The real story is not that companies hide liabilities. It is that accounting standards rely on probability thresholds that may not align perfectly with economic intuition in a rapidly evolving industry. For CPAs, controllers, auditors, and finance leaders, the lesson is simple. Read the footnotes. Model the scenarios. Do not rely solely on reported lease liabilities when assessing risk. AI infrastructure may power the next decade of growth. It also tests how well our accounting frameworks capture long-term commitments in a world moving at warp speed.