The escalating AI infrastructure arms race is quietly generating a massive and intricate web of off-balance-sheet risks that traditional financial statements largely fail to capture. These hidden obligations, spanning procurement deals, cloud capacity agreements, and leasing structures, represent a growing vulnerability for the world's largest technology companies.

According to a recent Morgan Stanley research report, hyperscale cloud providers including Alphabet (NASDAQ: GOOGL), Meta Platforms, Inc. (NASDAQ: META), Amazon.com (NASDAQ: AMZN), Microsoft (NASDAQ: MSFT), and Oracle (NYSE: ORCL), alongside chip giant NVIDIA (NASDAQ: NVDA), have accumulated over $1.3 trillion in long-term commitments. This figure includes more than $640 billion in purchase obligations and roughly $675 billion in leasing commitments. A significant portion of these commitments remains off the balance sheet, obscuring the true leverage of these corporations and creating a risk profile where off-balance-sheet growth far outpaces the expansion of on-book debt.

From a cash flow perspective, the scale of these obligations is staggering. For Meta, combined leasing and procurement commitments equal roughly 1.7 times its projected operating cash flow for the next twelve months. More alarmingly, for Oracle, this ratio surpasses seven times. These commitments are often deeply intertwined with infrastructure financing, leaving limited room for renegotiation if AI demand projections take a downturn. If computational demand continues to outpace supply, these contracts transform into competitive advantages for securing scarce resources. However, should demand expectations falter, many of these promises are already embedded within developer agreements, supplier contracts, private credit arrangements, and public bond structures, making exit or renegotiation extremely difficult.

Procurement Commitments: A Double-Edged Sword Locking Supply Chains

Procurement commitments represent the most direct form of off-balance-sheet pressure. The AI data center expansion has driven a dramatic surge in purchase obligations for hyperscalers and NVIDIA. Morgan Stanley estimates that the six aforementioned companies now hold over $644 billion in disclosed procurement commitments, more than doubling year-over-year and representing a sixfold increase over the past five years. These obligations cover GPUs, memory, land, electricity, data center shell leases, and other infrastructure components.

Under US GAAP and SEC regulations, companies must disclose "unconditional purchase obligations"—irrevocable payment commitments for fixed or minimum quantities of goods or services. However, these obligations typically remain off the balance sheet until the goods or services are actually delivered, appearing only as footnote disclosures. NVIDIA presents a particularly notable case. To secure wafer fabrication capacity and memory supply, the company has significantly advanced its procurement timeline. As of January 2026, its inventory and related purchase commitments have risen to approximately 32% of market expectations for FY27 revenue, up from historical levels of 15% to 20%. Morgan Stanley highlights that this strategy positions NVIDIA favorably to meet demand, but it also amplifies downside risk should demand decline.

Cloud Capacity Arrangements: Vast Obligations Residing Outside the Ledger

Cloud capacity arrangements form another critical structural component. In these deals, suppliers commit to providing computational power over a specified period, while customers commit to minimum spending levels. As long as contracts do not specify particular GPUs or racks and customers cannot control the use of specific assets, many computational contracts avoid being classified as leases. This means minimum payment obligations may not create lease liabilities, with related costs only entering the income statement or forming accounts payable when customers actually utilize the capacity and receive billing.

In disclosures, NVIDIA holds $27 billion in cloud service agreement commitments, while Oracle has $10 billion in cloud capacity arrangements. Meta has not quantified its exposure but notes that its $131 billion in contractual commitments includes third-party cloud capacity arrangements. Take-or-pay contracts are also prevalent in this category, requiring buyers to either purchase minimum service volumes or pay minimum amounts. Suppliers can leverage these contracts to support infrastructure financing, while customers sacrifice flexibility.

Take-or-pay contracts are the quintessential form of such arrangements, obligating buyers to pay minimum amounts regardless of actual usage. The agreement between Microsoft (NASDAQ: MSFT) and CoreWeave (NASDAQ: CRWV) serves as a primary example—CRWV discloses that the majority of its revenue stems from multi-year take-or-pay contracts, which serve as underlying assets supporting its data center financing debt. Once demand expectations shift, these contracts become exceptionally difficult to renegotiate due to their embedded third-party financing structures.

Leasing Commitments: $675 Billion Pending Unrecorded Entry

Leasing itself is not a novel concept; both operating and finance leases enter the balance sheet upon commencement. As of the latest disclosures, MSFT, ORCL, META, AMZN, and GOOGL collectively hold approximately $82 billion in finance lease liabilities and $175 billion in operating lease liabilities on their books.

The truly substantial exposure lies in unstarted leases, amounting to roughly $675 billion—a dramatic increase of $435 billion from approximately $240 billion a year earlier. Many of these are long-term data center shell leases, signed to support developer financing but remaining off-balance-sheet until lease commencement. As data centers are delivered, these obligations will progressively transition onto balance sheets. Among the companies, Oracle leads with $261 billion in unstarted lease commitments, followed by Microsoft at $155 billion, Meta at $104 billion, Amazon at $96 billion, and Alphabet at $59 billion. Alphabet has disclosed that its unstarted lease commitments will gradually enter the balance sheet through 2031.

Furthermore, accounting rules permit various lease payments to remain outside liability recognition. These include variable lease payments (such as electricity and maintenance costs), renewal options (only recognized when "reasonably certain" to be exercised), residual value guarantees, and third-party lease endorsements or guarantees. Variable lease payments now constitute 30% of Alphabet's total lease costs, approximately 25% for Meta, and over 10% for Amazon, with these percentages expected to rise as new data center leases commence.

Guarantees and Residual Value Assurances: Sequestering Risks in Contingency

Residual value guarantees commit lessees to ensuring an asset's end-of-lease value meets a specified threshold, with the lessee covering any shortfall. These function similarly to contingent liabilities, but only enter lease liabilities when payment is deemed probable. Such structures have appeared in transactions involving Meta and Blue Owl Capital.

In the realm of third-party lease endorsements, Alphabet stands out significantly. Morgan Stanley reports that Alphabet has provided approximately $17 billion in lease endorsements to bitcoin mining companies developing data centers, including four under-construction projects with Cipher, Hut 8, TeraWulf, and Flash Compute, collectively covering over 1,000 megawatts of computational capacity. These endorsements typically match the scale of project-level construction debt, become effective at lease commencement, and decline with amortization.

Under US GAAP, these endorsements are generally not recognized on the balance sheet until the probability of payment reaches the "probable" threshold. Rating agencies, however, adopt more cautious treatment. S&P has indicated it will adjust Alphabet's debt once endorsements become effective at lease commencement. In its December credit commentary, S&P noted that when contingent guarantees play a critical role in counterparty financing, it may adjust debt on a net-of-tax basis—for instance, a $100 contingent guarantee corresponding to approximately $79 in debt adjustment.

Power Purchase Agreements: Billions in Hidden Cash Outflows

Data centers require more than just GPUs—they demand substantial electricity. To meet the power needs of AI facilities, hyperscale cloud providers are aggressively signing long-term power purchase agreements, sometimes extending up to 20 years. These agreements typically lock in electricity supply at fixed prices, enabling energy producers to recover costs and achieve reasonable returns while providing credit support for related infrastructure financing.

Notable examples include: Microsoft's 20-year PPA with Constellation Energy supporting the restart of Three Mile Island nuclear plant; Alphabet's partnership with NextEra Energy to reactivate Iowa's Duane Arnold nuclear facility; and Meta's 20-year agreement with Vistra Corp covering over 2,600 megawatts of zero-carbon electricity. Morgan Stanley's utility analyst David Arcaro estimates that the single PPA between Meta and Vistra carries a total cost of $7 billion to $8.6 billion over its 20-year term, translating to $350 million to $430 million annually.

From an accounting perspective, PPAs may be classified as leases, standard purchase commitments, or derivatives depending on contract terms, leading to inconsistent disclosure practices. Alphabet disclosed a 20-year, $9.9 billion PPA signed in January 2026 to be treated as a lease. Meta, conversely, indicates that some of its PPAs lack fixed or minimum usage commitments, rendering reliable disclosure of future liabilities impossible. Morgan Stanley emphasizes that regardless of final accounting treatment, all commitments remain off balance sheet until electricity is actually delivered or the lease commences, creating substantial opaque future cash outflows.

SPVs Are Not Magic: Most Chip Financing Ultimately Becomes Debt Risk

High-grade cloud providers can purchase chips with cash and public debt. In contrast, high-yield-rated or unrated AI infrastructure companies increasingly rely on asset-backed structures. Common approaches include: supporting SPV loans with take-or-pay computational contracts, customer prepayments, chip leasing through SPVs, and other supply chain financing or securitization arrangements.

The critical factor is not whether an SPV exists, but who bears the risk. If SPV debt carries unconditional parent guarantees, that debt still enters the parent's balance sheet. When recourse is limited and long-term offtake contracts from high-credit customers provide support, off-balance-sheet treatment becomes more achievable—yet rating agencies may still view such arrangements as debt-like burdens due to economic dependency and potential parental support.

Customer prepayments can alleviate short-term cash flow pressure for GPU-as-a-service companies, potentially appearing as deferred revenue rather than direct debt, though they factor into credit quality assessments. Chip leasing operates more directly: leases enter the balance sheet upon commencement but remain off-balance-sheet beforehand.

This represents the most critical area to monitor in the AI capital cycle: money has already been committed, yet risks are dispersed across purchase obligations, unstarted leases, variable payments, PPAs, guarantees, SPVs, and rating adjustments. Evaluating AI companies requires looking beyond balance sheet debt alone. In the coming years, the numbers buried in footnotes may signal to the market—earlier than debt line items ever could—which companies have placed the heaviest bets on AI demand.