AI Capital Bureau: A Detailed Analysis of “Long-term Commitments” -- The “Invisible Leverage” of Nvidia and Cloud Giants

The AI infrastructure arms race is spawning a vast and complex off-balance sheet risk exposure, which is almost undetectable in traditional financial statements.

According to FOMO Trading Desk, Morgan Stanley's latest research report points out that hyperscale cloud vendors and chip giants such as Google (GOOGL), Meta, Amazon (AMZN), Microsoft (MSFT), Oracle (ORCL), and Nvidia (NVDA) have cumulatively formed more than $1.3 trillion in long-term commitments—of which purchase commitments exceed $640 billion and lease commitments are about $675 billion. A large portion of these commitments is off the balance sheet, making it difficult for investors to accurately assess the true leverage of these companies. Off-balance sheet leverage is growing much faster than on-balance sheet leverage for these companies.

From a cash flow pressure perspective, the scale of commitments by some companies is already eye-popping: Meta’s lease and purchase commitments are about 1.7 times its projected operating cash flow for the next 12 months, while Oracle's ratio exceeds 7 times. These commitments are often deeply tied to infrastructure financing, and once AI demand expectations reverse, there is very limited room for renegotiation.

If compute demand continues to outpace supply, these contracts will become a competitive advantage for securing scarce resources. But if demand expectations are revised downwards, many commitments are already embedded in the financing structures of developers, suppliers, private credit, and public bonds, making renegotiation or exit not necessarily easy.

Purchase Commitments: Six-fold Growth in Five Years, A Double-Edged Sword for Locking the Supply Chain

Purchase commitments are the most direct off-balance sheet pressure, with the AI data center expansion wave driving surging purchase commitments by hyperscale cloud vendors and Nvidia. According to Morgan Stanley, the six companies mentioned above have disclosed a total of over $644 billion in purchase commitments, more than doubling from a year ago and increasing roughly sixfold over five years. These commitments cover GPUs, memory, land, electricity, shell leases for data centers, and other infrastructure components.

According to US GAAP and SEC rules, companies must disclose "unconditional purchase obligations," i.e., irrevocable payment commitments for fixed or minimum quantities of goods or services. However, before delivery of the goods or services, these obligations are generally not recognized on the balance sheet and are only presented in footnotes.

Nvidia’s situation is particularly noteworthy. To secure foundry capacity and memory supply, Nvidia has significantly increased advance procurement. By January 2026, its inventory and related purchase commitments have risen to approximately 32% of market-expected FY27 revenue, compared to a historical range of just 15% to 20%. Morgan Stanley highlights that this strategy places Nvidia in a favorable position to meet demand, but if demand falls, the downside risk will also be magnified accordingly.

Compute Capacity Arrangements: Large Commitments Remain Off-Balance Sheet

Cloud capacity arrangements are another key structure. Suppliers commit to providing compute power over a certain period, while customers commit to minimum spending. As long as a contract does not specify particular GPUs or racks or grant the customer control over specific assets, many compute contracts are not treated as leases.

This means minimum payments may not form lease liabilities. Only when the customer actually uses the capacity and is billed do the related costs enter the income statement or accounts payable.

In disclosures, Nvidia has $27 billion in cloud service agreement commitments, Oracle has $10 billion in cloud capacity arrangements; Meta has not quantified, but states that its $131 billion in contract commitments includes third-party cloud capacity arrangements. Take-or-pay contracts are also in this category: the buyer must either purchase a minimum amount of services or pay a minimum amount. Suppliers can use such contracts to support infrastructure financing, while customers sacrifice exit flexibility.

Take-or-pay contracts are the classic form of this arrangement, requiring the buyer to pay a minimum amount regardless of actual usage. The contract between Microsoft (MSFT) and CoreWeave (CRWV) is one example—CRWV discloses that most of its revenue comes from multi-year take-or-pay contracts, which are used as underlying assets for data center financing debt. This means that once demand expectations change, these contracts are extremely difficult to renegotiate because they're already embedded in third-party financing structures.

Lease Commitments: $675 Billion Suspended Off-Balance Sheet

Leasing itself is nothing new. Operating and finance leases enter the balance sheet after commencement. As of the latest disclosure, MSFT, ORCL, META, AMZN, and GOOGL collectively hold about $82 billion in finance lease liabilities and $175 billion in operating lease liabilities.

But the real giant is uncommenced leases: about $675 billion, an increase of $435 billion from about $240 billion a year ago. Many are long-term shell leases for data centers, used at contract signing to support developer financing, but still off-balance sheet before the lease begins. As data centers are delivered, these will gradually enter the balance sheet.

Among them, Oracle has uncommenced lease commitments of $261 billion, Microsoft $155 billion, Meta $104 billion, Amazon $96 billion, and Google $59 billion. Companies such as Google disclose that these lease commitments will gradually be included on the balance sheet by 2031.

Moreover, accounting rules allow many types of lease payments to remain off-balance sheet, including: variable lease payments (such as electricity, maintenance fees, etc.), renewal options (included only when "reasonably certain" to be exercised), residual value guarantees (RVG), and third-party lease backing and guarantees. For example, variable lease payments made up 30% of Google’s total leasing cost, 25% for Meta, and more than 10% for Amazon, and this will continue to rise as new data center leases begin.

Guarantees and Residual Value Guarantees: Storing Risks in the "May Occur" Drawer

Residual value guarantees mean tenants commit that the leased asset’s value won’t fall below a certain level at the end of the lease; if it does, the tenant compensates for the difference. This is similar to a contingent liability but is only included as a lease liability when payment is deemed probable. This structure has appeared in Meta's transaction with Blue Owl Capital.

In terms of third-party lease backing, Google (Alphabet) stands out in scale. According to Morgan Stanley, Google has currently provided about $17 billion in lease backings to bitcoin mining companies developing data centers, including Cipher, Hut 8, TeraWulf, and Flash Compute—covering more than 1,000 megawatts in aggregate capacity. These guarantees usually match project-level construction debt, take effect when the lease starts, and decrease gradually as amortization occurs.

According to US GAAP, these guarantees are typically not recognized on the balance sheet until payment is "probable." However, rating agencies are more cautious: S&P has stated that once a guarantee takes effect (i.e. when the lease starts), it will adjust Google’s debt; in a December Q&A last year, S&P said when contingent guarantees play a key role in counterparty financing, it may adjust debt by the net after-tax amount (for example, $100 contingent guarantee would adjust debt by approximately $79).

Power Purchase Agreements: Billions in Hidden Cash Outflows

Data centers are not just short on GPUs, but also electricity.

To meet the power demand of AI data centers, hyperscale cloud vendors are signing large-scale long-term Power Purchase Agreements (PPAs), sometimes with terms as long as 20 years. These agreements lock in electricity supply at fixed prices to ensure energy producers can recover costs and earn reasonable returns, while also providing credit support for related infrastructure financing.

Typical cases include: Microsoft signing a 20-year PPA with Constellation Energy to support the restart of the Three Mile Island nuclear plant; Google working with NextEra Energy to restart the Duane Arnold nuclear plant in Iowa; and Meta signing a 20-year agreement with Vistra Corp, involving over 2,600 megawatts of zero-carbon power. According to Morgan Stanley utility analyst David Arcaro, Meta’s single PPA with Vistra will cost approximately $700 million to $860 million over 20 years, or roughly $35 to $43 million per year.

In accounting, PPAs may be classified as leases, normal purchase commitments, or derivatives, depending on contract terms, leading to mixed disclosure approaches. Google disclosed a 20-year, $990 million PPA signed in January 2026 that will be treated as a lease; Meta states that some PPAs cannot reliably disclose future liability amounts due to the absence of fixed or minimum usage commitments. Morgan Stanley points out that regardless of final accounting treatment, until power is delivered or the lease commences, these commitments are not recognized on the balance sheet, resulting in massive opaque future cash outflows.

High-rated cloud companies can buy chips using cash and public debt; high-yield or unrated AI infrastructure companies rely more on asset-backed structures. Common methods include: using take-or-pay compute contracts to support SPV loans, customer prepayments, chip leasing through SPVs, and other supply chain financing or securitization arrangements.

The key is not whether there is an SPV, but who bears the risk. If the SPV debt is unconditionally guaranteed by the parent company, it still goes on the parent company's balance sheet. If recourse is limited and supported by high-credit long-term offtake contracts, it's more likely to be kept off the balance sheet for accounting, but rating agencies may still treat it as a debt-like burden due to economic reliance and potential parent company support.

Customer prepayments can reduce short-term cash flow pressures for GPUaaS companies. These may show up on the books as deferred revenue, not directly as debt, but are considered in credit quality assessments. Chip leasing is more direct: after the lease starts, it goes on the balance sheet; before it starts, it remains off-balance sheet.

This is where the AI capital cycle needs most attention: money is already committed, but risks are scattered across purchase obligations, uncommenced leases, variable payments, PPAs, guarantees, SPVs, and rating adjustments. When analyzing AI companies, don't just look at on-balance sheet debt. In the coming years, the numbers in the footnotes may alert the market to who is betting most heavily on AI demand, before debt line items do.