The taste of an AI version of the subprime mortgage crisis? With $1.8 trillion in off-balance-sheet exposure, it is becoming a ticking time bomb in this round of frenzy.

Amidst the frenzy of AI infrastructure construction, an unprecedented scale of debt expansion is quietly taking shape - and the most dangerous part of it has never appeared on any balance sheet. Goldman Sachs' latest report predicts that by 2027, capital expenditures for super-large-scale cloud computing enterprises will reach 11 trillion to 14 trillion US dollars, far exceeding market consensus. However, according to in-depth research by Morgan Stanley, this staggering figure that has startled the market is only the tip of the iceberg. Nearly $1 trillion in procurement commitments, over $800 billion in unexecuted lease contracts, and billions of supplier financing arrangements together constitute about $1.8 trillion in off-balance sheet exposure - these liabilities are off the balance sheet but effectively lock in future cash outflows. The market has not yet fully priced in these risks. Morgan Stanley warns that the leverage ratio of super-large-scale cloud companies has soared from 0.9 times to 1.8 times in just two quarters, with capital expenditure growth rates consistently outpacing revenue and free cash flow growth rates, and the real impact of depreciation pressure has yet to come. At the same time, private credit institutions represented by Apollo and Blackstone are transferring leverage to the supply chain level through Special Purpose Vehicles (SPVs), forming highly cyclical and opaque financing structures. If the commercialization of AI falls short of expectations, or if enterprise customers shift to low-cost alternatives on a large scale, the vulnerability of the entire financing chain will be exposed. Debt issuance frenzy: AI has become the biggest variable in the public market According to Morgan Stanley's latest "AI Debt Financing Tracking Report", as of the end of May 2026, the global issuance of AI-related bonds had reached $236 billion, a 357% increase compared to the same period in 2025. Morgan Stanley expects the total annual issuance of AI debt to exceed $570 billion, and in the second half of the year, as the financing needs of capital expenditures become more concentrated, the pace of issuance will further accelerate. In April, monthly AI-related bond issuances exceeded $74 billion, reaching a new high for the year, with project financing structures (for data center construction) accounting for 85% of high-yield bond supply and 40% of investment-grade bond supply. At the same time, the five super-large-scale cloud companies - Amazon, Meta, Google, Microsoft, and Oracle - now account for 4% of the entire investment-grade bond index. At the leverage level, the overall gross leverage ratio of super-large-scale cloud companies has risen from 0.9 times in the third quarter of 2025 to the current 1.8 times, increasing by about 0.3 times per quarter, surpassing the leverage level of the entire energy industry. Morgan Stanley points out that due to supply pressure, credit spreads have drifted from the AA range to the A range, and may widen further. Meta's credit spread is currently wider than the CDX IG benchmark. At the free cash flow level, Morgan Stanley predicts that Amazon and Meta's free cash flow in 2026 will approach zero or even turn negative, and incremental financing will rely almost entirely on new debt. $1.8 trillion off-balance sheet exposure: invisible liabilities, locked cash outflows Todd Castagno, Global Valuation, Accounting, and Tax team at Morgan Stanley, pointed out in the report that focusing only on capital expenditure numbers severely underestimates the true financial commitments of the AI construction cycle. In addition to the disclosed capital expenditures, there are three key off-balance sheet exposures: Purchase commitments of approximately $982 billion. The total value of long-term purchase contracts for super-large-scale cloud companies and Nvidia approaches nearly $1 trillion. According to accounting standards, these obligations are not recognized as liabilities until the goods are delivered unless the company expects the contract to be loss-making. Therefore, nearly $1 trillion in future cash outflows does not currently appear as liabilities on any balance sheet. It is worth noting that Nvidia's own inventory and purchasing obligations have risen to approximately 32% of consensus revenue forecast for the 2027 fiscal year, significantly higher than the historical range of 15% to 20%, extending supply chain commitment risk to the chip supplier side. Unexecuted lease commitments of approximately $822 billion. Over $800 billion of lease contracts have been signed but not yet commenced, not included as current lease liabilities. Additionally, variable lease payments, renewal options, residual value guarantees, and other arrangements similarly remain off-balance sheet. Morgan Stanley estimates that if financing leases are included in the calculation, Microsoft's capital expenditures as a percentage of sales could jump from 33%/50% (fiscal years 2026/2027) to 44%/64%, while Oracle could potentially increase from 76%/115% to 101%/189%. Unpaid capital expenditures in accounts payable of approximately $110 billion. The days payable outstanding (DPO) of super-large-scale cloud companies have significantly lengthened - Oracle increased by 370% year-on-year, Meta by 73%, and Microsoft by 69% - indicating that the entire supply chain is effectively financing AI construction, with suppliers taking on liquidity pressures that should have been borne by buyers. SPVs and cyclical financing: transferring leverage into the dark Another core dimension of off-balance sheet risk is the cyclical financing structures built through SPVs. This week, Apollo and Blackstone jointly completed a $35 billion "chip-backed" private credit transaction for Anthropic, which notably exemplifies this model's operational logic: Broadcom provides endorsement for the SPV, Anthropic uses the raised funds to purchase Google chips made by Broadcom, and Google holds a 14% stake in Anthropic; Morgan Stanley, arranging the transaction, also provides loans to the participating investors. Morgan Stanley's AI ecosystem financing correlation map shows multiple cycles of funding, financing, and repurchasing relationships between OpenAI, Oracle, Nvidia, Microsoft, CoreWeave, AMD, and Amazon. The same funds circulate repeatedly among a few entities, with SPVs being the core tool to facilitate this circulation. Apollo's insurance subsidiary Athene is particularly active in the above-mentioned structure - raising funds by selling annuities to retirees and then injecting the funds into SPVs for AI infrastructure financing. This model transfers leverage from the visible balance sheets of super-large-scale cloud companies to suppliers and private credit ecosystems, making the real systemic risk exposure difficult for external observers to identify and aggregate. Depreciation cliff and monetization gap: the delayed impact Current financial data have a systematic optimistic bias. A large amount of capital expenditures is currently accounted for as "construction in progress" (CIP) and has not yet been depreciated, artificially inflating reported profit margins and underestimating future cost pressures. Oracle, Meta, and Google's CIP balances have increased by approximately 200%, 90%, and 55% year-on-year, respectively. Once these assets start depreciating, the impact will be felt. Morgan Stanley predicts that over the next three years, cumulative depreciation for Microsoft, Oracle, Meta, and Google will exceed $520 billion. For example, Oracle's depreciation as a percentage of revenue could increase from the current 7% to 28% by the 2028 fiscal year, while Meta could potentially increase from 9% to 19%. In this context, the only way to maintain profit margins is through significant revenue growth - and the current upward revisions of revenue forecasts lag far behind the upward revisions of capital expenditure forecasts. Data shows that Google's 2026 consensus capital expenditure forecast has been raised by 139% from a year ago, while Meta and Amazon have each increased by 85% and 81%, with Oracle seeing the largest increase at 175%. At the same time, the magnitude of revenue forecast revisions significantly lags behind, revealing a structural mismatch where capital expenditures are outpacing the commercialization progress. Furthermore, more than $2 trillion in remaining performance obligations (RPO) are highly concentrated in a few large long-term contracts, with concentrated counterparty risk that cannot be ignored - any issues with a major participant in the cyclical system could trigger a chain reaction. Mismatched timing rather than imminent solvency crisis Morgan Stanley's conclusion is that the risks mentioned above do not currently pose an imminent solvency crisis but rather a combination of timing mismatches and information disclosure gaps: the deferral of depreciation pressures, capital expenditures outpacing monetization progress, leverage shifting to suppliers and private credit layers, and significant discounts in the comparability of capital intensity between different companies due to accounting classification differences. Super-large-scale cloud companies clearly realize the limited window of market sentiment and are seizing the opportunity to maximize the size of their financing. Goldman Sachs analyst Ryan Hammond points out that if the scale of investment in AI infrastructure reaches 2% to 3% of GDP, analogous to the historical construction cycles of railways and automotive industries, capital expenditures could reach $11 trillion by 2027; in an extreme scenario, combining the cash flow of super-large-scale cloud companies with the capacity of the investment-grade credit market could reach a maximum of $14 trillion. However, all of this is contingent upon large language models (LLMs) being able to continually increase token prices and maintain sufficient customer stickiness. More and more companies are turning their attention to AI products with similar performance but significantly lower prices. Once there is a structural shift in demand, this carefully constructed financing system will face a fundamental stress test. This article is a reprinted translation from Wall Street Horizon, edited by GMTEight: Chen Yufeng.