AI版次贷危机？1.8万亿美元表外敞口，正成为这轮狂欢的定时炸弹
Original author: Bu Shuqing
Original source: Wall Street Journal
Amid the frenzy of AI infrastructure construction, an unprecedented wave 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 capital expenditure by hyperscale cloud companies will reach $1.1 trillion to $1.4 trillion by 2027, far exceeding market consensus. However, according to Morgan Stanley's in-depth research, this already staggering figure is just the tip of the iceberg.
Nearly $1 trillion in purchase commitments, over $800 billion in unstarted lease contracts, and tens of billions of dollars in supplier financing arrangements together constitute approximately $1.8 trillion in off-balance-sheet exposure – liabilities that exist off the balance sheet yet firmly lock in future cash outflows.
The market has yet to fully price in these risks.
Morgan Stanley warns that the leverage ratio of hyperscale cloud companies has surged from 0.9x to 1.8x in just two quarters, capital expenditure growth continues to outpace revenue and free cash flow growth, and the true impact of depreciation pressure has yet to arrive.
Meanwhile, private credit institutions such as Apollo and Blackstone are using SPVs (Special Purpose Vehicles) to shift leverage to the supply chain level, creating a highly circular and opaque financing structure. If the commercialization of AI falls short of expectations, or if enterprise customers massively shift to cheaper alternatives, the fragility of the entire financing chain will be exposed.
The Debt Issuance Frenzy: AI Has Become the Biggest Market Variable
According to Morgan Stanley's latest "AI Debt Financing Tracker," as of the end of May 2026, global AI-related bond issuance reached $236 billion, a 357% surge compared to the same period in 2025.
Morgan Stanley estimates that total AI debt issuance for the year will exceed $570 billion, with the pace accelerating further in the second half as demand for capital expenditure financing is concentrated.
In April alone, AI-related bond issuance exceeded $74 billion, a new high for the year, with project finance structures (for data center construction) accounting for 85% of high-yield bond supply and 40% of investment-grade supply. Meanwhile, the five hyperscale cloud companies – Amazon, Meta, Google, Microsoft, and Oracle – now account for 4% of the entire investment-grade bond index.
In terms of leverage, the overall gross leverage ratio of hyperscale cloud companies has risen from 0.9x in Q3 2025 to the current 1.8x, increasing by about 0.3x per quarter, which has already surpassed the leverage level of the entire energy sector.
Morgan Stanley points out that due to supply pressure, related credit spreads have drifted from the AA range to the A range and could widen further. Meta's credit spread is now wider than the CDX IG benchmark.
In terms of free cash flow, Morgan Stanley predicts that the free cash flow of Amazon and Meta will approach zero or turn negative in 2026, at which point incremental financing will rely almost entirely on new debt.
1.8 Trillion in Off-Balance-Sheet Exposure: Invisible Liabilities, Locked-In Cash Outflows
Todd Castagno of Morgan Stanley's Global Valuation, Accounting & Tax team noted in his report that focusing solely on capital expenditure figures would severely underestimate the true financial commitments of the AI construction cycle. Beyond disclosed capital expenditures, there are three key types of off-balance-sheet exposure:
Purchase commitments of approximately $982 billion. The total value of long-term procurement contracts at hyperscale cloud companies and Nvidia is close to $1 trillion. Under accounting standards, unless a company expects a contract loss, these obligations are not recorded as liabilities until the goods are delivered. Therefore, nearly a trillion dollars in future cash outflows are currently not reflected as any liability on the balance sheet.
Notably, Nvidia's own inventory and purchase obligations have risen to about 32% of consensus revenue forecasts for fiscal year 2027, far higher than the historical range of 15% to 20%. Supply chain commitment risk has extended to the chip supplier end.
Unstarted lease commitments of approximately $822 billion. Over $800 billion in lease contracts have been signed but have not yet commenced, which are not included in current lease liabilities. In addition, variable lease payments, renewal options, and residual value guarantees also remain off the balance sheet.
Morgan Stanley estimates that if finance leases were included, Microsoft's capital expenditure as a percentage of sales would jump from 33%/50% (FY2026/FY2027) to 44%/64%, and Oracle's could rise from 76%/115% to 101%/189%.
Unpaid capital expenditure in accounts payable of approximately $110 billion. Days payable outstanding (DPO) at hyperscale cloud companies have been significantly extended – Oracle increased by 370% year-over-year, Meta by 73%, and Microsoft by 69%. This means the entire supply chain is effectively fronting the cost for AI construction, with suppliers bearing the liquidity pressure that should belong to the buyers.
SPVs and Circular Financing: Leverage Shifted into the Shadows
Another core dimension of off-balance-sheet risk is the circular financing structure built through SPVs.
This week, Apollo and Blackstone jointly completed a $35 billion "chip-backed" private credit transaction for Anthropic, which exemplifies the operational logic of this model:
Broadcom provides a backstop for this SPV; Anthropic uses the funds to purchase Google chips manufactured by Broadcom; Google holds a 14% equity stake in Anthropic; Morgan Stanley, which arranged the transaction, also provides loans to the participating investors.
Morgan Stanley's AI Ecosystem Financing Correlation Map shows multiple circular relationships – client, investor, supplier financing, and buybacks – among OpenAI, Oracle, Nvidia, Microsoft, CoreWeave, AMD, and Amazon. The same capital flows repeatedly among a small number of entities, with SPVs serving as the core tool for achieving this circularity.
It is reported that Apollo's insurance subsidiary, Athene, is particularly active in these structures – raising funds by selling annuities to retirees and then injecting the capital into SPVs to participate in AI infrastructure financing.
This model shifts leverage from the visible balance sheets of hyperscale cloud companies to the supplier and private credit ecosystem, making the true systemic risk exposure difficult for external observers to identify and aggregate.
The Depreciation Cliff and Monetization Gap: The Deferred Impact
Current financial data suffers from systematic optimistic bias. A large portion of capital expenditure is currently recorded as "Construction in Progress" (CIP) and has not yet begun to depreciate. This artificially inflates reported profit margins and underestimates future expense pressure.
CIP balances at Oracle, Meta, and Google have increased by approximately 200%, 90%, and 55% year-over-year, respectively.
Once these assets are gradually transferred to depreciation, the impact will be concentrated and released.
Morgan Stanley predicts that the cumulative depreciation for Microsoft, Oracle, Meta, and Google over the next three years will exceed $520 billion. For example, depreciation as a percentage of revenue at Oracle could rise from the current 7% to 28% by fiscal year 2028; for Meta, it could rise from 9% to 19%.
Against this backdrop, the only way to maintain margins is through a significant simultaneous increase in revenue – but current upward revisions in revenue forecasts lag far behind the upward revisions in capital expenditure forecasts.
Data shows that the consensus forecast for Google's 2026 capital expenditure has been raised by 139% compared to a year ago, with Meta and Amazon seeing increases of 85% and 81%, respectively. Oracle saw the largest upward revision, at 175%.
Meanwhile, the magnitude of revisions in revenue forecasts lags significantly behind, clearly demonstrating a structural mismatch where capital expenditure precedes commercialization.
Furthermore, over $2 trillion in remaining performance obligations (RPO) are highly concentrated in a few large, long-term contracts, making counterparty concentration risk non-negligible. A problem with any major participant in this circular system could trigger a chain reaction.
A Timing Mismatch, Not an Imminent Solvency Crisis
Morgan Stanley concludes that these risks currently do not constitute an imminent solvency crisis, but rather a confluence of timing mismatches and information disclosure gaps: depreciation pressure is deferred, capital expenditure outpaces monetization progress, leverage is shifted to suppliers and the private credit layer, and the comparability of capital intensity across companies is severely undermined by accounting classification differences.
Hyperscale cloud companies are clearly aware of the limited window of current market sentiment and are racing to maximize their financing scale.
Goldman Sachs analyst Ryan Hammond points out that if AI infrastructure investment reaches 2% to 3% of GDP, analogous to historical construction cycles for railroads and the automobile industry, capital expenditure could reach $1.1 trillion by 2027. In an extreme scenario, combining hyperscaler cash flows with the capacity of the investment-grade credit market, the upper limit could reach $1.4 trillion.
However, all this hinges on large language models (LLMs) being able to continuously increase token pricing and maintain sufficient enterprise customer stickiness. A growing number of companies are turning their attention to AI products with comparable performance but significantly lower prices.
Once there is a structural shift on the demand side, this carefully constructed financing system will face a fundamental stress test.
