Morgan Stanley: AI Giants Face Operating Cash Flow Crunch, Off-Balance-Sheet Commitments Exceed $3.1 Trillion
nashnova research
Morgan Stanley warns that hyperscaler capex will exceed $1.2 trillion by 2027 while operating cash flow totals only about $1 trillion — Amazon and Google have already turned free-cash-flow negative, marking the moment AI investment shifts from self-funded to debt-fueled.
How big is the cash shortfall?
Hyperscalers now plow over 40% of revenue back into AI capex — far beyond what operating cash flow can support.
Amazon and Google posted negative free cash flow in Q2 2026; Meta expects to follow next quarter. This means → these companies are spending more on AI than they earn.
In plain terms = they used to fund investment from profits alone; now they must borrow and issue equity to close the gap.
How fast is the borrowing ramping up?
Hyperscalers began issuing bonds in late 2025; by 2026, both debt and equity issuance accelerated — their share of investment-grade non-financial corporate bond issuance jumped from 2% in 2025 to 19% in 2026 year-to-date.
Google cut annual buybacks from over $60 billion to zero and issued $50 billion in new stock in Q2 2026 — freeing up over $110 billion for AI. This means → shareholders are now absorbing dilution; Google had shrunk its share count by 13% over the prior decade, and that trend has reversed.
Total on-balance-sheet debt plus lease liabilities across hyperscalers has reached $770 billion.
Where is the $3.1 trillion hiding off-balance-sheet?
The larger risk sits off-balance-sheet: hyperscalers plus Nvidia and Broadcom have disclosed commitments and guarantees totaling over $3.1 trillion, covering leases, purchase obligations, guarantees, and credit-support structures.
Uncommenced lease payments alone total $1.1 trillion (undiscounted): Microsoft $329 billion, Oracle $261 billion, Meta $279 billion, Amazon $137 billion, Google $85 billion.
Purchase commitments add up to $1.7 trillion; Google alone accounts for $707 billion. In plain terms = the money hasn't been spent yet, but the contracts are signed — walking away would be extremely costly.
Are accounting choices obscuring the real burden?
Last quarter Microsoft extended data-center useful life from 15 to 25 years, reclassifying a large volume of finance leases as operating leases. This means → reported capex looks lower and free cash flow looks higher, but the actual cash outlay hasn't changed by a dollar.
Meta and Google both disclose that their data-center SPVs — special-purpose vehicles, standalone legal entities set up for a specific project — are kept off the balance sheet during construction, on the grounds that neither company is the SPV's "primary beneficiary."
Morgan Stanley warns this judgment is not permanent: if conditions shift, consolidation could be triggered, and reported leverage would jump overnight.
Creative financing — innovation or risk?
Oracle pioneered customer prepayments as a capex funding source, receiving $4.6 billion last quarter. This reflects traditional channels reaching saturation — the giants are now collecting rent in advance from their own customers.
These prepayments appear on the balance sheet as deferred revenue, but they function more like debt — Oracle must recognize interest at its incremental borrowing rate, with implied yields of 6.9% on 10-year and 7.9% on 30-year bonds.
Broadcom and Nvidia have each set up chip-financing SPVs that issue debt to purchase chips and lease them to unrated AI labs, with the chip makers providing residual-value guarantees. In plain terms = the chip vendors lend their credit ratings to smaller companies, letting them access hardware at near-investment-grade financing costs.
What is the ultimate question?
Morgan Stanley distills the issue into one equation: the return on new AI infrastructure must exceed its rising cost of capital.
Each new wave of commitments pushes financing costs higher. This means → every data center built later is more expensive than the last, and the break-even bar keeps climbing.
If that equation breaks down, the entire buildout thesis faces reassessment — this reflects the fundamental tension in the AI investment race: the question is not whether the technology works, but whether the math adds up.
市场有风险，内容仅供研究参考，不构成投资建议。