Morgan Stanley: Nvidia Captures $690B AI Capex Alone; Debt and Power Risks Underestimated
nashnova research
Morgan Stanley traced $1.58 trillion in AI capital spending to individual stocks and found Nvidia alone set to capture $690 billion (67%); strip Nvidia out, and AI capex lifts S&P 500 revenue by just 1.6% — while hidden lease debt and a widening power gap remain systematically underpriced.
Where does $1.58 trillion go?
Nine buyers — Alphabet, AWS, Microsoft, Meta, SpaceX, Oracle, CoreWeave, Nebius, and a leasing pool — spend a combined $1.58 trillion, flowing into semiconductors and servers, networking and optics, data-center infrastructure, and storage.
The largest single category is semiconductors and servers at $672 billion. Nvidia is the dominant recipient.
The second-largest category is "international suppliers" at $511 billion — Samsung, SK Hynix, TSMC-adjacent infrastructure, and Taiwanese ODMs — none of which sit inside the S&P 500.
This means → more than a third of AI capex flows outside the U.S. equity universe. Looking only at the S&P 500 significantly understates the full supply chain.
How much does Nvidia take — and what happens without it?
Nvidia alone is expected to capture roughly $690 billion, or 67% of the ~$1.03 trillion the S&P 500 absorbs from AI capex. That is 44% of the entire $1.58 trillion budget landing on one company's income statement.
The other 499 S&P 500 constituents split roughly $340 billion. Strip Nvidia out, and AI capex contributes just 1.6% to S&P 500 sales.
Nvidia derives 91.7% of revenue from AI infrastructure; Broadcom 81%, Vertiv 90%, Seagate and Western Digital each 60%, and even Caterpillar 21%.
In plain terms = more than a dozen large companies have their earnings virtually locked to AI capex. If spending is revised down, there is no offsetting business to cushion the hit — earnings estimates get cut directly.
What is "hidden capex"?
Data-center leases are economically equivalent to debt-financed capital spending but do not appear in the GAAP capex line item — so markets have long overlooked them.
Morgan Stanley estimates lease-based hidden capex at roughly $195 billion in 2027, rising to $246 billion in 2028. Including leases lifts 2027 total AI capex from $1.39 trillion (cash basis) to $1.58 trillion, a year-on-year increase of about 60%.
Oracle carries the largest 2027 lease exposure at $57 billion, followed by Microsoft $45B, Meta $37B, Amazon $28B, and Alphabet $27B.
This means → if you read only the "capital expenditures" line on a balance sheet, you miss nearly $200 billion in real commitments and the debt pressure that comes with them.
How severe is the debt pressure — and who is most exposed?
Hyperscalers have issued roughly $255 billion in unsecured bonds year-to-date, about 30% of estimated 2026 capex. Amazon ($98B) and Alphabet ($81B) lead.
Morgan Stanley's stress test: if AA-rated hyperscalers issue $400 billion at 7%, their interest-coverage ratio falls from 30.7× to 11.9× — still comfortable.
But if Oracle and SpaceX fund $120 billion at 8.5%, their combined coverage ratio drops from 4.0× to 1.8×. The report states explicitly that emerging cloud providers "rely almost entirely on debt issuance to fund their capex ambitions."
In plain terms = the big hyperscalers can handle the borrowing. Oracle and SpaceX — the "new entrants" — are running on a financial safety margin that is already dangerously thin.
Where does the power gap bind?
Morgan Stanley's "AI power" model shows that even accounting for fast-track power solutions (covering 34% of the required 97 GW), the U.S. faces a net shortfall of 32 GW through 2028, widening to roughly 72 GW through 2029.
The report frames this as a timing mismatch, not a demand collapse — power will arrive, but not on schedule. Storage, networking, optics, and ODM buildouts are most sensitive to power delays.
AI capex growth is projected to decelerate sharply from about 60% in 2027 to roughly 14% in 2028. In the slowdown phase, power-equipment names retain pricing power; storage, hard drives, and cooling equipment also make the list.
How underpriced is the downside risk?
Morgan Stanley cites the 2022 NAND memory cycle: shipment volumes fell only about 5%, but prices dropped 20–30%. The conclusion: "the market currently underestimates the risk of volume and price declining simultaneously."
The prevailing buyer psychology — "fear of missing a purchase window" — could reverse. If it does, the logic applies systemic pressure across the entire AI supply chain.
Nvidia's 91.7% AI-revenue share means its exposure to this risk is the most concentrated. This reflects a fundamental paradox: the biggest winner is also the most fragile link — once the capex cycle turns, concentration flips from advantage to vulnerability.
市场有风险，内容仅供研究参考，不构成投资建议。