Morgan Stanley: AI Capex Rises to $1.58 Trillion by 2027 After Including Leases
nashnova research
Morgan Stanley's latest research adds lease commitments back into AI capital spending, lifting the 2027 total to $1.58 trillion — roughly 14% above the cash-capex-only figure. That gap means investors have been systematically underreading nearly $200 billion of real commitment, and the question of who funds it, who earns it, and who buckles under it is sharpening fast.
Why does the traditional capex number "miss" so much?
Hyperscalers acquire data centers through leases — economically equivalent to debt-financed asset purchases — but these commitments never appear in the cash-flow statement's capex line and are excluded from management guidance.
This means → the standard capex framework systematically omits the lease "shadow bill."
Morgan Stanley notes that hyperscalers have disclosed over $1.1 trillion in lease commitments not yet commenced — money that will be spent, but investors cannot see it in the financials.
What does the $1.58 trillion growth curve look like?
Morgan Stanley's lease-inclusive estimates: ~$980 billion in 2026, ~$1.58 trillion in 2027, ~$1.81 trillion in 2028, with year-on-year growth of roughly 100%, 60%, and 14% respectively.
In plain terms = 2026–2027 is the steepest ramp — ~$700 billion of incremental AI investment in two years. By 2028, growth slows sharply to 14%, signaling a shift from "build-out surge" to "digestion phase."
Lease capex alone reaches $195 billion in 2027 and $246 billion in 2028 — the lease share keeps rising.
Where does that $1.58 trillion end up?
Morgan Stanley tracks the spend through to ~50 companies' revenues. Roughly 65% of AI capex flows into S&P 500 constituents, with Nvidia alone capturing about two-thirds of that flow.
Chips: Nvidia is in the early ramp of its next-generation Feynman platform; revenue per gigawatt of compute could exceed $50 billion. At ~15× FY2028 EPS, Morgan Stanley sees room for multiple expansion. Broadcom benefits from custom-silicon growth — the custom chip market is projected to surpass $1 trillion by 2028.
Networking: every $1 of hyperscaler capex generates roughly $0.25 in networking revenue, pointing to ~$374 billion in 2027 network revenue. Nvidia's networking unit, Broadcom, Innolight, and Arista Networks lead the hyperscaler-linked share.
Why is memory the "cheapest bottleneck"?
DRAM is a critical AI-build bottleneck, yet memory stocks trade at only ~3× trailing-twelve-month P/E, with near-zero implied growth premium.
This means → the market is pricing almost no AI upside into memory — the coldest segment across the entire AI supply chain.
Morgan Stanley expects the memory cycle to enter its late stage in Q4 2026, with price gains moderating and rising supply capping earnings surprises. SK Hynix and Samsung have already locked in most demand through long-term agreements.
The lease "shadow bill" — which company's number is biggest?
2027 single-company lease capex: Oracle at $57.2 billion leads, followed by Microsoft at $45.3 billion, Meta at $37.2 billion, Amazon at $28.4 billion, and Google at $26.6 billion.
This reflects Oracle's outsized reliance on the lease model in the AI infrastructure race — its cash-flow base is far smaller than the Big Four hyperscalers, yet its lease commitment is the largest.
On returns, GPU leasing (IaaS) delivers 60%–70% incremental EBIT margins and 25%–40% ROI; model-API monetization can push incremental EBIT margins above 70%, with ROI as high as 63%.
Who can carry this bill — and who cannot?
Investment-grade hyperscalers (Amazon, Meta, Google, Microsoft) carry an interest-coverage ratio of ~30.7×. Even with $400 billion of new debt in 2027, coverage stays at ~11.9× — largely insensitive to funding costs.
Lower-rated issuers (e.g., Oracle) start at only ~4× coverage. An additional $120 billion in debt would push that to ~1.8× — put simply = for every dollar of profit, nearly 60 cents goes to interest.
Emerging cloud players are the most fragile, relying on project-level financing and convertible debt, typically requiring credit support from investment-grade counterparties. This means → the sustainability of AI infrastructure expansion is diverging along credit lines: the cash-rich spend more comfortably, the cash-light borrow at an ever-tighter squeeze.
市场有风险，内容仅供研究参考，不构成投资建议。