According to a 65-page in-depth report released by Morgan Stanley on October 15, the bank for the first time broke down the capital expenditures of hyperscale cloud providers (Hyperscalers) and emerging AI cloud service providers (Neoclouds) into verifiable cash flows on a segment-by-segment basis.
Calculations show that related capital spending will reach approximately $1.58 trillion by 2027, with $672 billion flowing to semiconductors and servers, accounting for 42%. Morgan Stanley maintained its positive assessment of the AI investment cycle but explicitly cautioned that opportunity analysis can no longer rely solely on total capital expenditure figures — capital flow direction, genuine supplier value, and cash returns after computing power comes online must each be verified separately.
1. The Composition of $1.58 Trillion: Lease-Adjusted Figure 14% Higher Than Pure Cash Spending
Unlike the cash flow statement approach commonly used by the market, Morgan Stanley modeled lease investments separately and incorporated them into cash expenditures: 2027 cash capital expenditure is approximately $1.388 trillion, plus $195 billion in lease investments, bringing the total to $1.583 trillion — about 14% higher than the pure cash figure. Under the same methodology, total investment for 2026 is approximately $978 billion, and for 2028 approximately $1.812 trillion — with growth rates slowing from roughly doubling in 2026 to about 60% in 2027 and about 14% in 2028. Signed but unfulfilled lease commitments total approximately $1.118 trillion, representing a cross-period stock that cannot be directly counted as expenditures for a given year.
Breaking down the $1.58 trillion for 2027 by category: semiconductors and servers at $672 billion (42%), international suppliers at $511 billion (32%), U.S. storage at $217 billion (14%), networking and optical communications at $215 billion (14%), and data center infrastructure at $119 billion (8%).
2. Capital Flow Transmission: S&P 500 Absorbs 65%, NVIDIA Alone Captures Two-Thirds
After mapping expenditures to supplier revenue and eliminating double-counting (HBM already included in accelerators is counted under the chip segment, while server ODMs are credited with only about 4%–6% for assembly contributions), approximately $1.6 trillion in spending corresponds to roughly $1.7 trillion in supplier revenue. Of this, about 65% can be absorbed by S&P 500 companies, corresponding to approximately $1.029 trillion in revenue for 2027 — about 4.9% of the index companies' total revenue. NVIDIA accounts for approximately $690 billion, representing two-thirds of this inflow.
Under this model, S&P 500 revenue growth in 2027 would rise from approximately 6.6% (excluding direct AI impact) to approximately 8.6%, with direct AI capital expenditure contributing about 24% of that year's revenue growth. A separate broader measure (covering 42 AI enablers and adopters) shows: AI-related companies' revenue growth of approximately 28% in 2027 versus approximately 5% for the non-AI group; the AI-related group contributes approximately 78% of the S&P 500's profit增量 but accounts for only about 37% of that year's net profit — growth contribution does not equal profit stock, and the two must not be conflated.
3. Returns and Financing: A 25%–50% ROIC Path Meets $255 Billion in Bond Issuance
Morgan Stanley constructed a return framework for GPU leasing and API businesses, concluding that under certain conditions a return on invested capital path of approximately 25%–50% exists. Using a 1GW GB300 data center GPU leasing model as an example: total investment of $39 billion ($23 billion in IT equipment, $16 billion in non-IT facilities), under a 75% utilization assumption, GPU hourly rental rates rising from $7 to $8.5 and $10 correspond to ROIC of approximately 23%, 31%, and 39% respectively — return conclusions are highly sensitive to rental rate assumptions.
On the financing side, the gap between sample cloud providers' operating cash flow and capital expenditure will reach its widest point within the forecast period in 2027, with unsecured bond issuance already at approximately $255 billion year-to-date. Stress testing shows: the high-credit group's interest coverage ratio falling from 30.7x to approximately 11.9x remains manageable, while the weaker-credit group drops from 4x to approximately 1.8x — under the same technological opportunity, financing capacity divergence will determine who stays in the game.
4. Three Major Constraints: Power, Memory Cycle, and Optical Communication Coefficient
First, power. Morgan Stanley's power model estimates that even accounting for measures to shorten power supply wait times, the U.S. data center cumulative net power supply gap through 2028 will still be approximately 32GW — about one-third of the 97GW demand — and "racks bought but unable to get powered" may become the biggest physical constraint.
Second, memory. Capital expenditure upward revisions first transmit through pricing, and price increases also raise server procurement costs — higher capital expenditure figures do not necessarily correspond to increased computing power volume. Morgan Stanley judges that the memory cycle may enter its late stage in Q4 2026 but has not yet ended.
Third, optical communications. Each additional $1 of cloud provider capital expenditure generates approximately $0.15 of networking revenue (covering chips, switches, optical components, and interconnect equipment), with Zhongji Innolight and Eoptolink both within the report's tracking scope; however, 1.6T/3.2T upgrades, silicon photonics, and CPO architecture changes may alter per-unit bandwidth component intensity, and capacity expansion outpacing deployment could turn shortage into oversupply.
Additionally, the server segment warrants caution regarding a "revenue surge, gross margin decline" combination: VR200 versus GB300 increases ODM per-rack value by 35%–40%, but the complete rack selling price nearly doubles, with incremental materials flowing more toward memory, GPU, PCB, and other segments.
Conclusion
Morgan Stanley believes the market's focus may gradually shift from pure infrastructure providers to enterprises that combine both supply and application capabilities and genuinely use AI to improve operations. For investors, the tracking framework can be summarized in four steps: align cash expenditures, leases, and actual deployments; break down supplier revenue into volume, price, and share; verify utilization and payment for facilities that have come online; and verify financing costs and cash collections. $1.6 trillion is a demand forecast — who can convert orders into cash and who retains competitiveness after demand slows will be the watershed for the next phase of pricing.
Comments