When the traditional capex lens is no longer complete, lease financing is quietly reshaping the full picture of AI infrastructure investment.

According to the latest research report from Morgan Stanley's Nicholas Lentini team, after incorporating lease capex into the tally, total AI computing investment by hyperscalers and neoclouds in 2027 rises to approximately $1.58 trillion, about 14% larger than the cash capex measure. That figure implies an AI investment increment of roughly $700 billion between 2026 and 2027, a growth rate of about 60%.

Morgan Stanley notes that this $1.58 trillion in capex will directly drive about 5% of S&P 500 revenue in 2027 and contribute roughly 25% of revenue growth. More broadly, the "Enablers" and "Adopters" that Morgan Stanley classifies as core AI beneficiaries will together drive 78% of S&P 500 earnings growth in 2027. Notably, NVIDIA Corp (NASDAQ: NVDA) alone is expected to capture about two-thirds of the AI capex inflows into the S&P 500.

Research shows hyperscalers have disclosed more than $1.1 trillion of lease commitments that have not yet commenced. Because data center leases are asset purchases of a debt-financed nature, they do not appear as capex in the cash flow statement and are not included in management guidance, causing the market to chronically underestimate the true scale of AI investment. After including lease items, Morgan Stanley estimates total AI capex for 2026, 2027 and 2028 at approximately $980 billion, $1.58 trillion and $1.81 trillion, with year-over-year growth of about 100%, 60% and 14%, respectively.

Lease Capex: How Large Is the "Shadow Bill"

In this report, Morgan Stanley systematically introduces hyperscaler lease capex forecasts for the first time. According to its estimates, lease capex reaches $195 billion in 2027 and expands further to $246 billion in 2028. On a single-company basis, Oracle Corp (NYSE: ORCL) contributes the most in 2027 at $57.2 billion, followed by Microsoft Corp (NASDAQ: MSFT) at $45.3 billion, Meta Platforms Inc (NASDAQ: META) at $37.2 billion, Amazon.com Inc (NASDAQ: AMZN) at $28.4 billion and Alphabet Inc (NASDAQ: GOOGL) at $26.6 billion.

The core of the issue lies in accounting treatment. When hyperscalers acquire data centers through leases, the economic substance is equivalent to debt-financed asset purchases, but this is not reflected under the capex line in the cash flow statement, nor is it included in management's capex guidance. Morgan Stanley believes this makes traditional capex forecasts "increasingly unable to fully reconstruct the complete picture of capital投入."

From a return-on-investment perspective, Morgan Stanley is optimistic about this batch of investment. The report estimates the return on invested capital (ROIC) of new computing capacity at between 25% and 50%, and models two business modes — GPU leasing (IaaS) and model APIs. The former can achieve marginal EBIT margins of 60% to 70% and ROIC of 25% to 40%, while the latter can have incremental EBIT margins above 70% and ROIC as high as 63%.

Where AI Capex Flows: Who Is Sharing the Pie

Using a bottom-up approach, Morgan Stanley traces the $1.58 trillion in capex to the revenue of about 50 companies, covering sectors including technology, communication services, industrials, energy and overseas ODMs. By its estimates, about 65% of AI capex ultimately flows into the revenue of S&P 500 constituents, with NVIDIA Corp (NASDAQ: NVDA) accounting for about two-thirds of the inflow.

Among the beneficiary sectors, AI chips, memory, networking and ODMs (original design manufacturers) capture the largest share, with NVIDIA Corp (NASDAQ: NVDA), Samsung Electronics Co Ltd (KRX: 005930), SK hynix Inc (KRX: 000660) and Micron Technology Inc (NASDAQ: MU) together taking the vast majority.

At the chip level, Morgan Stanley believes NVIDIA Corp (NASDAQ: NVDA) offers the best risk-reward, arguing that the company is in the early ramp of a new product cycle, its Fenyman platform could push revenue per GW of compute beyond $50 billion, and its current valuation of about 15x FY28 EPS leaves room for multiple expansion. Broadcom Inc (NASDAQ: AVGO) also benefits from expanding custom silicon business, with the XPU market expected to exceed $1 trillion by 2028.

At the memory level, DRAM has become a key bottleneck for AI buildouts, but related stocks trade at only about 3x NTM P/E, implying almost zero growth premium. Morgan Stanley judges that the memory cycle will enter its late stage in the fourth quarter of 2026, with price increases moderating and added supply limiting room for earnings surprises. SK hynix Inc (KRX: 000660) and Samsung Electronics Co Ltd (KRX: 005930) have locked up most demand through long-term agreements (LTAs). SK hynix Inc (KRX: 000660) has committed to repurchasing and cancelling 40 billion won of shares, while Samsung Electronics Co Ltd (KRX: 005930) has set a shareholder return target of 90 billion to 110 billion won for 2026.

At the networking level, every additional $1 of hyperscaler capex corresponds to about $0.25 of networking revenue, with total networking revenue expected at about $374 billion in 2027. Among the major companies covered, NVIDIA Corp (NASDAQ: NVDA) (networking business), Broadcom Inc (NASDAQ: AVGO), Innolight Technology Corp (SZSE: 300308) and Arista Networks Inc (NYSE: ANET) are expected to rank among the leaders in hyperscale cloud-related networking revenue.

At the design software level, hyperscalers continue to increase custom silicon investment, providing incremental opportunities for Synopsys Inc (NASDAQ: SNPS) and Cadence Design Systems Inc (NASDAQ: CDNS), especially Synopsys Inc (NASDAQ: SNPS)' newly launched application-optimized IP (AOIP) model, which ties revenue more directly to chip shipments through production royalties.

Power demand triggered by data center construction is becoming an important growth engine for industrial and power equipment suppliers. Morgan Stanley notes that the transmission path of AI capex to the industrials sector is concentrated in power supply and cooling equipment.

On power equipment, GE Vernova Inc (NYSE: GEV) gas turbines are sold out through 2029, with prices rising from $1,000 per kW in the early 2020s to $3,000 per kW now, and data center-related orders already account for about one-third of its pipeline orders. INIO and BE benefit respectively from demand for high-speed reciprocating engines and fuel cells. Morgan Stanley believes that higher capex will further extend power OEM backlogs and improve margins.

Financing Pressure: Divergence Between High-Rated and Low-Rated Issuers Intensifies

As capex continues to expand, the free cash flow gap is widening. Morgan Stanley estimates that 2027 will be the year with the largest cash flow deficit in the forecast period. Hyperscalers have issued about $255 billion of unsecured bonds so far this year, roughly 30% of estimated full-year capex, but the full-price yield on 10-year bonds has risen about 100 to 180 basis points from the end of 2025.

Morgan Stanley divides financing entities into two tiers: high-rated hyperscalers (Amazon.com Inc (NASDAQ: AMZN), Meta Platforms Inc (NASDAQ: META), Alphabet Inc (NASDAQ: GOOGL), Microsoft Corp (NASDAQ: MSFT)) currently have interest coverage of about 30.7x. Even with an additional $400 billion of bond issuance in 2027, coverage could still remain at 11.9x, leaving them largely insensitive to financing costs. Low-rated entities (Oracle Corp (NYSE: ORCL), SPCX) have coverage of only about 4x. If they add $120 billion of bond issuance over the same period, coverage would fall to about 1.8x, exposing them to significant financing constraints.

Neoclouds mainly rely on project-level financing and convertible bonds, making them the group most sensitive to financing costs. They typically need credit support from high-rated counterparties to obtain financing.

Looking at the evolution of financing structure, Morgan Stanley expects bond issuance maturities to tilt toward the short end in 2027. Because chips and servers generate cash flow relatively quickly, they are better suited to asset-backed financing and private placement formats, offering flexibility tied to contracted demand.