AI data center development has exceeded hyperscalers' operating cash flow capacity, with financing shifting toward leases, debt, equity, and off-balance-sheet SPVs

Overview

The report examines how North American AI computing capacity and data center development can continue to be financed when operating cash flow is insufficient. Its core conclusion is that leases, on-balance-sheet debt, equity, customer prepayments, and chipmaker-supported SPVs are expanding the pool of available funding, but each arrangement entails future cash flow commitments and makes project returns more dependent on external capital costs and accounting presentation.

Core views

AI infrastructure development has broken through internal cash flow constraints. Morgan Stanley expects hyperscalers' cash capital expenditures to exceed $1.2 trillion in 2027, while aggregate operating cash flow will be approximately $1 trillion; these companies are also reinvesting more than 40% of sales revenue into AI capital expenditures. Even with the high margins of their traditional businesses, this level of investment cannot be sustained solely by current operating cash flow. Amazon's and Google's free cash flow turned negative in the second quarter of 2026, and Meta is expected to turn negative in the following quarter, so development funding has begun shifting sequentially toward leases, bond issuance, reduced share repurchases, and equity issuance. The first layer of external financing primarily comes from off-balance-sheet leases and long-term purchase commitments. Data center developers or suppliers can borrow against long-term leases, guarantees, or purchase commitments provided by investment-grade hyperscalers, enabling projects to begin construction before customers make payments or recognize liabilities. The latest disclosures show that hyperscalers' undiscounted lease payment commitments for leases not yet commenced have reached $1.1 trillion; purchase commitments by hyperscalers, Nvidia, and Broadcom for GPUs, memory, wafer capacity, networking equipment, and other equipment exceed $1.7 trillion. Hyperscalers' total undiscounted commitments exceed $2.7 trillion, equivalent to approximately three years of current operating cash flow; including support from Nvidia and Broadcom, the related off-balance-sheet commitments and guarantees exceed $3.1 trillion. These contracts help secure capacity when supply is constrained, but once projects are completed, they will flow through free cash flow and increase operating leverage; if supply and demand normalize earlier than expected, companies may have to pay for excess capacity or renegotiate terms. The second layer is growing on-balance-sheet debt. Once data centers are completed, the corresponding operating or finance lease liabilities enter the balance sheet; hyperscalers also began issuing bonds in late 2025 and accelerated issuance in 2026. Their aggregate long- and short-term debt and lease liabilities have now reached $770 billion, while their issuance as a share of nonfinancial investment-grade bond supply increased from 2% in 2025 to 19% year-to-date in 2026. Although on-balance-sheet leverage still appears limited relative to off-balance-sheet commitments, as upward revisions to capital expenditures increasingly rely on debt rather than cash, the economics of AI projects will become more dependent on interest rates and credit spreads; if new commitments raise financing costs, returns on new infrastructure must cover a higher cost of capital. The third layer is cash retention and equity financing. Hyperscalers currently rely more heavily on leases and debt, but equity financing may become more important if interest rates rise, credit spreads widen, or elevated AI investment persists longer than expected. Before directly issuing shares, companies are retaining cash by reducing buybacks, meaning dilution from equity compensation is no longer fully offset. Microsoft is the only hyperscaler that has not yet reduced buybacks or issued debt. Google, by contrast, has reduced annual buybacks of more than $60 billion to zero and issued $50 billion of equity in the second quarter of 2026, with the two changes together freeing up more than $110 billion for AI infrastructure; after its share count had declined 13% from its peak over the past decade, shareholders are beginning to face net dilution. The report also expects related dilution to rise further if equity compensation per employee continues to grow. Although this expense does not directly consume cash, it affects shareholders' economic interests. New financing sources are also bringing the balance sheets of customers and chip suppliers into the development ecosystem. Oracle disclosed $4.6 billion of customer prepayments for capital expenditures in its most recent quarter; such payments are recorded as deferred revenue, but because more than one year separates cash receipt from revenue recognition, their economic characteristics are closer to debt, requiring interest accrual at Oracle's incremental borrowing rate. Oracle's 10-year and 30-year bonds imply yields to maturity of 6.9% and 7.9%, respectively, and the report believes interest on prepayments should reflect a similar financing cost. The report illustrates that if a customer prepays $10 billion for a two-year term and the seller's incremental borrowing rate is 10%, the seller would recognize approximately $1 billion of interest expense annually, compounded in the second year, and recognize approximately $12 billion of revenue upon delivery. Broadcom and Nvidia, meanwhile, are using chip-financing SPVs to help unrated AI labs obtain equipment. The SPV issues debt to purchase chips and then leases the chips to AI labs; the chipmaker provides residual-value support, covering any shortfall if lease payments cease and the equipment's resale value is insufficient to repay designated bondholders. This enables unrated customers to lease chips at a cost closer to that of investment-grade suppliers, but it also brings risk back to the chipmaker: proceeds from sales to the SPV may need to be allocated between the chip sale and the residual-value guarantee, and the manufacturer may also recognize a guarantee liability at the fair value of the stand-alone guarantee transaction at the time of sale. If the guarantee ultimately requires no payment, the related guarantee income is generally not regarded as operating revenue. Broadcom has announced a chip-leasing arrangement supporting up to 20 GW of computing-capacity capital expenditures, and Nvidia is also reported to be developing a similar structure. Accounting treatment determines when these financing arrangements enter the financial statements. During construction, data centers being built for and leased to hyperscalers generally remain off balance sheet, while the related SPVs constitute VIEs because the tenants' influence over their economic benefits and decisions exceeds their nominal ownership and voting rights. Consolidation depends on whether the hyperscaler both has the power to direct the most important economic activities and has an obligation to absorb losses or a right to receive economic benefits. Meta and Google disclose that they do not consolidate certain data center VIEs because they cannot direct the activities that most significantly affect economic performance; for example, Meta cannot control negotiations with future tenants or property sales. However, this judgment must be continually reassessed. As the likelihood of residual-value guarantees being triggered, operational control, or economic risk changes, projects that are off balance sheet today may later enter the consolidated financial statements. Microsoft extended the estimated useful lives of data centers from 15 years to 25 years, causing some leases to be reclassified from finance leases to operating leases. A typical 15-year lease term for a new data center previously equaled 100% of its former useful life, but under the new estimate it represents only 60%, below the 75% finance lease test threshold cited in the report. Because Microsoft's free cash flow definition includes finance lease capital expenditures but excludes operating leases, the change will reduce reported capital expenditures and increase free cash flow, even though the economic substance of the lease has not changed and remains similar to debt-financed capital expenditures. For cross-company comparisons, the report recommends including the value of data centers obtained through long-term operating leases in capital expenditures and free cash flow, adding back operating lease costs, and treating them similarly to depreciation; rating agencies regard lease liabilities as debt regardless of lease classification. Microsoft's fiscal 2026 10-K disclosed $24.6 billion of assets obtained through finance leases in the prior year, while Oracle disclosed $18.2 billion of assets obtained through operating leases, showing that relying solely on reported classifications may create significant comparability differences.

Analysis framework

The report first compares projected capital expenditures with operating cash flow to identify the internally funded shortfall in AI development; it then breaks down leases and purchase commitments, on-balance-sheet debt, reduced buybacks, and equity issuance in the order in which financing sources emerge, before analyzing newer structures such as customer prepayments and chipmaker-supported SPVs. Finally, through VIE consolidation criteria, lease classification, and free cash flow adjustments, the report explains that legal or accounting off-balance-sheet treatment does not eliminate future cash obligations and uses this analysis to assess financing costs, shareholder dilution, and financial statement comparability.

Methodology notes

• Company Fundamentals and Financial FrameworkFree cash flow analysis

  Analysis of operating cash flow, capital expenditures, and the free cash flow gap

  The report compares hyperscalers' operating cash flow with AI cash capital expenditures and tracks when free cash flow turns negative to determine whether development projects can be self-funded and when external capital will be required.

• Company Fundamentals and Financial FrameworkOperating/Financial Leverage Analysis

  Combined leverage from on-balance-sheet debt and off-balance-sheet long-term commitments

  The report examines not only recognized debt and lease liabilities but also leases not yet commenced, purchase commitments, and guarantees as future cash obligations, revealing the economic burden that reported leverage may understate.

• Fixed Income and Credit AnalysisSpread analysis

  Analysis of bond yields and external financing costs

  The report uses hyperscalers' share of bond issuance, AI financing spreads, and the implied yields to maturity on Oracle bonds to assess the cost of capital that new debt and customer prepayments should bear.

• Industry/Sector Analysis FrameworkSupply-demand framework

  Computing equipment supply constraints and long-term purchase commitments

  Long-term contracts can secure capacity when supplies of GPUs, memory, and networking equipment are constrained, but if supply and demand normalize early, commitments may translate into excess-capacity costs or renegotiation pressure.

• Company Fundamentals and Financial Framework

  VIE primary-beneficiary consolidation assessment

  Based on control over key economic activities and the extent of exposure to losses or benefits, the report assesses whether data center SPVs should be consolidated and emphasizes that this conclusion must be continually reassessed throughout the transaction's life.

• Company Fundamentals and Financial Framework

  Comparability adjustments for operating lease capital expenditures and free cash flow

  The report reassesses data centers obtained through long-term operating leases according to their economic substance as structures resembling debt-financed development, reducing distortions in capital expenditure and free cash flow comparisons caused by lease classification and changes in useful-life estimates.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• Microsoft（MSFT）

  Supports data center development with leases and internal cash and changes the classification of certain leases by extending useful lives.

  Strengths

  The report states that it is the only hyperscaler that has not yet reduced share repurchases or issued debt.

  Weaknesses

  The new accounting estimate will reduce reported capital expenditures and increase free cash flow, weakening direct comparability with companies that use different lease classifications.

  Comparison

  It disclosed $24.6 billion of assets obtained through finance leases in fiscal 2026; the report compares this with Oracle's operating lease assets.

  Risks

  The economic substance of leases remains similar to debt financing, and reported free cash flow alone may understate capital investment.

• Google（GOOGL）

  Frees up funding for AI infrastructure development by halting buybacks and issuing equity while using data center VIEs for financing.

  Strengths

  Adjustments to buybacks and equity financing together freed up more than $110 billion in cash for AI development.

  Weaknesses

  Free cash flow turned negative in the second quarter of 2026, while buybacks that previously offset dilution from equity compensation have declined significantly.

  Comparison

  Its development primarily supports internal models, and its commitments exceed remaining performance obligations.

  Risks

  The net share count may rise, and VIE consolidation judgments require continual reassessment.

• Meta Platforms（META）

  Supports the development of internal AI models through reduced buybacks, data center SPVs, and long-term commitments.

  Strengths

  It can use investment-grade credit and long-term leases to help SPVs obtain financing.

  Weaknesses

  The report expects its free cash flow to turn negative in the following quarter, while lower buybacks make equity compensation more likely to cause dilution.

  Comparison

  Like Google, its computing capacity primarily serves internal models, so commitments exceed remaining performance obligations.

  Risks

  If control over key VIE activities or the economic risks assumed changes, off-balance-sheet projects may need to be consolidated.

• Amazon（AMZN）

  As a computing-capacity provider, it expands computing capacity through leases, debt, and long-term commitments.

  Strengths

  As a seller of computing capacity, its future commitments are supported by corresponding contractual revenue.

  Weaknesses

  Free cash flow turned negative in the second quarter of 2026, making incremental capital expenditures more dependent on external funding.

  Comparison

  Like Microsoft, its commitments are approaching remaining performance obligations; unlike Meta and Google, which primarily build for internal models.

  Risks

  Project economics are increasingly sensitive to external capital costs and the realization of contractual revenue.

• Oracle（ORCL）

  Uses customer prepayments, debt, equity, and SPVs to finance data centers and chip purchases.

  Strengths

  It received $4.6 billion of customer prepayments in the most recent quarter, providing a new source of funding for capital expenditures.

  Weaknesses

  Long-term prepayments have debt-like characteristics and require interest accrual at the incremental borrowing rate.

  Comparison

  It disclosed $18.2 billion of assets obtained through operating leases in fiscal 2026, while Microsoft disclosed $24.6 billion of assets obtained through finance leases.

  Risks

  Implied bond yields of 6.9% to 7.9% indicate that prepayment financing could entail significant interest expense.

• Nvidia（NVDA）

  Uses chip-financing SPVs and residual-value support to help unrated AI labs lease chips at costs closer to those available to investment-grade suppliers.

  Strengths

  Supplier credit support expands customers' access to chip financing and facilitates computing-capacity project development.

  Weaknesses

  Sales proceeds may need to be allocated between the chips and the guarantee, with a guarantee liability recognized at the time of sale.

  Comparison

  It participates with Broadcom in chip-leasing and residual-value support structures.

  Risks

  If lease payments cease and chip resale values are insufficient, residual-value support may require it to cover the SPV's debt-service shortfall.

• Broadcom（AVGO）

  Provides chip-leasing financing and residual-value support and has announced an arrangement supporting up to 20 GW of computing-capacity capital expenditures.

  Strengths

  It can use its own credit to help customers and SPVs obtain lower-cost equipment financing.

  Weaknesses

  It must bear guarantee valuation, accounting recognition, and potential asset residual-value risks.

  Comparison

  It uses a chip-financing SPV model similar to Nvidia's.

  Risks

  If equipment resale values or lease receipts are insufficient, residual-value guarantees may turn into actual payment obligations.

Key data

• Hyperscalers' 2027 cash capital expendituresMore than $1.2 trillionAbove projected aggregate operating cash flow of approximately $1 trillion
• AI capital expenditure intensityMore than 40% of sales revenueThe proportion hyperscalers are reinvesting in AI capital expenditures
• Off-balance-sheet commitments and guaranteesMore than $3.1 trillionAggregate amount disclosed by hyperscalers, NVDA, and AVGO
• Hyperscalers' undiscounted commitmentsMore than $2.7 trillionEquivalent to approximately three years of current operating cash flow
• Lease payment commitments for leases not yet commenced$1.1 trillionUndiscounted amount disclosed by hyperscalers
• Purchase commitmentsMore than $1.7 trillionCommitments by NVDA, AVGO, and hyperscalers for chips, capacity, and equipment
• On-balance-sheet debt and lease liabilities$770 billionHyperscalers' aggregate long- and short-term debt and lease liabilities
• Share of nonfinancial investment-grade bond issuance19%Hyperscalers' year-to-date share in 2026, versus 2% in 2025
• Changes in Google buybacks and equity financingAnnual buybacks reduced from more than $60 billion to zero; $50 billion of equity issued in the second quarter of 2026Together freeing up more than $110 billion for AI infrastructure
• Oracle customer prepayments$4.6 billionDisclosed in the most recent quarter and used to finance capital expenditures
• Implied yields to maturity on Oracle bonds6.9% for 10-year bonds and 7.9% for 30-year bondsThe report believes interest on customer prepayments should reflect similar borrowing rates
• Scale of Broadcom chip-leasing financingSupporting up to 20 GW of computing-capacity capital expendituresA financing arrangement backed by chipmaker residual-value support
• Microsoft data center useful-life estimateExtended from 15 years to 25 yearsThe ratio of a typical 15-year lease term to useful life declined from 100% to 60%
• Assets obtained through leasesMicrosoft $24.6 billion; Oracle $18.2 billionAssets obtained through finance leases and operating leases, respectively, as disclosed for fiscal 2026

Impact & implications

The report believes AI data center development will not stop immediately because of insufficient operating cash flow, as companies are mobilizing lease markets, public bonds, equity, customer funding, and supplier credit support. However, as funding shifts from internal cash to long-term commitments and external capital, project returns must cover higher interest rates, credit spreads, and guarantee costs; shareholders may also bear greater dilution. Meanwhile, off-balance-sheet SPVs and operating leases can make reported capital expenditures, free cash flow, and leverage appear better than their economic substance, so cross-company comparisons need to incorporate leased assets, future commitments, and potential guarantee liabilities.

Risks

• If supply and demand for AI equipment normalize earlier than expected, long-term lease and purchase commitments may result in payments for excess capacity or force companies to renegotiate contracts.
• Higher interest rates or wider credit spreads will increase financing costs for the next phase of development, making it more difficult for project returns to cover the cost of capital.
• Reduced buybacks and increased equity issuance will cause equity compensation and other share issuance to translate into net share growth, increasing shareholder dilution.
• If lease payments to chip-leasing SPVs cease and equipment residual values are insufficient, residual-value guarantees provided by suppliers may create payment obligations.
• If VIE control or economic risks change, data center SPVs currently kept off balance sheet may later need to be included in consolidated financial statements.
• Operating leases and changes in useful-life estimates may understate capital expenditures, overstate free cash flow, and weaken cross-company comparability.

What to watch

• Track the gap between 2027 capital expenditures and operating cash flow, as well as the new funding sources required for subsequent upward revisions to capital expenditures.
• Monitor AI financing interest rates, credit spreads, and hyperscalers' share of investment-grade bond issuance.
• Watch the effect of reduced buybacks, equity issuance, and growth in equity compensation per employee on net share counts.
• Track whether aggregate long-term leases, purchase commitments, and guarantees continue to exceed contractual revenue and remaining performance obligations.
• Monitor the scale, interest-accrual methods, guarantee liabilities, and residual-value support terms of customer prepayments and chip-leasing SPVs.
• Watch the continual reassessment of VIE primary-beneficiary determinations and the impact of operating leases on capital expenditure and free cash flow definitions.