Off-balance-sheet AI infrastructure commitments are expanding rapidly, raising real leverage pressure

Overview

The report discusses the rapidly expanding off-balance-sheet financing structure in AI infrastructure buildout. Large technology companies and NVDA have signed substantial long-dated procurement, lease, guarantee, and power purchase commitments to secure GPU, memory, cloud-computing capacity, data-center footprint, and power supply. These commitments help suppliers, data center developers, power companies, and private credit providers obtain financing, but many obligations will not appear on the balance sheet until goods are delivered, leases commence, or payments become unavoidable.

Core views

The core view is that financing for AI supply-chain buildup is not fully reflected on the balance sheets of large tech companies and NVDA; instead, it is transferred to suppliers, developers, banks, public debt markets, and private credit systems through contractual commitments. Off-balance-sheet structures reduce short-term reported leverage, but they do not eliminate underlying economic risk. As contract size grows, tenors lengthen, and structures become more complex, investors who look only at reported liabilities may underestimate true leverage, future cash-flow usage, and downside risk if demand weakens.

Analysis framework

The report breaks down six types of AI infrastructure commitments from accounting-recognition and credit-analysis perspectives: unconditional purchase obligations, inventory procurement commitments, take-or-pay or fixed-purchase-quantity contracts, off-balance-sheet lease payments, power purchase agreements, and chip financing through contractual support, chip leasing, and prepayments. The analysis focuses on determining when obligations enter the balance sheet, when they remain off-balance commitments, and how rating agencies may adjust debt metrics.

Methodology notes

• Accounting recognitionLease identification framework

  Control of specific assets

  If a contract gives the customer control of specific GPUs, racks, or production facilities for a period and the right to obtain substantially all economic benefits, minimum payments are more likely to be accounted for as lease liabilities on the balance sheet; if the contract only purchases compute capacity or power services, it usually remains an off-balance-sheet commitment.

• Credit analysisOff-balance-sheet commitment adjustments

  Economic liabilities

  Even when procurement commitments, guarantees, or PPAs do not create on-balance-sheet debt, rating agencies and investors may still treat them as future cash-flow pressure or debt-equivalent risk.

• Lease accountingNot-yet-started leases and variable lease payments

  Off-balance-sheet lease payments

  Signed but not-yet-started leases, variable lease payments billed by power or usage, renewal options, and residual-value guarantees may not be fully reflected in current lease liabilities.

• Energy contractsPower purchase agreement accounting treatment

  PPA classification

  PPAs may be classified as leases, normal purchase commitments, or other contractual arrangements; whether they are recognized on balance sheet depends on whether the company controls related power generation assets and whether contract payment terms are fixed.

Asset mapping & comparison

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

• META.US

  Key participant in AI data-center and power procurement commitments

  Strengths

  Long-term commitments help secure compute and energy supply, supporting AI product expansion.

  Weaknesses

  Debt-to-future operating cash flow ratio of around 1.7x; unstarted leases, variable lease costs, and PPAs may understate true commitments.

  Comparison

  Pressure appears lower than ORCL, but expansion of off-balance-sheet obligations still warrants close monitoring.

  Risks

  AI demand below expectations, lease liabilities being brought on balance sheet, and rising fixed costs from power purchase agreements.

• NVDA.US

  Core company in AI chip supply-chain and inventory procurement commitments

  Strengths

  Securing foundry capacity and memory supply early helps support AI product demand.

  Weaknesses

  Inventory and procurement spend as a share of expected revenue has risen to about 32%, above the usual 15%-20% range.

  Comparison

  Compared with internet platforms, NVDA’s risk is more concentrated in inventory procurement, supply-chain prepayments, and demand volatility.

  Risks

  Demand declines could create loss risk around inventory, procurement commitments, or supply-chain prepayments.

• GOOGL/Alphabet

  Major credit backer for data-center lease support and power purchase agreements

  Strengths

  Investment-grade credit can help data-center projects obtain higher leverage and lower financing costs.

  Weaknesses

  Third-party lease support and guarantees are typically off-balance-sheet, but rating agencies may adjust debt metrics when obligations become effective.

  Comparison

  Its guarantee structure is an important case in the report on how agencies may treat AI infrastructure liabilities.

  Risks

  Triggering guarantee obligations, debt remeasurement after lease commencement, and long-term fixed-cost pressure from PPAs.

• MSFT

  Participant in cloud-capacity contracts and long-term power purchase agreements

  Strengths

  Secures AI compute and clean power supply through long-term contracts.

  Weaknesses

  Fixed-volume or take-or-pay contracts may limit future flexibility.

  Comparison

  Like CRWV-related contracts, it shows that large customer commitments can underpin supplier data-center financing.

  Risks

  Minimum payment or fixed-purchase obligations may still apply if demand changes.

• ORCL

  Participant in AI data-center, power agreement, and cloud infrastructure expansion

  Strengths

  Long-term customer demand and PPAs can support infrastructure financing.

  Weaknesses

  Debt-to-future operating cash flow ratio above 7x indicates elevated financial commitment pressure.

  Comparison

  Among the sample disclosed in the report, ORCL shows the most pronounced leverage pressure relative to future cash flows.

  Risks

  Oversized contract commitments, fixed-cost obligations from PPAs, and insufficient future cash flow coverage.

• Power utilities and independent power producers

  Beneficiaries of AI data-center PPAs

  Strengths

  Long-term PPAs can provide financing support for nuclear, gas microgrid, and renewable projects.

  Weaknesses

  Projects depend on long-term performance by large tech customers and continued data-center electricity demand.

  Comparison

  Compared with traditional power demand, AI data-center demand is more concentrated, longer-duration, and more contract-structure dependent.

  Risks

  PPA renegotiation, regulatory changes, and data-center electricity demand below expectations.

Key data

• Total AI infrastructure-related commitmentsMore than $1.3 trillionIncludes approximately $640 billion in procurement spend and approximately $675 billion in lease spend.
• META debt-to-future operating cash flow ratioabout 1.7xUsed to gauge the scale of contract and leverage pressure relative to future cash flows.
• ORCL debt-to-future operating cash flow ratioabove 7xThe report says contract tenors and business scaling materially increase pressure.
• Procurement commitment growthMore than doubled over the past year and up about sixfold over the past five yearsReflects the trend of locking in AI infrastructure resources through long-term contracts.
• NVDA inventory and procurement spend as a share of expected revenueAs of Jan 2026, about 32% of FY2027 expected revenueThe usual range is about 15%-20% of forward revenue; this level indicates a higher degree of early procurement.
• Large tech future lease commitmentsOver $67 billionClearly higher than about $24 billion a year ago, with many leases not yet put into effect.
• GOOGL lease supportabout $10 billionSupports leasing for four data centers under construction; the report also cites about $17 billion of support by Alphabet for data-center buildouts tied to bitcoin miners.
• Data-center electricity cost example15-year power cost of a 200 MW facility is about $2.5 billionEstimated using a national average electricity price of about $80/MWh.
• GOOGL PPA example20 years, $9.9 billionThe report states that the agreement signed in January 2026 will be accounted for as a lease.

Impact & implications

For investors, the key risk in AI infrastructure investment is not only capex itself, but also the potential impact of off-balance-sheet contracts on future cash flows, financing flexibility, and credit ratings. If AI demand remains strong, long-term commitments can secure supply and create a competitive advantage. If demand weakens, fixed procurement, data-center leases, power purchases, and guarantee obligations could translate into losses, cash-flow stress, or rating revisions. For suppliers, developers, and power companies, credit support from large tech firms helps with financing but also increases chain risk by making the sector more dependent on continued demand from a small set of high-credit customers.

Risks

• The expansion of off-balance-sheet commitments may lead to underestimation of true leverage.
• If AI demand slows, fixed procurement, leasing, and PPAs may reduce financial flexibility.
• When not-yet-started leases begin to take effect, lease liabilities on the balance sheet may rise sharply.
• Variable lease payments, residual-value guarantees, and third-party guarantees may create cash-flow pressures that are not fully disclosed.
• Supplier and developer financing depends on large-tech customer commitments; contract renegotiation can become difficult if expectations change.
• Rating agencies may adjust debt metrics for guarantees and off-balance-sheet commitments.

What to watch

• The pace of growth in future procurement commitments, lease commitments, and PPA disclosures by mega-cap tech companies and NVDA.
• When unstarted leases enter the balance sheet and the scale of newly added lease liabilities.
• Whether AI training and inference demand is sufficient to absorb the contracted compute, chips, and power supply.
• Whether rating agencies adopt debt-adjustment metrics for lease guarantees, residual-value guarantees, and contingencies.
• Whether data-center developers, power suppliers, and chip supply chains continue to rely on customer long-term commitments for financing.
• Whether companies disclose more information on contract tenors, minimum payments, renewal options, and termination costs.