Off-Balance Sheet AI Financing: Smart Structuring or Hidden Leverage?
Big Tech is pouring billions into AI data centers, and a growing slice of that spend is being kept off core balance sheets through special purpose vehicles (SPVs). The approach is legal and disclosed, but it adds layers of debt, leases, and guarantees that don't show up as traditional borrowings.
For finance teams, the question isn't solvency today. It's whether 20-year demand assumptions pay off-and who eats the risk if they don't.
How the structures work
SPV financing routes debt through an external entity that builds the infrastructure and leases it back to the sponsor. On the sponsor's books, you see lease liabilities and payments rather than direct project debt, even though the economic commitment is similar over decades.
Terms can include residual value guarantees and partial equity stakes in the SPV, which tighten the tie between sponsor and asset performance. For background, see SPVs explained by Investopedia and lease accounting under IFRS 16.
Who's using it-and at what scale
- Meta: Roughly $27bn of AI infrastructure borrowing held off balance sheet, with about $6.5bn in incremental financing costs to do so. A $30bn Louisiana data center was largely funded by private credit groups including Blue Owl, Pimco, BlackRock, and Apollo. Meta owns around 20% of the vehicle and provides a residual value guarantee.
- Oracle: Tens of billions routed through similar structures, including a ~$38bn package tied to its OpenAI partnership.
- xAI (Elon Musk): About $20bn raised with debt reportedly secured against Nvidia chips.
- Nvidia: In some cases has taken equity in customers who then purchase its hardware-supporting revenue while pushing liabilities elsewhere.
Why markets care
Forecasts are huge, and funding is following. Morgan Stanley sees hyperscalers issuing roughly $400bn in corporate bonds in 2026. JPMorgan estimates AI and data-center names now make up about 14.5% of its $10tn investment-grade bond index (~$1.5tn exposure). UBS counts roughly $450bn of private capital in tech infrastructure by early 2025.
Veteran investors flag the familiar trio of risks: leverage, complexity, and opacity. If AI returns trail the spend, the pressure will show up first in private credit vehicles and long-dated leases.
Balance sheets are still thick with cash
Across the largest platforms, debt metrics remain solid. Debt-to-capital sits near 8.3% for Nvidia, 10.3% for Alphabet, 27.9% for Meta, and about 83.9% for Oracle (still investment grade, but on negative watch). Only Oracle and Apple currently carry more long-term debt than cash and short-term investments.
Capex from Amazon, Alphabet, Meta, and Microsoft is projected to top $600bn in 2026, while combined operating cash flow could approach $700bn. Demand for compute looks stubbornly strong-cloud providers are reportedly still renting out six-year-old A100s.
Risk map for finance teams
- Demand risk: Data centers are underwritten on 20-year demand curves. If AI spend (Gartner pegs 2026 at ~$2.52tn, up ~44% YoY) stalls, lease coverage thins.
- Lease and recourse risk: Residual value guarantees and keepwell-style support tighten exposure beyond reported lease liabilities.
- Counterparty risk: Some marquee AI customers, including OpenAI, may not reach profitability until later in the decade.
- Refinancing risk: SPVs often depend on rolling private credit and ABS-like structures; higher base rates or tighter spreads can crimp project IRRs.
- Asset obsolescence: Rapid chip cycles can erode collateral value faster than modeled, stressing guarantees and lease coverage.
- Private credit contagion: Concentrated lenders across multiple AI SPVs can transmit stress if one large project underperforms.
- Accounting transparency: Disclosures vary; obligations can be economically debt-like even if labeled differently.
What to monitor through 2026
- Bond supply and spreads: Track hyperscaler issuance versus Morgan Stanley's ~$400bn estimate.
- Lease footnotes (ASC 842/IFRS 16): Look for term extensions, variable payments, and residual value guarantees.
- SPV documentation: Recourse, step-in rights, and equity kickers that change risk transfer.
- Compute pricing and utilization: GPU rental rates (A100s and newer) as a demand proxy.
- Rating actions: Especially on highly levered names (Oracle on negative watch) and core SPV lenders.
- Private credit flows: Fundraising and deal terms for data-center vehicles.
- Customer milestones: Profitability timelines for AI platforms that anchor long-term contracts.
- Power and permitting: Grid constraints and delays that push out revenue ramps.
Practical actions for CFOs, treasury, and IR
- Treat leases as debt in risk limits: Include present value of lease payments in leverage, coverage, and covenant headroom tests.
- Set hurdle rates with buffer: Add risk premia for technology obsolescence, refinancing, and power-cost volatility.
- Run downside cases: 20-30% below-plan utilization; slower price declines; higher maintenance capex; delayed customer go-lives.
- Negotiate protections: Cap residual value guarantees, add rate step-downs on extension, and secure termination/upgrade options.
- Avoid circular financing traps: Limit vendor equity or credit that artificially boosts near-term sales.
- Diversify funding: Blend unsecured bonds, leases, and project debt to reduce single-channel stress.
- Upgrade disclosures: Plain-English bridges from GAAP figures to economic obligations; help investors price risk without guesswork.
Bottom line
SPVs let Big Tech build faster without headline debt spikes. That's efficient capital allocation-until growth assumptions slip. Treat these commitments as debt-like, price the risk, and keep optionality high.
Further reading for finance teams: AI for Finance
Your membership also unlocks:
