Big Tech's AI SPVs: $120B Off Balance Sheet and a Growing Private Credit Risk
U.S. tech giants are pushing AI build-outs through special purpose vehicles, shoving as much as $120 billion of liability off their balance sheets. It keeps headline leverage clean, protects credit ratings, and shifts risk to asset-backed structures funded by private credit.
The setup is simple on paper: an SPV finances data centers, GPUs, and power-hungry infrastructure; investors buy its debt or equity; the operating company leases capacity without booking the full liability. The risk, of course, doesn't disappear - it just moves.
Who's using SPVs - and at what scale
According to reporting cited by the Financial Times, Oracle has leaned in the most, raising about $66 billion via SPVs to fund data centers in Texas, Wisconsin, and New Mexico. Each SPV owns the assets; Oracle leases them.
Meta created an SPV dubbed "Beignet Investor" to fund its Hyperion data center in Louisiana, stacking $30 billion in total financing from Pimco, BlackRock, Apollo, and Blue Owl. Elon Musk's xAI used an SPV to secure $20 billion for Nvidia GPUs, while CoreWeave raised $2.6 billion through an SPV tied to an OpenAI contract.
Non-recourse is the key phrase. If the deals sour, lenders typically have claims on sites, facilities, and chips - not the parent's balance sheet. It's clean until it isn't.
Why it matters for finance teams
Off-balance-sheet treatment flatters leverage, interest coverage, and ROIC optics, but can hide economic exposure. Lease commitments, take-or-pay contracts, capacity guarantees, or performance obligations can still behave like debt under stress.
Meanwhile, Wall Street's exposure grows. UBS estimates private credit tied to big tech has reached about $450 billion, up roughly $100 billion year over year, with about $125 billion flowing into long-term project finance like data centers. The broader private credit market is now near $1.7 trillion.
Concentration risk is not theoretical
AI demand is led by a handful of buyers and counterparties. One data point making allocators uneasy: OpenAI has reportedly signed more than $1.4 trillion in long-term compute commitments. If a single anchor customer stumbles, the shock can cascade across SPVs, funds, and counterparties at the same time.
There are other pressure points: power constraints, policy shifts, and fast chip obsolescence - the very assets backing the debt. If valuations reset, recoveries will, too.
Not every giant is using SPVs
Some operators are building with cash or direct debt - notably Google, Microsoft, and Amazon. That reduces optical risk transfer but keeps terms simple and disclosures cleaner.
How the structures work (and fail)
- Structure: SPV owns the site, shells, power gear, and semis; investors fund via secured debt and equity; the tech company leases capacity.
- Accounting aim: keep SPV debt non-recourse and outside consolidation; book lease expenses instead of full project debt.
- Failure mode: if demand or pricing drops, SPV cash flows crack first; lenders seize assets with uncertain resale value and short tech half-lives.
- Contagion path: multiple AI SPVs mark down at once, hitting private credit funds with correlated exposures and similar underwriting.
What to review on your side
- Consolidation risk: variable interest entity triggers, guarantees, residual interests, or control rights that could force consolidation later.
- Hidden leverage: total lease and offtake commitments, step-up clauses, make-wholes, and cross-default linkages with corp-level facilities.
- Asset risk: chip depreciation curves, upgrade cycles, and secondary market values; power procurement terms and curtailment exposure.
- Counterparty maps: reliance on a few AI tenants or model providers; concentration in single funds, managers, or structures.
- Funding fragility: floating-rate tranches, refi walls, build-risk during construction, and covenant tripwires tied to utilization.
- Disclosure quality: segment reporting, lease detail, related-party exposure, and SPV-level performance metrics.
Risk scenarios to model
- Utilization slips 10-30% while power and labor costs rise; test lease coverage, DSCR, and equity cures at the SPV level.
- GPU values fall with a new chip cycle; haircut collateral and re-run recoveries on non-recourse facilities.
- Private credit spread widening of 150-300 bps; assess refi feasibility and knock-on effects to parent cash flows.
- Single-name tenant shock (e.g., delayed rollout or funding stress) triggering simultaneous marks across peers.
Practical actions for CFOs, treasurers, and credit teams
- Inventory all SPVs and leases; build a consolidated "economic leverage" view that includes committed capacity payments and minimums.
- Push for non-recourse purity: narrow guarantees, avoid keepwells, and cap performance obligations that invite consolidation risk.
- Tighten counterparty risk: set limits per AI tenant and per private credit manager; require transparency on fund-level concentrations.
- Stage-gate capex: tie funding draws to pre-sold capacity and power availability; penalize delays.
- Stress-test disclosures: assume lower AI demand, higher cap rates for data centers, and faster chip obsolescence.
- Pre-wire refis early: hedge rates, diversify lenders, and avoid clustered maturities across SPVs.
The market read
For now, balance sheets and ratings at the largest platforms look solid. But the mix - long-lived assets, fast tech cycles, heavy private credit participation, and concentrated demand - is a fragile stack if conditions turn.
If SPV liabilities move in sync, even a small shock can ripple through funds and banks that financed the build. It's smart to assume correlation rises at the worst time.
Further reading
- Financial Times reporting on AI infrastructure financing
- UBS research on private credit and project finance flows
Useful resource
If your finance team is evaluating spend tied to AI adoption, these curated tools may help frame ROI and vendor risk: AI tools for finance.
Your membership also unlocks:
