AI's Hidden Price Tag: Record Debt and Off-Balance-Sheet Financing
AI is changing how the data centre build-out gets financed. Debt is now the default, and a growing slice of that debt lives off corporate balance sheets. For finance teams, the headline is simple: leverage is rising, transparency is thinning, and the refinancing cycle will arrive faster than you think.
Debt becomes the default
Big Tech used to run fortress balance sheets, funding growth from cash flow. That's flipped. Global tech companies issued a record $428.3bn in bonds through early December 2025 to fund AI data centres, shifting from cash-funded capex to market-funded capex.
Leverage is moving up. Across roughly 1,000 tech firms over $1bn in market cap, median debt-to-EBITDA reached 0.4 by September 2025-nearly double the pandemic-era spike in 2020. The driver: AI infrastructure ages fast, and accelerator refresh cycles force continuous reinvestment.
How off-balance-sheet SPVs work
To mute reported leverage, companies are pushing billions into special-purpose vehicles (SPVs). More than $120bn in AI data centre spend has been shifted off balance sheets via SPVs owned by investors, while the tech companies sign long-term capacity leases.
In these structures, the SPV owns the land, buildings, power gear, and often the GPUs. Institutional investors (e.g., Pimco, BlackRock, Apollo) and private credit funds (e.g., Blue Owl) provide debt and equity to the SPV. The tenant gets capacity without booking traditional debt, preserves ratings metrics, and keeps optionality.
Case studies: Meta, xAI, CoreWeave
Meta's ~$30bn Hyperion project used an SPV (Beignet Investor) backed by Blue Owl. Roughly $27bn came as loans from investors including Pimco, BlackRock, and Apollo, plus ~$3bn in equity from Blue Owl. The obligation didn't show up as debt on Meta's balance sheet, which helped clear the way for a subsequent ~$30bn corporate bond raise.
Meta still has skin in the game-owning ~20% of the SPV and covering losses under certain conditions if the lease isn't renewed and asset values fall. Similar structures are underway at xAI (raising ~$20bn to buy GPUs and lease them back) and CoreWeave (SPVs tied to ~$11.9bn in compute contracts with OpenAI).
Private credit steps into the spotlight
Private credit has become a core pillar of the AI build. By early 2025, tech companies had borrowed roughly $450bn from private funds, up about $100bn in a year. For tenants, the appeal is flexible, large-scale capital without headline leverage. For investors, the pitch is infrastructure-like cash flows backed by blue-chip counterparties.
The risk: demand concentration. If AI workloads underdeliver, the pain concentrates in a small customer set and specialized assets with uncertain resale values. Equity markets are already questioning timelines-shares of some AI-exposed names, including Oracle, have faced pressure as return visibility lags spend.
Why finance leaders should care
- Look-through leverage: Off-balance-sheet SPVs mask economic leverage. Include lease liabilities, purchase options, residual guarantees, and take-or-pay terms in your effective net leverage math.
- Consolidation risk: Assess whether SPVs trigger consolidation (ASC 810 VIE tests or IFRS 10). A shift in control or guarantees can pull debt on balance sheet at the worst time.
- Refinancing and rate risk: Many SPVs will need refis within 3-7 years. Model spreads, base rates, and covenant headroom through a higher-for-longer scenario.
- Asset obsolescence: GPU refresh cycles can be 2-4 years. Residual value risk sits with someone-know if it's you via residual guarantees or upgrade obligations.
- Counterparty concentration: A handful of tenants and workloads drive cash flows. Stress test utilization, renewal probabilities, and step-down rents after initial terms.
- Power and uptime: Capacity means nothing without power. Scrutinize PPAs, grid interconnect timelines, curtailment clauses, and penalties for downtime.
- Disclosure quality: Off-balance-sheet commitments, lease tables, and related-party SPVs deserve a fine-tooth review. Inconsistent footnotes are a red flag.
- Exit liquidity: If demand softens, who buys used GPUs or stranded capacity? Secondary markets for accelerators are unproven at scale.
What to scrutinize in filings and deal docs
- Lease economics: Term, escalators, capacity ramp schedules, take-or-pay minimums, and renewal options (including penalties if not exercised).
- Credit protection: Parent guarantees, keep-well agreements, residual value guarantees, and step-in rights for lenders.
- Replacement capex: Contractual obligations to upgrade chips; timing and who funds the swap cycle.
- Hedging: Interest rate policy at both corporate and SPV levels; commodity hedges tied to power costs.
- PPA alignment: PPA tenor vs. lease tenor; curtailment risk; pass-through mechanics for energy price spikes.
- Look-through metrics: Fixed-charge coverage including lease payments; effective WACC including SPV debt; lease-adjusted EBITDA and leverage.
- Disclosure gaps: Off-balance-sheet notes, related-party transactions, and VIE determinations; rating agency commentaries vs. company narratives.
Implications by seat
- CFO/Treasury: Treat SPV liabilities as quasi-debt in planning. Lock in staggered maturities, negotiate early upgrade price curves, and pre-wire refi options.
- Equity PMs: Build scenarios where AI unit economics lag 12-24 months vs. capex. Re-rate on look-through leverage and renewal risk, not headline cash balances.
- Credit investors: Underwrite to lease-adjusted coverage and tenant dependency. Favor structures with strong residual protections and step-in rights.
- Banks/Private credit: Tighten covenants around utilization, upgrade schedules, and PPA milestones. Monitor chip supply cycles and resale channels.
Policy and disclosure
Off-balance-sheet financing isn't new, but the scale and asset specificity are. Regulators and standard setters should push for clearer, comparable disclosure of SPV exposures, residual guarantees, and look-through leverage. Consistency across lease accounting, VIE analysis, and rating methodologies would reduce blind spots for investors.
Upside vs. downside
Upside: AI demand meets forecasts, GPUs hold value better than feared, and SPV financing lowers WACC while protecting ratings. In that case, tenants enjoy capacity on call and investors clip steady coupons.
Downside: Workloads ramp slower, renewal rates soften, and refis bite just as upgrade cycles accelerate. Residual risk moves from footnotes to P&L, and structures that once hid leverage start to reveal it.
30-day action checklist
- Map all SPVs and leases tied to AI capacity; compute lease-adjusted leverage and fixed-charge coverage.
- Run a refi stress at +200-300 bps and a 12-24 month delay in AI revenue ramp.
- Inventory residual guarantees and upgrade obligations; quantify worst-case cash calls.
- Audit PPA and interconnect timelines vs. capacity go-live; flag timing mismatches.
- Engage rating agencies with look-through metrics to avoid surprise outlook changes.
Bottom line
AI infrastructure is capital hungry, fast aging, and now financed with a mix of bonds, leases, and SPVs. If you ignore the look-through leverage, you'll misprice risk. If you price it well, you can still fund growth-without handing your future to the next refi window.
If you're mapping practical AI use cases to finance workflows, this curated list may help: AI tools for finance.
Your membership also unlocks:
