A reader wrote in with a useful question: if an outside vehicle borrows to build an AI data center and leases it to a hyperscaler, who really carries the debt and the risk?
Rather than answer privately, we are sharing the framework with every reader. The legal debt can sit in the special-purpose vehicle while the economic obligation still follows the hyperscaler's lease, guarantees and customer economics.
SPVs are legitimate infrastructure-finance tools, not free risk transfer. A lease can move spending away from visible upfront capex, but it remains a contractual claim on future cash flow.
The structure in plain English
Private credit funds, bondholders and infrastructure investors provide debt and equity to a project company. The SPV owns the land, building, power, cooling and connectivity. The hyperscaler leases the capacity and often owns or controls the servers and accelerators inside it.
The structure can be a build-to-suit, sale-leaseback, joint venture or layered capacity agreement. In the last case, the hyperscaler may resell compute to an AI lab. That creates two underwriting questions: can the hyperscaler pay the property lease, and can the ultimate customer create enough value to use the capacity profitably?
What “off balance sheet” actually means
If the hyperscaler is not the primary beneficiary of the project entity under accounting rules, the SPV's bonds may not be consolidated as corporate debt. The hyperscaler may still sign a long-term lease, guarantee residual value, contribute equity, own the computing equipment and supply nearly all project cash flow.
Credit analysts therefore look beyond the corporate debt line. Fixed rent, termination payments and residual-value guarantees can behave like debt even when the legal label is different.
Hyperion shows where the risk goes
Meta agreed to lease the entire campus and provided a residual-value guarantee for the first 16 years of operations. It later disclosed approximately $12.31 billion of initial lease commitments, a declining residual-value-guarantee threshold of approximately $28 billion and maximum accounting exposure to loss of $45.95 billion.
The maximum exposure is not an expected loss. It shows why a 20 percent ownership stake does not necessarily mean 20 percent economic exposure. Legal debt, accounting debt and economic support are three different measures.
The bull and bear cases
Bull case
- Outside capital accelerates capacity without forcing the tenant to fund every building internally.
- Long-duration investors can match their capital to infrastructure cash flows.
- Specialist developers handle construction and operations.
- Project debt can isolate claims beyond negotiated leases and guarantees.
- Renewal options preserve flexibility as computing architecture changes.
Bear case
- Fixed rent creates synthetic leverage even when reported corporate debt looks lighter.
- Short firm leases may support much longer debt through refinancing assumptions.
- One campus can depend on one tenant, one lab or one workload family.
- Power delays and overruns can occur while interest keeps accruing.
- Specialized facilities can lose economic value faster than their shells depreciate.
The seven risks that matter
1. Utilization
The tenant owes rent even if AI customers do not consume the capacity. A strong lease moves first-loss utilization risk to the hyperscaler, not out of the system.
2. Lease and debt mismatch
Compare the non-cancelable lease term with debt maturity and amortization. Optional renewals are not committed rent.
3. Residual value
If the tenant leaves, can another customer use the site without a major retrofit? Power, cooling flexibility, fiber and local market depth decide the answer.
4. Power and completion
Announced megawatts are not financed megawatts. Financed megawatts are not interconnected, commissioned or utilized capacity.
5. Concentration
Several contracts can still depend on one underlying demand factor. Solvent tenants can renegotiate or abandon marginal expansion.
6. Economic obsolescence
A functional facility can become uneconomic if newer campuses offer cheaper power, better cooling or higher rack density.
7. Accounting opacity
Uncommenced leases are generally disclosed as commitments before they become recognized lease liabilities. The debt line can lag the economic decision by years.
Leading indicators
The scale investors should not ignore
These figures are undiscounted, contain different asset mixes and should not be added together as though they were directly comparable corporate debt. They are shown because they represent future claims on cash flow that a capex-only analysis misses.
What changes the thesis
More constructive
- Firm leases cover debt repayment.
- Debt amortizes with limited balloons.
- Parent guarantees are explicit.
- Meaningful equity absorbs completion and residual risk.
- Power is secured before major spending.
- Commissioned capacity converts into revenue.
- Facilities can support several tenants and technology generations.
More cautious
- Commitments outgrow AI gross profit and free cash flow.
- Debt maturity extends well beyond firm rent.
- Projects finance before power or demand is binding.
- Customers repeatedly cancel, relocate or retrade capacity.
- Utilization and GPU pricing weaken as supply accelerates.
- Refinancing spreads widen or equity requirements rise.
- Hyperscalers sublease capacity or impair specialized sites.
Otto's verdict
SPVs extend the financing runway for the AI infrastructure buildout. They do not make the buildout capital-light.
The right question is not where the debt is reported. It is who must keep paying if utilization disappoints, and what the asset is worth if that party stops.
What should we explain next?
This issue exists because a reader emailed a question that deserved a shared answer. Tell us what you want more of, what you want less of, which formats help, what remains confusing or which subject you want unpacked next. A sentence is enough.
Email OttoPrimary sources
- Meta and Blue Owl announce the Hyperion joint venture, October 21, 2025.
- Meta 2025 Form 10-K, Hyperion, leases and VIE notes.
- Meta Q1 2026 Form 10-Q, lease commitments.
- Microsoft Q3 fiscal 2026 Form 10-Q, lease commitments.
- Alphabet Q1 2026 Form 10-Q, leases and VIEs.
- S&P Global Ratings on Meta's data-center joint venture, October 2025.
- Fitch Data Center Securitizations Rating Criteria, September 16, 2025.