Zettabyte
AI Economics

AI Data Center Financing: Buying, Leasing and Partner-Financed Infrastructure

The best AI data center financing structure depends on GPU utilization: buying or partner-financing generally suits predictable, sustained workloads, while flexible leasing suits uncertain demand.

Investor concern is growing. In Bank of America's October 2026 Global Fund Manager Survey, approximately 39% of fund managers identified AI capital expenditure as a potential source of a systemic credit event. Lenders therefore need to assess contracted demand, cash flow and residual asset values before approving financing.

Four Types of Data Center Financing Structures

The National Law Review outlines four major financing structures used in data center development:

  • Project finance: Establishing a special purpose vehicle (SPV) to develop the data center and securing debt against projected cash flows from long-term leases. Lenders may require completion guarantees, funding for construction overruns and other sponsor credit support.
  • Private credit: Private credit and infrastructure debt funds provide customized construction or bridge loans, often with greater flexibility but potentially higher interest rates and fees.
  • Capital markets and securitization: Issuing investment-grade or high-yield bonds, or securitizing assets and contracted cash flows. These structures can offer longer financing terms but require substantial documentation and disclosure.
  • JV and equity partnerships: Partnering with institutional investors who provide capital while the developer contributes development expertise and assets. This reduces direct capital requirements but introduces shared governance and negotiated exit rights.

Zettabyte's GPU financing service helps companies evaluate ownership, leasing and partner-capital options for their AI infrastructure deployments.

Buying vs. Leasing GPU Infrastructure

Buying outright works best for predictable, sustained workloads, but depreciation and hardware obsolescence remain important considerations. High-end NVIDIA H100 configurations can still cost over $30,000, despite newer GPU generations entering the market. Recent GPU resale data also shows that used HGX H100 servers have retained substantial value, although individual GPU prices vary by configuration and transaction type. Rapid GPU development creates uncertainty around long-term competitiveness even when existing hardware retains resale value.

Two key questions for lenders, buyers and auditors are:

  • What is the full cost of the data center buildout, including GPUs, networking, storage and facilities?
  • What are the useful life, depreciation schedule and expected residual value of the GPUs?

Leasing works best when GPU demand is uncertain or variable. Common arrangements include:

  • Dedicated GPU servers: Leasing physical servers equipped with GPUs.
  • Dedicated GPU clusters: Leasing interconnected servers for larger training or inference workloads.
  • GPU-as-a-Service: Paying for GPU capacity through on-demand or reserved cloud services.
  • Long-term leases: Reserving dedicated capacity for sustained workloads, often with more predictable pricing.

Partner-Financed AI Data Centers

Partner financing typically combines equity and debt. Long-term customer offtake agreements and creditworthy counterparties can support more favorable financing terms.

Meta has used two major structures:

Meta El Paso: In July 2026, Meta announced a joint venture with BlackRock for a 1 GW, approximately $14 billion data center development.

  • BlackRock-managed funds own 80% of the equity; Meta retains 20%.
  • Meta contributes approximately $2.3 billion in assets, while BlackRock contributes approximately $4.9 billion in cash, including funding supported by debt.
  • A $12.5 billion debt financing supports the transaction.
  • Meta leases the campus for an initial four years, with extension options providing up to 20 years.

Meta Hyperion: In October 2025, Meta and Blue Owl Capital announced an approximately $27 billion joint venture for the Hyperion data center campus in Louisiana.

  • Blue Owl-managed funds own 80%; Meta retains 20%.
  • Blue Owl contributed approximately $7 billion in cash, while Meta received a roughly $3 billion distribution.
  • Meta will operate and lease the facility, with contractual arrangements providing long-term support for financing.

These structures allow Meta to secure infrastructure capacity without directly funding the entire development.

Residual-value guarantees and other contractual protections can also support AI infrastructure debt, although their coverage and obligations vary by transaction.

Two GPU Financing Facilities Priced in 2026

1. CoreWeave DDTL 4.0

In March 2026, CoreWeave closed an $8.5 billion delayed draw term loan, the first investment-grade-rated HPC infrastructure financing secured by GPU infrastructure and an associated customer contract.

  • Facility: $8.5 billion, with draws linked to asset deployment and stabilization.
  • Ratings: A3 by Moody's and A (low) by DBRS.
  • Pricing: SOFR + 2.25% floating; approximately 5.9% fixed.
  • Maturity: March 2032.
  • Collateral: Substantially all assets of CoreWeave Compute Acquisition Co. VIII, LLC.

2. Lambda Term Loan B

In August 2026, Lambda closed a $926 million senior secured Term Loan B to finance GPU infrastructure supporting a committed customer deployment with an investment-grade offtaker.

  • Facility: $926 million.
  • Rating: Baa2 by Moody's.
  • Pricing: SOFR + 3.00%, issued at 99.5% of principal.
  • Maturity: December 31, 2030.
  • Repayment: Fully amortizing, aligned with contracted cash flows and the useful life of the GPU infrastructure.

Both transactions demonstrate how contracted GPU demand can support investment-grade infrastructure financing.

Why GPU Utilization Matters

GPU utilization determines how much of an infrastructure investment generates productive capacity. If a company purchases 500 GPUs but consistently uses only 200, 60% of its GPU fleet sits underutilized while still incurring financing, depreciation and operating costs.

The financial impact depends on demand certainty, contract duration, operating costs, financing terms and expected GPU residual value.

Five Underwriting Questions for AI Data Center Financing

Lenders must assess what supports repayment and what the underlying infrastructure will be worth if the original projections fail.

  1. What exactly is being financed? Land, grid connections, generation, buildings, mechanical and electrical systems, and IT equipment have different collateral values, depreciation curves and replacement cycles.
  2. What obligation makes the debt serviceable? Is repayment supported by a lease, offtake agreement, capacity reservation, parent guarantee or residual-value support agreement?
  3. What does the guarantee cover? What are its limits, covered assets, valuation methods, triggering conditions and exclusions? Are other lenders relying on the same guarantee?
  4. Who bears obsolescence risk? Will the hardware retain sufficient residual value, and is there a viable secondary market if the borrower defaults?
  5. What proof of power is financeable? Grid connections, contracted power, facility capacity and usable IT power are different measures. Financing assumptions must distinguish between planned and deliverable capacity.