GPU financing for contracted AI infrastructure

Because the contract is the source of repayment, its terms carry more weight than any other document in the financing. Lenders look first for a non-cancellable term at least as long as the debt, and for take-or-pay payments, which are owed for reserved capacity whether or not the customer uses it. They then look at the customer. An investment-grade or well-capitalized customer can carry a financing on its own credit, while a younger customer usually needs a parent guarantee, a letter of credit or a larger prepayment.

Three further terms often decide the outcome. The customer must consent to the contract being assigned to lenders as security. Service credits for downtime must be capped below the level that would impair debt service. And if the customer terminates early, the termination payment should be large enough to repay the debt still outstanding.

These terms are negotiated for commercial reasons, and they are difficult to add once a contract is signed. A contract can be long, creditworthy and fully priced, and still support less debt than the operator planned for. The practical conclusion is to test the draft against a lender’s requirements before signature. The full list is on the compute financing page.

The deployment is usually placed in a project company that owns the GPUs and holds the customer contract, separate from the operator’s other business. Institutional investors lend to that company on a senior secured basis, with a first lien on the equipment, an assignment of the contract and control over the accounts into which the customer pays.

Customer payments then flow through a waterfall. Power, colocation and other operating costs are paid first, followed by interest and principal on the debt and any required reserves. Only then does the remaining cash reach the operator. A debt service reserve, typically three months of payments and up to six for weaker credits, is funded at closing to absorb short interruptions.

The debt amortizes from the contract’s cash flow and is scheduled to reach zero before the contract ends, so repayment does not depend on renewing the contract or selling the hardware. The operator contributes its own equity, at least 20% of equipment cost, which sits beneath the debt and absorbs losses first.

The proceeds can pay for GPU servers, networking and storage, installation and commissioning, and the reserve. Where agreed, part of the customer’s prepayment can also be financed, and bridge or vendor financing used before signature can be refinanced once the contract is in place.

Lenders size the debt twice and lend the lower amount. The first test is a share of equipment cost, typically 65% to 80%. The second is the amount the contract’s net cash flow can repay within its term with room to spare, measured as debt service coverage of at least 1.25x. On a three-year contract, the second test usually decides.

In this example the contract supports about $78 million, slightly below the $80 million cap, so the contract rather than the hardware sets the size of the loan. A longer term, a larger customer prepayment, a stronger customer or value insurance on the hardware would each increase it.

• The customer. Its credit quality, its total compute commitments across providers and any parent or letter-of-credit support.
• Net cash after costs. The contract’s payments after power, colocation and operating costs, not the headline contract value.
• Coverage. Net cash that covers debt service by at least 1.25x in every period.
• The operator. Its record of deploying and running GPU capacity, and whether a replacement operator could step in.
• The site and power. Contracted power, cooling and colocation for the full term.
• The hardware. NVIDIA or AMD systems, their delivery schedule, warranties and expected value over time.

Corporate credit is lent against the whole company’s balance sheet. It is the cheapest route for the largest buyers of compute, but most operators cannot reach it. Equipment leasing is priced mainly on the hardware and the lessee, which suits smaller clusters but limits the amount to what a lessor will advance against the equipment’s value.

Contracted GPU financing lends against a specific deployment and its customer contract. It is most useful when the customer is a stronger credit than the operator, because the customer’s credit, not the operator’s, carries the loan.

Park Street Global structures GPU financings and arranges them with institutional credit investors in the United States and Europe. We do not lend our own money, so the size of a financing follows the transaction rather than our balance sheet.

We are most useful early. Before signature, we read the draft contract, mark the terms the institutional capital will need and propose language for each. After signature, we design the structure, prepare the credit materials and bring the transaction to investors.