How GPU Financing Works at USD.AI

How Independent Institutions Connect Onchain Capital to Physical AI Infrastructure

A company has signed a contract to provide a customer compute. The customer is ready. The data center is ready. The purchase order for the GPUs is ready.

There is only one problem: the company needs millions of dollars to buy the machines.

This is why companies come to USD.AI.

From the outside, these transactions look simple. A company wants GPUs, USD.AI provides the loan, and then the operator pays back the loan over time with interest.

In reality, the GPUs are only one part of the transaction. The real challenge is underwriting the broader set of risks: the borrower’s financial condition, operational risk, the customer contract supporting the deployment, the data center and operating arrangements, the expected revenue and cash flows, and the legal and structural protections that connect all of these elements throughout the life of the loan.

This article examines USD.AI’s underwriting, structuring, and execution methodology, purpose-built to fund the AI capex boom. For a deeper look at borrower diligence, collateral standards, loan terms, and ongoing risk monitoring, read Underwriting and Risk Management.

The Life of a USD.AI Loan

All loans originated through USD.AI follow a similar process:

1. A GPU operator, the “OpCo,” wants to buy GPUs, so they apply for USD.AI financing. The first step is to submit documents for underwriting and ensure the transaction meets USD.AI’s requirements. These include an approved hardware purchase order, the required equity contribution, an executed colocation agreement, and either an executed offtake agreement or qualifying evidence of on-demand revenue.
2. A dedicated borrower SPV is established beneath a newly formed Parent HoldCo, each a wholly owned subsidiary. The intervening Parent HoldCo makes the SPV bankruptcy-remote, so an OpCo insolvency does not stay enforcement of USD.AI's pledge over the SPV equity.
3. The financed GPUs, offtake agreement, colocation agreement, and revenue accounts are assigned to the SPV structure.
4. The SPV places the approved order with an OEM or authorized supplier.
5. USD.AI establishes its first-priority claim over the GPUs and all material contracts of the SPV.
6. USD.AI places the committed loan funds into escrow with Wilmington Trust.
7. The OEM builds and ships the approved servers to the designated data center.
8. The servers are installed, independently verified, and brought online.
9. Once the release conditions are satisfied, escrow pays the OEM or authorized supplier.
10. The Offtaker makes contracted payments into the SPV’s controlled revenue account.
11. The SPV pays principal and interest to USD.AI.
12. Remaining revenue may then be distributed to the OpCo.

The structure prevents any single party from controlling the entire transaction. The borrower cannot unilaterally redirect the purchase funds, substitute the collateral, move the equipment, or divert the deployment’s revenue.

The individual components are examined below to understand how they integrate into a cohesive structure.

The SPV

Consider the first problem: USD.AI needs to know exactly what it is financing and what it can recover if the borrower defaults.

A GPU operator cannot simply borrow $50 million from USD.AI as an unsecured corporate loan. If the money were commingled with the operator’s other funds, and the GPUs were purchased alongside its other assets, USD.AI would end up competing with other creditors in the event of insolvency. The lender would have no clean path to recovering the GPUs and the related contracts that support their operations.

Defaults could be messy. Typically there are multiple claims against the distressed balance sheet. A data center might be owed back rent, a contractor unpaid for services rendered. All of these counterparties become creditors in a bankruptcy.

As a lender, USD.AI must ring-fence the collateral supporting its loan so that the overall structure is bankruptcy-remote. A bankruptcy of the OpCo should not disturb the loan, the operation of the GPUs, the offtaker revenues, or the tenancy at the data center. Separating the collateral from the operator's broader business makes it easier to enforce against.

So the first step in ring-fencing is establishing a dedicated special-purpose vehicle, or SPV.

USD.AI holds a first-priority claim over the relevant SPV assets and the equity in the SPV. The SPV is also restricted from taking on unrelated debt or voluntarily entering bankruptcy without independent consent. What this means is that all revenue the GPUs create is routed through the SPV, and USD.AI has first claim for repayment of the loan.

If the borrower defaults, the lender has a clear path to take control of the specific machines and the economics attached to them.

The Offtake

An offtaker is a customer that agrees to purchase compute generated by the financed GPUs. Depending on the deployment, that customer may be a hyperscaler, inference provider, AI model company, enterprise, or another cloud platform requiring dedicated capacity.

The offtaker is the primary source of credit support for the senior financing, as the offtaker is the party ultimately making rent payments that go toward paying off the loan. USD.AI reviews offtaker financials, backers, and financial health metrics, including cash runway and leverage levels, to determine the offtaker’s ability to perform on its contractual obligations. Then, it reviews the offtake contract itself to ensure there are no “outs” for the offtaker to break the contract. The combination of the offtaker’s credit profile and the offtake contract’s legal strength make up ~70% of the overall underwriting and approval process.

The Data Center

GPUs have to live somewhere. For USD.AI financings, GPUs have to be located in a Tier 3 or better data center, which represents the highest levels of security, redundancy, and power requirements. This requirement is established by property and casualty insurers, who require this level of tiering to place insurance over the cluster.

State-of-the-art data centers are purpose-built to house millions of dollars of financed chips. Their physical-security controls typically include biometric access, mantrap entry systems, continuous surveillance, and restricted access zones. In an event of default, USD.AI has contractual step-in rights with the data center operator: it can assume performance under the colocation agreement and take control of the GPUs. The operator separately waives any lien over the hardware, so nothing sits ahead of USD.AI's claim. Keeping the operator paid matters as much as holding the hardware. Powered rack space is a major bottleneck in AI expansion, and GPUs that lose their slot are less valuable than ones still running.

The OEM

When USD.AI lends millions of dollars for GPUs, the operating company never actually touches that capital. Instead, the funds are sent to the OEM, such as Dell, Lenovo, or Supermicro, to purchase built servers.

USD.AI finances the entire server rack, including memory, storage, cables, switches, and all other parts of the server. OEMs naturally build their servers with differing internal hardware components.

Wilmington Trust: Restricting Capital Until the GPUs Are Online

USD.AI uses Wilmington Trust, a subsidiary of M&T Bank, as its institutional escrow agent. Large GPU purchase orders require credible proof that committed capital is available before an OEM begins building and shipping the systems.

When placing an order with an OEM, the operating company pays 20-30% as a down payment on its purchase. When the chips are ready for use, USD.AI releases the remaining 70-80% from escrow to the OEM.

Under USD.AI’s arrangement, if the transaction does not fund within a permitted longstop period (120-150 days), the escrowed capital and applicable interest are returned to the protocol.

More detail is available in USD.AI’s explanation of the Wilmington Trust escrow arrangement.

Independent Verification and Monitoring

USD.AI monitors loans throughout their entire lifecycle using telemetry devices installed at the datacenter which provide uptime and server health metrics in real time, enabling USD.AI to independently verify servers are active and in good working order.

This article describes USD.AI’s general published framework. Individual loan terms, counterparties, collateral packages, monitoring arrangements, and enforcement rights may vary. Asset-backed lending involves risk, including borrower default, collateral depreciation, counterparty failure, enforcement delays, and losses during liquidation.