How to Invest in GPU Loans: Four Routes for Institutional Capital

Summary

• Institutions invest in GPU loans four ways: direct origination, a position at origination or in a syndicate, a seasoned loan, or a rated facility.
• In every route, contracted offtake revenue repays the loan and the GPUs, held in a special-purpose vehicle (SPV), secure recovery. Pricing follows the offtaker's credit.
• Diligence covers offtaker credit and jurisdiction, concentration, deployment and hardware value; a buyer should get a complete credit file and servicing that stays in place for the full term.

Introduction

Institutions that want to invest in GPU loans are entering a credit market still being built around a very large capital need. In April 2025, McKinsey projected that data centers will need about $6.7 trillion of capital spending worldwide by 2030, $5.2 trillion of it for AI. We expect a growing share of that spending to be financed with debt secured by the hardware itself.

Holding GPU credit has usually required a GPU underwriting team that many asset-backed and private credit mandates don't have. Many deals also sit below hyperscaler scale, in the Compute Middle Market of GPU cloud operators between startups and hyperscalers.

This guide covers the four routes in, what each position holds, what to check and what to expect as a buyer. It covers getting exposure and leaves aside who finances GPUs.

Four Routes to Invest in GPU Loans

You can invest in GPU loans four ways: originate, enter at closing or via a syndicate, buy a seasoned loan, or hold a rated facility. Originating means building the capability yourself, while a position at closing puts you in a new financing beside other capital. A seasoned loan already has servers online and a payment record, and a rated facility rests mainly on a large operator's investment-grade customer.

The choice comes down to how much deployment risk you carry and how much you build in-house.

Originate directly

Direct origination gives you full control and makes you responsible for sourcing, hardware valuation, the security package and servicing. Those costs barely change with deal size, so direct lenders concentrate on large deals. In the Compute Middle Market, that fixed cost of making each loan is the main constraint on lending.

Direct origination also lets you set the structure, including corporate credit alongside the collateral. IREN's 10-Q describes its $363 million GPU Financing 3.0 as recourse to IREN under guarantees, so its lenders hold IREN's corporate credit alongside the collateral.

Take a position at origination or join a syndicate

A position at origination puts you in the loan at closing, so you carry deployment risk until the servers are online. Lenders usually hold the capital in escrow until the servers are installed, burned in (run under load to catch early failures) and independently verified online. Joining a syndicate works the same way for part of a larger facility.

One public example is the $106 million NexGen Cloud facility arranged by USD.AI in October 2026, a four-year amortizing senior secured loan backed by 1,728 NVIDIA B300 GPUs in Finland. It is fully contracted under a take-or-pay agreement, so the customer pays for reserved capacity whether or not it uses it. Silver Point Capital and Victory Park Capital took positions at origination.

Buy a seasoned loan

A seasoned loan takes construction and delivery risk out of your position, because the hardware is earning before you buy. In a season-and-sell, the originating lender funds the loan and holds it while the hardware is installed, verified and brought online. It then sells all or part of the loan to an institutional buyer and recycles the capital into new financings.

One public example is USD.AI's sale of the $98.1 million Duos Edge AI facility, originated in June 2026 and sold at GLX's October launch once it was seasoned and performing. The USD.AI loan history lists every publicly announced GPU loan originated through USD.AI.

Hold a rated facility

Rated facilities suit large tickets, and their ratings rest on an investment-grade offtaker, the customer contracted to buy the compute. CoreWeave's DDTL 4.0, a delayed-draw term loan (DDTL) drawn in stages after closing, closed in March 2026. CoreWeave's March 2026 announcement reports A3 (Moody's) and A (low) (DBRS) ratings for a facility "secured by HPC infrastructure and an associated customer contract."

IREN reported that its $3.65 billion financing, anchored by Microsoft's offtake, closed in June 2026 rated A by Fitch and A(low) by DBRS. Its fixed rate is equivalent to SOFR (the Secured Overnight Financing Rate, the main US dollar benchmark) plus 2.13%. Its floating tranche pays SOFR plus 2.25%.

The trade-off is tighter spreads, concentration in one offtaker and large tickets. In our view, the rated route is mainly exposure to the offtaker's credit, and the middle market is where the loan structure does the work.

What a GPU Loan Position Holds

In a standardized middle-market GPU loan, you hold a first-priority claim on a ring-fenced SPV's GPUs, offtake contracts and revenue accounts. Contracted offtake revenue repays the loan, and the hardware is the recovery source. The package typically includes the following, as set out in USD.AI's published loan structure of July 2026:

• Bankruptcy-remote SPV: A company holding only the deployment, with an independent manager, so the operator's other creditors can't reach its assets.
• Assigned offtake contracts: The offtake contracts are assigned to the SPV.
• Controlled accounts: SPV bank accounts sit under a deposit account control agreement (DACA), which gives the lenders control over the cash.
• Lien and equity pledge: A first-priority lien on the GPUs and an SPV equity pledge, perfected by UCC-1 filing (a public lien notice).
• Step-in rights: The colocation provider signs a lien waiver, giving up its own claim on the servers and letting lenders step in at the facility.
• Debt-service reserve: Months of debt service funded at closing, so a late customer payment doesn't become a missed loan payment.

For a fuller walk-through, see how GPU financing works. Middle-market loans can carry the same security package as larger deals; what changes is who pays under the offtake contract.

Pricing follows the offtaker's credit: CoreWeave's DDTL 4.0 priced its floating tranche at SOFR plus 2.25% in March 2026. Under the earlier DDTL 2.0, CoreWeave's March 2025 prospectus set the spread at 13.00% for non-investment-grade contracts.

For a mid-size comparator, Sharon AI's 8-K, dated September 2026, covers a facility of up to $356 million in two tranches. Facility A starts at 7.25% and steps up to 9.95%, and Facility B is fixed at 9.95%. The company's announcement of the facility names Goldman Sachs among its investors, alongside private credit funds.

What to Diligence Before You Invest

The customer's credit and the terms of the offtake contract carry most of the review, because contracted revenue repays the loan. Four checks cover the full review, with the hardware mattering most in recovery:

• Offtaker credit and jurisdiction: Who pays, contract term against loan term, termination rights and where the offtaker is based.
• Concentration: How much revenue depends on one customer, at the SPV and at the operator.
• Deployment and execution: Whether servers are delivered, accepted and online before your capital is at risk.
• Hardware value: What the GPUs earn today and what they could fetch in an enforcement.

One customer accounted for 67% of CoreWeave's 2025 revenue, according to CoreWeave's 2025 10-K. On deployment, IREN's 10-Q says its draws are funded upon acceptance of the GPUs, the same logic behind escrow. In October 2026, HIVE Digital Technologies announced BUZZ HPC's loan of about $129 million to buy an NVIDIA GB200 cluster for a three-year contract. Delivery of the GPUs was expected toward the end of Q4 2026.

Server lives in 10-K filings are accounting estimates, and Meta uses five to 5.5 years for servers and network assets. Amazon cut a subset of its servers and networking equipment from six years to five, citing the pace of AI development. Oracle moved from five years to six, and CoreWeave uses six years for computing equipment.

Earnings are easier to observe: Silicon Data's neocloud H100 index was $2.99 per GPU-hour on November 6, 2026. Public residual-value data is thin, and the clearest filing datapoint is a contract term. One lease in IREN's FY2025 10-K sets a lessee purchase option at the lower of fair market value and 18% of cost. Where the deal includes it, a residual-value warranty covers part of a shortfall if servers sell below an agreed value, subject to its terms.

In an illustrative example, a $6.0 million deployment is financed with a $4.8 million loan at 80% and $1.2 million of operator equity. A three-month debt-service reserve is funded at closing, and straight-line amortization over three years cuts principal by about $133,000 a month. About $3.2 million remains after 12 months, and the loan is fully repaid at month 36, inside the five-to-six-year accounting lives above.

What to Expect as a Buyer

Whoever sells you the loan, ask for a complete credit file and servicing that stays in place for the full term. The file should contain:

• Underwriting assessment: The lender's credit analysis and how it set the loan's pricing, size and structure.
• Structure: The ring-fenced SPV, the lender group's first-priority lien, the coverage tests and how the cash is monitored.
• Funding conditions: The conditions the deployment had to meet before capital left escrow.
• Loan details: The principal, rate, maturity and GPUs securing the loan.

A dedicated servicer should manage the loan for its lenders for the full term, with monitoring of uptime and server health. If the loan defaults, the lenders should be able to step in and enforce against the collateral.

Funds use the credit package to support their review of seasoned loans within their mandate, as set out in GPU credit for private credit funds. Banks examine the documentation, structure and servicing behind GPU loan positions for banks against their credit policies, and each determines its own regulatory treatment. Family offices considering GPU credit for family offices generally start with the asset, the structure and the ongoing reporting.

Conclusion

Choose your route by how much deployment risk you want to carry and how much you want to build in-house. In every route, diligence the offtaker, concentration, execution and hardware value, which decide how the loan repays and what lenders recover.

USD.AI is built for the Compute Middle Market, which we put at $50 million–$250 million per deal for financing through USD.AI. The Compute Middle Market tracker uses a wider band, counting GPU-backed financings of $500 million or less. As of November 3, 2026, it showed 46 financings since January 2023, totaling $5.04 billion from 23 named lenders, with a $44.2 million median.

GLX, USD.AI's institutional marketplace, lets you take a position at origination, join a syndicate or buy all or part of a seasoned loan without building a GPU underwriting team. Every financing is scored by Compute IQ, which rates the operator, the customer's credit and the contract on a 100-point scale and sets the pricing tier, size and structure. Buyers get that assessment in the credit file, servicing stays in place after a sale, and Aravolta's telemetry monitors each deployment.

By November 3, 2026, $531 million of GPU financing had been originated through USD.AI across 20 financings, with more than $200 million through GLX in its first month, and no USD.AI loan has missed a payment as of November 2026. Opportunities are shared with qualified institutional counterparties under NDA, so Talk to USD.AI about your mandate.

FAQ

What is the main collateral in a GPU loan?

The main collateral is the SPV's GPUs, offtake contracts and revenue accounts, plus a pledge of the SPV's equity. The offtake contract is the primary source of repayment, and the GPUs are the source of recovery if the loan defaults.

Can institutions buy GPU loans that are already performing?

Yes. Some lenders sell seasoned, performing loans to institutional buyers once the hardware is online; USD.AI does this through GLX, and servicing stays in place for the full term after the sale.

Do GPU lenders get a parent guarantee?

It varies by deal: IREN's 10-Q describes its GPU financing as recourse to IREN under guarantees. Financings through USD.AI are structured around the SPV. In the Duos Edge AI facility originated through USD.AI, Duos pledged its equity in the SPV and is liable only for specified bad acts.

How much of a deployment does a GPU loan through USD.AI finance?

A GPU loan through USD.AI finances up to 80% of the deployment cost, and the operator contributes at least 20% as equity.

Editor Notes

1. Link the GPU private credit article for the full CoreWeave DDTL spread table once it is published (not yet live, so no link added).
2. HIVE/BUZZ cites HIVE's 8-K Exhibit 99.1 on sec.gov; the rate is held back.
3. "GLX's October launch" now links the GLX announcement post instead of the October recap (which covers CHIP buyback content).
4. Held back: the GLX page's lender IRR range, average DSCR and blended interest rate; per-loan rates for NexGen and Duos; escrow agent and residual-value warranty provider names; GLX live counters; CoreWeave DDTL 5.0 parent guarantee (pending credit-agreement confirmation); IREN's stated rate on GPU Financing 3.0. Ratings are attributed to the issuers.