4 Conditions for Reasonable Project Finance AI Data Center Loans

When lending billions of dollars for AI data centers, lenders look closer at the contracts than at the company’s brand name.

Because project finance AI data center loans are structured to use the equipment and the customer contracts utilizing that equipment as collateral, they can offer lower interest rates than general corporate bonds if the customer’s credit rating and contract term are well-aligned.

Conversely, if either of those two factors is misaligned, the story changes.

How Project Finance AI Data Center Loans Differ from General Corporate Bonds

General corporate bonds are loans granted based on trust in the company as a whole.

Investors view the company’s entire future profit as the source of repayment. Even if one business fails, the company can repay the debt with profits from other ventures, so lenders may provide funds without specific collateral. In return, if the credit rating is low, the interest rate rises.

Project finance is different.

Project finance refers to a loan where the cash flow of a single, specific project serves as the repayment source.

In the data center business, a separate legal entity (an SPV, or Special Purpose Vehicle) is established to hold the equipment and sign contracts with customers, and the loan is issued only to that entity.

While I refer to this as “project finance-style” in this article, it is an industry classification; CoreWeave press releases refer to it as ‘HPC infrastructure-backed financing.’

CoreWeave’s March 2026 press release illustrates this structure well.

The $8.5 billion DDTL 4.0 facility uses all assets of a separate entity called CoreWeave Compute Acquisition Co. VIII as collateral, and the company itself called it ‘first-of-its-kind non-recourse’ financing. Non-recourse means that if the loan is not repaid, the lender cannot claim assets from the parent company. Lenders essentially lent money relying only on that entity’s equipment and customer contracts, not on CoreWeave as a whole.

From a lender’s perspective, the difference lies in repayment priority and information.

Corporate bond investors judge based on the company’s entire financial statement, whereas secured lenders can track the cash flow entering that specific entity, allowing them to monitor contract performance more narrowly and specifically.

Conversely, corporate bond investors trust the entire company, so if one business falters, profits from other businesses act as a buffer. One approach achieves certainty by narrowing the scope, while the other gains a buffer by broadening it.

It is easy to compare this to housing.

An unsecured loan taken by an employee is based on their entire monthly salary. A loan taken by a landlord using a building as collateral relies only on the rent from that building. The landlord’s other financial circumstances matter less; the key is whether the tenants will consistently pay their rent. In data center loans, the role of the tenant is occupied by large customers using AI services.

This leads to the first point of the counter-argument. Even if the total debt appears large, not all of it represents a claim against the entire company. CoreWeave’s total debt as of the end of June 2026 was reported at approximately $35 billion, divided into about $31.4 billion in recourse debt and $3.7 billion in non-recourse debt. One must recognize from these figures that the non-recourse portion is still small. Stating that a structure is reasonable is different from saying the entire debt is safe.

Why are Delayed Draw Term Loans (DDTL) drawn down in stages to match equipment deployment?

A Delayed Draw Term Loan (DDTL) is a method where a limit is set in advance, and funds are drawn in installments as needed. If the money were taken all at once, the period of paying interest before the equipment is even installed would be unnecessarily long, and the lender’s capital would be tied up for a long time without being used. CoreWeave’s DDTL 5.0 press release explains that the drawdown schedule was aligned with the equipment deployment schedule and the GPU service life.

The flow works like this: First, a contract is signed with a customer, and the SPV takes out a loan to purchase the equipment to fulfill that contract. Once the equipment is installed and operational, the customer pays usage fees, which are used to repay the interest and principal. It is similar to construction loans, where loan disbursements are tied to the progress rate of the building.

For example, assume a 5-year customer contract and a 7-year loan maturity.

After the contract ends, the equipment needs to find a new customer to continue paying interest for those 2 years. Conversely, if the loan maturity falls within the contract period, this risk is significantly reduced.

The first question to determine if the structure is reasonable is this ‘relationship between loan maturity and contract maturity’.

I could not confirm the exact term of each contract from CoreWeave’s public filings. However, I can confirm from press releases that the DDTL 3.0 maturity is August 2030, DDTL 4.0 is March 2032, and DDTL 5.0 is about 5.5 years; comparing these numbers with customer contract terms is the first step in auditing them.

How much does customer credit rating affect secured loan interest rates?

There are cases where the same company with the same loan structure has significantly different conditions.

The effect of customer credit is revealed when comparing two loans CoreWeave closed in 2026.

DDTL 4.0 received a Moody’s A3 rating, with a floating rate of SOFR+2.25% and a fixed rate of approximately 5.9%. DDTL 5.0 was rated Moody’s Ba2 (below investment grade) with an interest rate of SOFR+4.50%, and the press release noted that these funds were used for contracts with two large, non-investment-grade customers. Investment grade refers to a credit rating tier considered to have sufficient capacity to repay principal and interest.

Why does one customer move the interest rate this much? Lenders are essentially asking two questions.

What is the likelihood the customer won’t pay the usage fees, and what can be recovered if they don’t?

General corporate bonds answer the second question with ’the entire company.’ Data center secured loans answer with ’that entity’s equipment and contracts.’

Because the scope of the answer is narrow, the first question—confidence in the customer’s ability to pay—becomes much more important. If there is one creditworthy customer, it effectively becomes a loan using that customer’s credit. If the customer is weak, the quality of the loan drops to that customer’s level, even with collateral. It is more accurate to say that collateral doesn’t replace credit, but rather concentrates credit to make it visible.

The spread difference is 2.25 percentage points. To get a sense of what this difference means, let’s do a simple calculation.

Assuming the full $3.1 billion limit of the 5.0 facility is drawn and multiplying only by the spread difference, it amounts to approximately $69.75 million per year. This excludes base rate fluctuations and fees, so it differs from the actual interest gap. Still, it shows the scale—how a single customer’s credit can shift tens of millions of dollars in annual interest for the same equipment investment.

There are two things to be cautious about here.

First, the customer for 4.0 was only called a ’leading AI company’ in the press release, and their name and credit rating were not disclosed. One analytical article summarized it as a single investment-grade customer, but this was not verified directly from company documents.

Second, it is difficult to explain the interest rate difference solely by customer credit. 4.0 received an A3 rating in a separate credit assessment, while 5.0 was sold via a public syndicate (a method of selling to multiple investors), and the closing dates of the two loans differed by two months. Therefore, while it is possible to interpret customer credit as a major factor in the interest rate, there is no basis to quantify its exact weight.

DDTL 3.0 is also a useful reference. It was a $2.6 billion facility closed in July 2025 with an interest rate of SOFR+4%, used to purchase equipment to fulfill a long-term contract with OpenAI. The press release did not mention OpenAI’s credit rating. The key takeaway from this case is that the spread fluctuated between 2.25% and 4.50% as the customer and timing changed within the same company.*

*

*What are the conditions that make a secured loan structure appear reasonable?

Gathering these facts, the conditions under which this secured loan structure can be more reasonable than general corporate bonds boil down to four factors. If even one is missing, the statement ‘it’s safe because it’s a secured loan’ does not hold. These conditions are my interpretation based on public disclosures, not official criteria from credit rating agencies.

First is customer credit.

The party paying the usage fees must be robust for the repayment source to be robust. It is natural to read that there are such factors behind 4.0 receiving an A3 rating.

Second is the alignment of contract and loan terms.

As seen earlier, if the loan maturity ends within the contract period, the reliance on the used equipment’s value or re-contracting is reduced.

Third is collateral isolation.

The more the loan is issued only to the SPV and is close to non-recourse, the less the parent company’s other circumstances affect the loan’s repayment. This is explicitly stated in the 4.0 press release, while the recourse status for 5.0 could not be confirmed from the press release text.

Fourth is the residual value of the equipment.

When a customer fails to uphold a contract, what remains is the value of the GPU server itself. However, I could not find in public documents how this value was assessed. I also could not verify CoreWeave’s exact public figures for GPU useful life and depreciation periods this time. This item is the final safety valve of reasonableness, and the fact that it is difficult for outsiders to verify is the weak link in this structure.*

*

*What risks remain even in the project finance AI data center structure?

Just as the counter-argument holds weight, one must also look at the grounds for the opposing side to be fair. Even if the conditions for reasonableness are met, risks remain.

First is customer concentration.

According to reports, Microsoft accounted for approximately 67% of CoreWeave’s 2025 revenue. While high credit from one customer is a safety factor, the concentration of revenue in that customer is a weakness in another direction. Even if loans are isolated at the SPV level, the revenue base upon which the company survives is singular.

Second is the case of term mismatch.

If the loan duration is longer than the useful life of the equipment provided, re-contracting is required in the final phase.

Third is the difficulty of valuing collateral.

As noted earlier, the resale value of GPUs is difficult to verify publicly.

When viewed from the perspective of the counter-argument, the more sophisticated the structure, the more variables there are to check.

For corporate bonds, one only needs to see if the company’s profits are declining, but in a secured loan structure, the customers, maturities, and guarantee conditions differ for every entity. Since investors or readers cannot read every entity’s contract, they end up relying on the summaries the company discloses. This is why structural reasonableness and structural transparency are two different things.

Fourth is the fact that secured loans are only a part of the total debt.

In the second quarter of 2026, CoreWeave issued over $10 billion in unsecured bonds and convertible notes in addition to the $3.1 billion secured loan (DDTL 5.0), and the net proceeds from debt issuance were $13.457 billion. [Q2 net interest expense was $640 million](https://investors.coreweave.com/news/news-details/2026/CoreWeave-Reports-Strong-Second-Quarter-2026-Results/), up from $267 million in the same period the previous year.

Unsecured bonds are funds purchased based on company credit, not a contract-isolation structure. According to reports, the market yield on 2032 maturity unsecured bonds was reported at approximately 14.8% (coupon rate 9.625%). This cannot be directly compared to the SOFR spread of secured loans. I could not verify the SOFR level this time, and one is a floating rate while the other is a yield calculated based on market price. However, the direction is clear: market evaluation differs between funds tied to collateral and funds borrowed on company credit even within the same company.

A $104 billion backlog is the biggest number this structure relies on.

However, this backlog is not all held as collateral for individual loans. The structure read from press releases is that individual contracts linked to specific SPVs are collateralized, and I could not find evidence that the entire backlog is collateralized at once. Nvidia’s $6.3 billion obligation to purchase capacity not used by customers is also mentioned as a factor that reduces the risk of demand gaps, but it is not confirmed whether it is included in the collateral structure of each loan.

Are project finance AI data center loans safer than corporate bonds?

Starting with the conclusion: it depends on the conditions.

One cannot generalize that secured loans are structurally safer than corporate bonds; there are simply items to check to see if the structure is designed reasonably. This article summarizes how to read loan structures, not investment advice, and I do not address forecasts for CoreWeave’s stock price or bond prices.

Saying it is different from general corporate debt does not mean it is safer, nor does it mean it is more dangerous. It means the risks are located in different places. In corporate bonds, risk is placed on the company’s overall profitability, while in secured loan structures, it is placed on specific items like a customer’s ability to pay, contract term, and equipment value. The trade-off is that while items are specific and therefore easy to check, there is no other buffer if one item collapses.

If it receives an investment-grade rating like 4.0, has a non-recourse structure, and is backed by customer contracts, a lower interest rate than general unsecured bonds can be reasonable. Loans that rely on two non-investment-grade customers, like 5.0, formed high rates at SOFR+4.50% despite having collateral. This can be read as a signal that the market places greater weight on customer credit than on collateral.

Next time you encounter an article about AI data center loans, check these five things: which entity the loan was issued to, whether the customer’s credit rating is disclosed, whether the loan maturity aligns with the customer’s contract maturity, whether a parent company guarantee is attached, and how equipment value was assessed. Among these, the first three are immediately verifiable via public documents, while the latter two often require reading the loan agreement.