Are GPU Loans Safe From Rising Rates?

The Federal Reserve’s recent rate hike, and rising expectations of another soon, put a fresh spotlight on financing arrangements for the wide array of companies borrowing money to pay for AI data centers and chips. 

Rates are rising just as investors have been growing more skeptical about AI-related debt in general, a potentially potent mix for higher borrowing costs for new deals across the board. But in one corner of the AI financing market that might sound the worst off—floating-rate loans used by major GPU buyers—some of the biggest loans have already been engineered to blunt the impact of higher benchmark rates, for now at least.

CoreWeave has borrowed billions in recent years to pay for Nvidia’s graphics processing units via delayed draw term loans, a structure others including Iren and Nscale have begun to embrace too. Many of these deals have hedges built in that are meant to effectively turn much of their floating-rate exposure into fixed rates. 

These DDTLs, which are generally drawn down as the GPUs are purchased and deployed, are often initially structured with floating interest rates for at least some of the borrowing capacity. For instance CoreWeave’s largest DDTL, an $8.5 billion facility, is roughly split between floating rate and fixed-rate portions. That means that the interest rate on some of the financing can change frequently, unlike that on a fixed-coupon bond, commonly used in AI data center financing. 

The interest rate on the floating portion of many DDTLs is based on short-term gauges like the secured overnight financing rate, which typically moves with changes in the Fed’s benchmark rate, plus a fixed spread on top. SOFR was 3.87% as of Sept. 23, up from a recent low of around 3.5% in May, though it remains below its year-ago level. Some GPU financings use a version called Term SOFR, a forward-looking rate that’s around 4% right now on a three-month basis. 

Meanwhile, recent big GPU loans have carried spreads ranging from roughly 2.25 percentage points to 5.5 percentage points added to SOFR-based benchmarks. Spreads can vary widely based on factors like the creditworthiness of the customer ultimately using the GPUs. 

But for the floating-rate component, some of the biggest GPU borrowings also have significant hedges against rising rates. Those are generally derivatives such as interest rate swaps, where a borrower pays a counterparty a fixed rate while receiving a floating rate.

For example, CoreWeave was hedging at least 75% of anticipated floating-rate borrowing under a $2.6 billion GPU facility from last year, filings show. The floating portion of the company’s newer $8.5 billion facility, meanwhile, will be subject to a hedging requirement of at least 95% of anticipated borrowing.  

And Nscale’s $1.85 billion GPU facility for a data center in Ward County, Texas, requires interest-rate hedges covering between 90% and 110% of its projected floating-rate borrowing, according to credit agreements the company disclosed as part of its recent initial public offering filing. Nscale’s other large U.S. GPU facility, a $1.2 billion loan for a North Carolina project, has a similar hedge requirement.

There are arguably some benefits of the DDTL structure itself. Interest on DDTLs generally starts accruing only once the money is drawn, though borrowers may pay a smaller, separate fee on the undrawn commitment. So a buyer of GPUs can secure a big pool of money but only pay interest as it draws the funds to buy and deploy the hardware. 

But all that doesn’t mean borrowers can ignore rate risk entirely, especially when it comes to contemplating new borrowing. 

Hedge requirements have already been written into some loan agreements, which limits how much exposure companies can leave floating. But for new facilities, companies or their lenders may still have to decide how much exposure to leave unhedged.

Of course, AI infrastructure providers can look to tap a wide range of financings beyond DDTLs, and in general companies shift between sources of financing based on investor appetite and market pricing. There’s a buffet of options for finding cash that includes vendor financing and private credit as well as bonds and convertible debt, plus issuing equity. 

But those aren’t all necessarily well suited to financing GPUs specifically, since borrowers spend money on the GPUs over time. And CoreWeave, for its part, has told investors that its DDTLs are structured so debt payments line up with contractual payments from its own customers. 

Spreads have widened on investment-grade AI credit lately, and new issuance has been coming in pricier than the broader market, while some lower-rated AI data center project debt has seen investors demand especially high yields. Convertible debt can keep interest costs down but come with potential dilution for shareholders. 

In other words, finding attractive financing may be getting tougher regardless. And part of why AI borrowing is getting more expensive isn’t related to rates—at least not directly. Investors have been swamped with new supply while questions about the ultimate payoff of AI investments remain largely unanswered.

Overheard: Who Buys All the AI Debt? 

Jas Khaira, head of Blackstone’s AI-focused investment team, said on Wednesday that there are still some big questions about what kind of investors will step up to continue financing the AI industry’s infrastructure buildout. 

“We have not mapped out who’s going to buy all the debt,” he said at The Information’s AI Agenda event in San Francisco, responding to an audience question about whether financing costs will have to rise in order to fund the buildout. “We have some hypotheses, but the honest answer is we don’t know where all of it will come from.”

He noted debt markets for investment-grade companies and the private credit market as potential sources of capital. He said that while public markets are large, debt will “probably clear at a higher price because of the quantity getting pushed.” 

And how exactly big companies like OpenAI and Anthropic will finance their compute needs remains uncertain. “A lot of that is going to depend on their ability to go public,” he said, “and their ability to be investment grade over time.”

New From Our Reporters