Hot Stacks is 9fin’s weekly take on data center finance — the deals moving the market, the people behind them, and the risks and opportunities buried in the stack.
The hot stack
Welcome to Hot Stacks — so named because in the data center world, the capital stack is as fast and hot as the server racks themselves. Every Thursday, I’ll be pulling back the curtain on the deals and dealmakers driving the AI hardware buildout.
We’re kicking off our inaugural issue with a reality check that’s shaking up the market — debt in the capital-heavy AI infrastructure race is getting pricier.
Take data center development firm Novva for example. Its founder and chief executive Wes Swenson had expected the $800m of debt he’s planning to raise for a pair of Nevada data center projects in Las Vegas and Reno to price at interest rates not far from investment grade.
“In the 6s to 8s,” Swenson told Hot Stacks. Instead, “we’re seeing rates in the 10s to 11s on high yield. It’s been a sea change in the last six weeks.”
Swenson isn’t alone. Although bankers indicated that they’ve lined up $40bn in high yield data center debt, a sudden surge in pricing is giving borrowers a severe case of sticker shock.
Look no further than the rapid spread widening over recent months — Applied Digital, which is building a data center project in Harwood, North Dakota, raised $2.15bn of senior secured notes back in March for an interest rate of 6.75%. That project, called Polaris Forge 2, encompasses a 200 megawatt data center that was leased by Oracle, one of the Big Five hyperscalers. (Applied Digital’s cost of borrowing will be tested again as it seeks $1.6bn of high yield bonds via Goldman Sachs to fund a data center development in Louisiana, with sources indicating that preliminary price talk suggests a yield in the low to mid-7% range).
Compare that to CleanSpark’s $2.276bn of senior secured notes that priced at 7.875% this month for a Georgia data center that will be occupied by Meta. Despite Georgia being a top-five data center market compared to far-flung Harwood and Meta having superior credit to Oracle, CleanSpark paid over 100bps more for its debt.
Asymmetric risk
High yield investors are realizing that funding the AI rush has come with unequal risks and rewards. As Pimco managing director Lotfi Karoui wrote in a recent report, debt investors have begun to wake up to the possibility they’re bearing “fundamentally asymmetric” downside.
“They finance the AI buildout without directly participating in much of its economic upside,” Karoui wrote. “Returns are largely contractual, driven by coupon, principal, and, at most, some spread compression. On the other hand, the risks span leverage, execution, utilization, technological obsolescence, and refinancing.”
In most corners of the lending market, debt sits comfortably behind a protective layer of equity and the assets that act as collateral have well-established implicit value. Lenders can comfortably price yield. But the data center boom has complicated that calculation.
Builders like CleanSpark and Applied Digital — like many builders in this development boom — have short track records of delivering multi-billion dollar data center projects. If their projects are delayed or if the complex facilities experience operational difficulties, might their tenants walk away?
The chances for such disruptions have only risen as energy, labor, and key equipment such as turbines and transformers have grown scarce. Political and community opposition has surged.
Look at the situation brewing in New Mexico between Oracle and Stack Infrastructure, a data center developer owned by private credit powerhouse Blue Owl. We learned today that Oracle sent Stack a “notice citing force majeure,” according to Bloomberg, in a bid to put off payments for the facility if its delivery is delayed in 2028. $18bn of debt was taken to fund it, meaning a lot is on the line.
Developers say that demand for data center space is so hot, tenants will be in a forgiving mood when inevitable setbacks and complications arise. But if there’s a shift and demand becomes less voracious, could a tenant like Oracle pivot from its requests for rental relief to demands to cancel its lease? If such a scenario were to play out, data centers would become a far more expensive — and economically disastrous — version of the retail apocalypse of the 2010s that left countless shopping malls across the country vacant and crumbling.
In a 2027 market outlook survey recently released by AlixPartners that polled “nearly 500 investors, lenders, builders, operators, and neoclouds,” 64% of respondents “expect distressed situations to increase in next 12–18 months.”
It’s certainly on the mind of some buysiders.
Stefan Balog, a portfolio manager at Impax Asset Management, said that he avoided buying bonds tied to deals that allowed tenants an offramp.
“We try to assess and look at the terms and also the termination rights of the tenant,” Balog said. “Under what circumstances can they actually terminate? We’re looking at all of these things to see how are we shielded or protected from an event where the tenant could walk away.”
Balog attributed some of the recent upswing in high yield rates to deal terms that have shifted to favor borrowers, including lighter amortization and higher leverage levels.
The question of what a vacant data center might be worth in a market that is demonstrably less frothy than today is a focus point for ratings agencies too. Swenson said that during a recent conversation with ratings analysts in August, he was quizzed on what he thought about the residual value of data centers in increasingly unconventional locations (think areas like North Dakota, West Texas, Wyoming, and other obscure data center markets that have suddenly become hotspots for development) when their leases run out.
“The ratings agencies are an indicator of what they get questions on from the bondholders,” Swenson said.
Balog said that yields are becoming more equity-like in that they track the shifting market sentiments around a particular borrower’s progress towards — or retreat from — AI riches and the mega-bucks they plan to spend to get to that pot of gold.
“Bonds are starting to behave closely to the public equity of that company in question,” Balog said.
Under these circumstances, even hyperscalers with fortress balance sheets have inspired concerns given the enormity and duration of their immense AI spending.
“Hyperscalers have seen spreads move wider — that excess supply is being absorbed by wider spreads,” Pimco’s Karoui said. “How much [debt] will come to market in the next 12 to 18 months? When you get to more predictable volume, you’ll see more stabilization in pricing.”
Swenson said that bond investors have also become savvier in their understanding of the AI infrastructure business — and the premiums they can charge amid such a borrowing binge.
There were early concerns, for instance, that data center operators and hyperscalers would need to spend huge sums to frequently upgrade their AI computer chips to remain competitive. Swenson said there’s growing evidence that chips have longer utilization cycles before obsolescence, especially as China has begun to offer cut-rate AI compute capacity.
“The prediction that chips will need to be replaced every three years hasn’t played out and the business will follow traditional 5 to 7 year replacement cycles,” Swenson said. “The bond market started getting more expert on these cycles and it caused the high yield market to go higher. These investors realized that if developers are willing to pay 8%, they’re willing to pay 10%.”
The hype
- Crusoe GPU: Goldman Sachs accelerated and tightened pricing on a DDTL tied to an Icelandic data center
- Digital Drive: Blue Owl-backed Digital Drive priced a $1.12bn senior secured bond via Goldman Sachs to fund the construction of a project in Chesterfield County, Virginia
- Nscale: The company is working with banks including JP Morgan and Goldman Sachsfor debt financing for a data center in West Virginia that it has reportedly agreed to lease to Anthropic
Other news
Google takes the AI data center race to outer space (New York Times)
All in on AI — Why freight borrowers are hitching their wagons to the data center boom (9fin)
The AI buildout is becoming the biggest economic bet in US history (WSJ)
OpenAI’s agents breached Australian government data. Its human response may do more damage (Politico)






