BREAKINGVIEWS-Shaky data centre tenants could choke off AI boom

By Yawen Chen

LONDON, Dec 10 (Reuters Breakingviews) - The global boom in artificial intelligence has unleashed a giant wave of data centre construction. Developers are committing vast sums of capital to build facilities capable of powering large language models. For now, banks and other lenders are willing to finance the expansion. However, their confidence rests on a fragile assumption: that the companies renting the facilities remain creditworthy.

The scale is enormous. Morgan Stanley analysts reckon global data centre capacity will need to grow six-fold by 2035 to meet the demands of cloud computing and AI. This translates into an estimated $3 trillion investment in infrastructure between 2025 and 2028. The United States alone will need around 82 gigawatts of additional electricity-generating capacity through 2030 to power the facilities, Goldman Sachs estimates. That's roughly equivalent to all the offshore wind power capacity installed globally.

Much of this historic splurge is being funded by cloud computing giants like Microsoft MSFT.O, Amazon AMZN.O and Google, which benefit from vast operating cash flows and strong credit ratings. However, new entrants to cloud computing could account for about 20% of the global rental market for AI-grade graphics processing units (GPUs) by 2030, Goldman analysts calculate . The strains on those upstarts' balance sheets are already beginning to show.

New suppliers like CoreWeave CRWV.O and Nebius NBIS.O, often referred to as neo-clouds, are the middlemen of the AI gold rush. They lease data centres, fill them with GPU-heavy clusters, and then rent out the processing capacity to customers ranging from large technology groups to venture-backed AI startups. The neo-clouds' rise has helped to quickly scale up the supply of computing. But it has also concentrated credit risk around the developers that build the data centres. These companies tend to borrow heavily to finance construction, now backed by long-dated leases from neo-cloud tenants.

The financing mechanics leave little margin for error. Developers pay for the construction costs with a mix of common equity and substantial amounts of debt from banks and private-credit lenders. Infrastructure investors often provide preferred equity or mezzanine capital above common equity, earning fixed returns. The residual risk sits with the developer and, ultimately, its creditors.

The neo-clouds that rent the data centres reduce their risk by signing take-or-pay contracts with their biggest customers, ensuring they receive payment whether or not the capacity is fully used. Nevertheless, there's a timing mismatch. While client contracts typically run for four to five years, the neo-cloud provider will usually lease a data centre for 15 years or longer, leaving it on the hook if customers flee.

This model is very different from the projects backed directly by cloud computing giants, also known as hyperscalers. Facilities leased to Microsoft or Amazon benefit from those companies' robust balance sheets and diversified revenue streams, allowing them to access cheaper financing, although a recent rush of borrowing by peers from Google owner Alphabet GOOGL.O to Meta Platforms META.O is testing bondholders' appetite. When the tenant is a less financially robust intermediary, however, developers have to pay more for debt.

These borrowing costs are the biggest variable when determining the likely return on investment for a new AI-focused data centre. To see how it works, consider Applied Digital’s APLD.O partnership with CoreWeave, a fast-growing specialist cloud provider that rents Nvidia GPUs to customers including hyperscalers and AI start-ups. The $40 billion New Jersey-based company has a junk credit rating. Texas-based Applied Digital is constructing purpose-built facilities for the company under long-dated leases. One flagship site under construction in North Dakota will have a power capacity of 100 megawatts and cost an estimated $1.2 billion.

Including other facilities, CoreWeave has committed to rent 400 megawatts of Applied Digital capacity for 15 years, paying roughly $135 per kilowatt per month, Breakingviews estimates. Assume the same payment for the next 100-megawatt build-out and it will earn revenue of about $160 million a year once operational. After deducting operating costs equivalent to roughly 12% of revenue, net operating income will be about $140 million a year.

Assume Applied Digital borrows $900 million of the construction cost, in line with industry norms, at a 7.5% interest rate as lenders typically demand a spread of 4.5 percentage points over SOFR, the secured overnight interest rate, for tenants like CoreWeave compared to just 2 to 2.5 percentage points for the hyperscalers. Also assume revenue rises by 2% a year while costs consume the same proportion of income.

The other important factor is the value of the facility at the end of CoreWeave's lease. This is calculated using a capitalisation rate which is standard in real estate financing. At a 6% cap rate used for prime locations, the data centre would be worth $3.1 billion and Applied Digital would earn a respectable 16% internal rate of return on its investment, according to Breakingviews calculations.

However, even small shifts in these assumptions make a big difference. Assume the interest rate rises to 10%, and apply a 10% cap rate typical for subprime locations, and the IRR drops to 12%. That's below the 12.75% that Macquarie Asset Management is charging for preferred equity on the Applied Digital-CoreWeave project.

Credit markets are already signaling that tenants like CoreWeave have become riskier. Five-year credit-default-swap spreads referencing the company's debt, which traded in the 250 to 300 basis-point range earlier this year, spiked to roughly 720 basis points in November after CoreWeave modestly trimmed its revenue guidance for 2025. That cut was due to capacity delays at a third-party data centre provider rather than weaker customer demand. Even so, credit investors focus on near-term cash flow and fixed data centre lease obligations, meaning delays to revenue quickly translate into higher perceived risk.

That raises the prospect of a negative feedback loop. Slower delivery of new facilities can weaken server farm tenants’ near-term cash flows, which in turn scares lenders and pushes up financing costs for developers, making future projects harder to justify.

Similar distress signals are visible elsewhere. CDS spreads referencing debt of Oracle ORCL.N, which has also piled into data centres, have widened despite the company’s investment-grade credit rating. The concern is not just Oracle’s balance sheet, but execution risk tied to the large volumes of data centre capacity it is building for OpenAI, and whether the owner of ChatGPT will be able to pay for it.

For now, there's no sign of a letup in the data centre construction bonanza. Banks like JPMorgan, Deutsche Bank, MUFG, and Citigroup, and private credit providers such as Blackstone, KKR, Apollo Global Management and Ares continue to pile into new projects. Meanwhile, developers expect to be able to refinance their projects at higher valuations and lower interest rates when the facilities are complete.

Even so, the biggest risk to the data centre boom is not waning demand for AI. It is that lenders, faced with stressed tenants and rising signs of credit strain, decide to charge more, which in turn would choke off equity investment for new projects. The physical underpinnings of the AI bonanza are looking fragile.

Follow Yawen Chen on Bluesky and LinkedIn.