Why AI's $15 Trillion Buildout Is Rewriting the Rules of Data Center Financing

The unprecedented scale of capital and resources dedicated to the AI buildout is breaking the mold of how data centers get financed and constructed. It’s also raising new questions about how investors assess their risk, as new data and my interviews over the last week showed.

At The Information’s AI Agenda Live last week, I asked two experienced infrastructure developers—Ricard Boada, co-founder and CEO of new computing and funding platform Volta, and Chris Dolan, chief data center officer of developer Crusoe—how approaches are changing to keep up with this velocity of spending and building.

Boada argued capital has to come from new pools of long-term investors that traditionally backed assets like toll roads and telecom towers. Dolan said operators must use more of the computing and power assets they claw together instead of the old habit of overbuilding and underutilizing.

Columbia Business School economist Stijn Van Nieuwerburgh wrote in a paper published by Brookings last week that AI spending will reach 3.63% of U.S. GDP, larger than relative historic bets on canals, railroads, electrification, highways and internet infrastructure. He predicts investment in data center buildings, power, networking, chips and other equipment will total $10.3 trillion between 2025 and 2032.

Boada, whose firm secures access to financing, energy and compute capital for data centers, says he expects the infrastructure buildout to top that and consume $15 trillion over an even shorter span, from 2026 to 2030. That is twice the $7 trillion figure that McKinsey threw out last year. (U.S. GDP is now about $32 trillion.)

The huge sums at stake are forcing markets and builders to adapt, Boada and Dolan said.

Debt investors are already struggling to assess risk from the buildout. A Moody’s report last week detailed an eye-popping $2.8 trillion in off-balance sheet commitments by Amazon, Microsoft, Google, Meta and Oracle, in the form of future leases, purchase agreements and guarantees. That’s an eight-fold jump from levels in 2023, and it still probably undercounts the size of these commitments, since the Moody’s data only goes to June 30. It also doesn’t include any off-balance sheet commitments by Nvidia. Data center project guarantees have only grown since then, especially by Nvidia, as we’ve reported.

“Credit analysis is going to have to get a lot more advanced” to catch up to all the new bespoke financial commitments, said Moody’s analyst David Gonzales.

These less-visible obligations are growing as the same tech companies exhaust their free cash flow and take on hundreds of billions of dollars in on-balance sheet debt. Morgan Stanley estimates $1.4 trillion in AI-related capital expenditures from six companies, Microsoft, Amazon, Google, Meta, Nvidia and SpaceX, for the next 12 months, triple the levels of a year ago.

Investment-grade U.S. debt issuance related to AI now stands at around $500 billion, with $230 billion issued just this year, six times last year’s levels, says HSBC. Analysts warn traditional debt markets’ ability to absorb more borrowing—to, say, fund upcoming leases not yet on the balance sheet—is reaching a saturation point.

These projections explain why financial firms like Blackstone and KKR are assembling multiple vehicles to raise additional hundreds of billions to help cover more of the spend for non-investment grade AI labs and new data center developers.

It’s also why the companies on my panel, Crusoe and Volta, and vehicles like KKR’s Helix Digital Infrastructure, are attempting to bring in-house the procurement of energy, land, chips and capital so that it is easier to orchestrate delivery of AI data centers and avoid penalties for missed milestones.

Helix’s co-founder and CEO Adam Selipsky, who will speak on a panel I’m moderating at the Yotta conference this week, tells me “scale breaks everything” and requires new capital and operating models.

Boada argues that deep-pocketed infrastructure funds, sovereigns and insurers will fund AI data centers and energy sites at a lower cost of capital than what traditional equity and debt investors have been demanding. He argues that AI infrastructure projects will soon approach the characteristics of essential public works like power and fiber lines and bridges, with standardized building formats and long-dated contracts.

“All of these features create… incredibly high levels of bankability,” he asserts.

Some might disagree on that bankability point, since low-cost financing typically demands predictability. AI deployments are so new they’re still unpredictable—just look at margins of error on those spending forecasts! And we are also making revenue assumptions on the current narrow set of AI super users, which makes forecasts especially vulnerable to revision.

Torsten Slok, chief economist at Apollo Global Management, emphasized last week that the top 10% of AI model and inference customers account for 99.5% of the spending. We are still waiting to see what spending patterns we can expect from the rest of the market.

We are also still determining whether communities will allow AI data centers to move forward on a predictable schedule. Oracle’s recent move to declare force majeure on a data center project in New Mexico to give itself the right to postpone payments to lenders because of a permitting delay is a reminder of how each jurisdiction has site-specific political risk.

Still, there are reasons to believe this spending will pay off. Boada has sketched out how these top labs can reap multiples on their investments in cloud spending, chips and data centers.

He figures that the frontier AI labs are generating about $75 million in revenue per megawatt of use a year for inference output compared to an annualized cost to build and provide the infrastructure of $10 million to $15 million per megawatt a year.

Not bad. But first, somebody’s got to cough up the billions to build and power the data center and buy the chips to train the models that enable this revenue stream. That’s mostly coming from public and private stock investors, bond holders and lenders.

With this much money on the line, it then falls on developers like Dolan, who’s overseeing Crusoe’s delivery of Oracle and OpenAI’s Stargate frontier campus in Abilene, Texas, to wring more real-world efficiencies out of labor, energy and chips, which are all rising in cost.

“We’ve all been designing and engineering data centers the same way for the last 25 years,” Dolan said. That meant layering on backup power systems for what’s called “five nines” of uptime, or 99.999% reliability. The scale of this buildout doesn’t tolerate idle capacity, Dolan said. “You have a lot of stranded infrastructure, which means you have stranded capital that you're not leveraging for compute.”

So Crusoe has, for example, rethought how it can rely on additional energy sources at a campus, like batteries, to perform multiple functions —not just act as backup power for the grid but also as a buffer to absorb fluctuating computing workloads. It is also contemplating how it can bring in modular data center units as swing computing capacity to utilize extra power on hand, or add smaller computing units in locations with tight power availability.

Engineers are pretty black and white, but this moment is prompting them and his customers “to think differently, ” Dolan said.

Ann Davis Vaughan is the author of the AI Infrastructure newsletter for The Information. She is a former senior Wall Street Journal investigative reporter turned investment strategist who has tracked the energy, industrial and financial sectors for three decades.