Capital and resources pouring into AI construction are unprecedented, breaking the traditional financing and building models for data centers.

Over the past week, combining the latest data with my own interviews, this wave has also created entirely new challenges for how investors assess risk.

At a forum, I interviewed two veteran infrastructure developers: Volta co-founder and CEO Ricard Boada, and Crusoe data center director Chris Doran, to discuss how the industry needs to adjust its thinking to keep up with the current frenzied pace of spending and construction.

Boada proposed that capital needs to come from entirely new pools of long-term investors, capital that historically focused on assets like toll roads and communication towers. Doran said operators must fully activate the computing and power assets they have consolidated, abandoning the old approach of overbuilding and underutilizing.

Columbia Business School economist Stijn Van Nieuwerburgh published research last week stating that AI-related spending will account for 3.63% of U.S. GDP, a relative investment scale exceeding major infrastructure projects in history such as canals, railways, electrification, highways, and internet infrastructure. He predicted that between 2025 and 2032, total investment in data center buildings, power, networking, chips, and other equipment will reach $10.3 trillion.

Boada's institution specializes in connecting data centers with financing, energy, and computing capital. He expects infrastructure investment to exceed that forecast, reaching $15 trillion in total over just a few years from 2026 to 2030, double McKinsey's $7 trillion forecast last year.

The massive capital stakes are forcing markets and builders to adapt, Boada and Doran said. Debt investors are already struggling to assess the various risks in this infrastructure boom.

A Moody's report last week disclosed that Amazon, Microsoft, Google, Meta, and Oracle have a staggering $2.8 trillion in off-balance-sheet commitments, in the form of future lease agreements, procurement contracts, and guarantees. That figure is 8 times higher than in 2023, and it likely still understates the actual scale — Moody's statistics only go through June 30 and do not include any of Nvidia's off-balance-sheet commitments. Since the cutoff date, data center project guarantees have continued to expand, especially for Nvidia.

Moody's analyst David Gonzalez said: "Credit analysis frameworks must be significantly upgraded to keep pace with these entirely new customized financial commitments."

These hidden liabilities continue to climb, while major tech companies exhaust their free cash flow and add hundreds of billions of dollars in new debt to their balance sheets. Morgan Stanley estimates that Microsoft, Amazon, Google, Meta, Nvidia, and SpaceX will spend $1.4 trillion on AI-related capital expenditures over the next 12 months, triple the level of last year. HSBC data shows that U.S. investment-grade bond issuance related to AI currently stands at about $500 billion, with $230 billion issued this year alone, six times last year's pace.

Analysts warn that the traditional debt market's capacity to absorb new borrowing — for example, to finance lease projects not yet on the books — is approaching saturation.

The above forecasts explain why financial institutions such as Blackstone and KKR are setting up multiple funds to raise hundreds of billions more to provide capital for non-investment-grade AI labs and emerging data center developers. This is also why companies like Crusoe and Volta, as well as KKR's Helix digital infrastructure platform, have chosen to internalize energy, land, chip procurement, and financing altogether, making it easier to advance AI data center delivery in a unified way and avoid penalties from project milestone delays. Helix co-founder and CEO Adam Selipsky will speak at a forum panel I am moderating this week. He told me: "Scale changes everything, and we need an entirely new capital model and operating model."

Boada believes that well-capitalized infrastructure funds, sovereign funds, and insurance capital can invest in AI data centers and energy sites at a lower cost of capital than traditional equity and debt investors. He proposed that AI infrastructure projects will soon take on the characteristics of basic public works like power, fiber optics, and bridges: standardized construction plans plus long-term contracts. He asserted: "All of these characteristics together will greatly enhance project financing feasibility."

But not everyone agrees that projects have this kind of financing certainty. Low-cost financing typically requires predictable returns, while AI deployment is still an emerging matter with enormous uncertainty — just look at the wide margin of error in these spending forecasts themselves. Moreover, revenue forecasts are based on a small number of heavy AI customers today, and once demand shifts, the forecasts can easily be revised sharply. Apollo Global Management chief economist Torsten Slok pointed out last week that the top 10% of AI model and inference customers account for 99.5% of related spending. The spending patterns of the rest of the market remain to be seen.

In addition, the pace of local community approvals will also affect whether AI data centers can be delivered on schedule. Oracle recently triggered force majeure at a data center project in New Mexico, obtaining the right to delay payments to lenders due to approval permit delays. This incident reminds people that localized policy risks exist in every region.

Nevertheless, there are still reasons to believe this massive investment can pay off. Boada outlined how leading labs can earn high returns from investments in cloud resources, chips, and data centers. He estimates that frontier AI labs generate about $75 million in annual inference revenue per megawatt of computing power, while the annualized construction and operating cost of supporting infrastructure is only $10 million to $15 million per megawatt. The returns look quite good. But the premise is that someone must put up billions of dollars to build powered data centers, procure chips, and train models that can generate this revenue.

The funding mainly comes from public market and private equity investors, bondholders, and lenders. This massive bet ultimately falls on the shoulders of infrastructure developers like Doran. He is responsible for delivering Crusoe's Stargate frontier computing campus in Abilene, Texas, for Oracle and OpenAI, and must continuously extract real operational efficiency in an environment of rising labor, energy, and chip costs.

Doran said: "For the past 25 years, our approach to designing and building data centers has not changed." To achieve so-called "five nines" (99.999%) system availability, layers of backup power are built in. But this round of hyperscale construction cannot afford idle capacity. "A large amount of infrastructure sitting idle means capital is trapped and cannot be put to use for computing."

For example, Crusoe redesigned the campus energy utilization plan: batteries are no longer merely backup power for the grid, but can also serve as a buffer to absorb fluctuating computing loads. The company is also considering introducing modular data center units as flexible computing capacity to absorb surplus power, or deploying small computing units in power-constrained regions. Engineers usually think in black and white, but the current situation is pushing engineers and their customers to "think differently," Doran said.