Physical requirements of AI data centers

Behind every AI model training run and cloud query sits a physical support structure: purpose-built facilities that require land, power, cooling and connectivity at industrial scale. “That’s digital infrastructure. Most people don’t realize it, but there’s billions of dollars of physical assets supporting virtually every technology experience that you enjoy today,” said Scott Wilcoxen, head of Global Digital Infrastructure Investment Banking at J.P. Morgan.

Historically, this infrastructure spanned foundational physical networks including cell towers, fiber-optic cables and satellites. As AI adoption and cloud computing grow, data centers are emerging as the most critical facilities to support increased demand.

Not all open land can support industrial-grade utilities and long-term expansion possibilities. Developers must closely evaluate data center build sites based on unique requirements, including:  

How AI infrastructure is financed

AI infrastructure financing requires both scale and bespoke structuring. The capital needs are large, build timelines are long, and the cash flow profile often differs from what traditional investment-grade markets are accustomed to underwriting.

“What’s particularly exciting for us is the level of creativity that is required to finance all of this. Thinking about how to raise this capital in the most efficient way requires a lot of collaboration across teams and markets,” said Servidea.

How debt financing structures AI infrastructure capital

In AI infrastructure, debt most often comes in two core structures: corporate (balance-sheet) debt and project-level debt. Corporate debt allows for speed and flexibility: proceeds can fund capex across data center buildouts, power-related spending, networking and compute without tying the lender to one specific project’s performance.

Project-level debt, by contrast, is underwritten primarily to a defined set of project cash flows, often supported by long-term contracts or otherwise high-visibility revenues. The trade-off is less flexibility and more structural complexity, in exchange for tighter lender protections and debt sizing that is anchored to the asset’s own cash generation.

The June 2026 financing of Hut 8’s Beacon Point data center in Nueces County, Texas illustrates how project-level debt can be structured around durable, contracted cash flows. J.P. Morgan acted as lead left bookrunner on the $4.25 billion senior secured notes offering, which was 100% pre-leased under a 15-year triple-net lease to a high investment-grade tenant. The transaction achieved a 95% loan-to-cost ratio — the highest ever for a high-performance computing data center bond — and priced at T+165, the tightest spread ever recorded for a data center construction bond.

Equity’s role in multi-layered AI infrastructure capital stacks

Equity is the risk-absorbing capital in AI infrastructure, used to fund growth and shoulder uncertainties that debt investors typically discount, such as technology obsolescence, ramp-up timing, utilization volatility and customer concentration. In practice, equity is often used to get projects built and scaled, with the expectation that the business can later add more debt once revenues are contracted and performance is more predictable. As capital needs grow, equity financing structures have become increasingly complex, often drawing on multiple layers of the capital stack simultaneously.

“More often, forming this capital means going to multiple pockets, across multiple regions and, in many instances, looking across multiple layers of the capital stack and incorporating both structured as well as common equity to ultimately fund these investments,” said Wilcoxen.

CoreWeave’s April 2026 capital raise demonstrates how AI-first companies are increasingly drawing on multiple layers of the capital stack simultaneously. J.P. Morgan acted as lead left and active bookrunner on a $5.25 billion combined offering comprising of $1.75 billion in senior unsecured notes and $3.5 billion in convertible senior notes. This deal represents one of the largest combined high-yield and convertible offerings ever executed. It also reflects both the scale of capital required to build and operate AI infrastructure at speed and the sophistication of the financing structures now available across the sector.

Key risks in AI infrastructure financing

Why time to power is the critical constraint for AI data centers

As the demand for data center capacity grows, pressure builds on key components of the supply chain. According to Wilcoxen, the most crucial of these is power supply. Demand for compute is rising faster than many regional grids can support, and the friction shows up in aging transmission infrastructure, limited local capacity and multiyear wait times to connect to power that can push project timelines out materially. As a result, time to power has become a limiting factor for data center and compute deployments. “Accelerating time to power is a top priority for many of our clients, but the grid infrastructure isn’t there,” said Wilcoxen. “There’s a lot of work to do.”

Beyond the grid itself, the supply chain is constrained by natural resources and physical inputs that are harder to scale quickly than capital. Builders must secure available land in the right locations, water access and cooling requirements, and appropriate permitting that determines whether a project can move from plan to energized facility. Clients are responding by securing sites and utility pathways sooner, evaluating alternative cooling approaches where feasible and bringing energy and infrastructure specialists into the diligence process to underwrite power availability and execution risk with the same rigor as the underlying compute demand.

How AI infrastructure developers are managing over-capacity risk

Over-building is a parallel risk: AI capital expenditure has, in many cases, outpaced monetization. “The uncertainty around how, when and in what magnitude end-user demand materializes in a monetizable way does make it difficult to plan out over a multiple-year cycle,” said Wilcoxen.

As a result, long lead times for power and construction are colliding with rapidly evolving technology cycles and shifting customer needs. In response, many operators and sponsors are trying to structure buildouts in phases, pursue more contracted or pre-committed demand where possible, and maintain optionality so capacity can scale with real adoption.

Debt concentration risk as AI financing scales

Debt concentration is an emerging systemic consideration as AI’s share of the overall financing market grows rapidly. “I don’t think we’re at a point yet where concentration limits are a huge concern, but it is part of the dialogue because as a portion of the overall market, AI is growing very quickly,” said Servidea.

Practically, clients are working to keep funding channels resilient by diversifying capital sources, matching debt structures to cash-flow visibility and avoiding overly correlated exposures to a single counterparty, technology node or demand narrative. The throughline is that the next phase of AI infrastructure buildout will be shaped as much by execution discipline as by the availability of capital.

US data center outlook

Investor appetite for AI infrastructure is strong, but it’s not indiscriminate. “Investors are not blindly buying everything. There’s a tremendous amount of capital being deployed, but investors are taking the time to do the credit work, understand the stories and figure out relative value and pricing. To me, that’s a healthy sign for the market,” said Servidea.

Rather than jumping on the bandwagon, investors are positioning themselves to finance a long-term buildout, considering downside protection as much as upside. “The red lights would start flashing if it felt like there was blind euphoria. That’s not what we’re seeing,” Servidea said. 

While AI is a major incremental driver, infrastructure buildout is broader than a single theme. “This entire AI dynamic is additive to an already extraordinarily large investment base,” said Wilcoxen. For instance, public cloud and software-as-a-service (SaaS) models are expected to expand meaningfully in the near future. Compute growth won’t map one-to-one with revenue or usage because efficiencies improve over time, but the direction of travel remains clear: more workloads, more data movement and more always-on digital services will translate into continued expansion of data centers, networks and supporting hardware.

The next phase of AI infrastructure will reward platforms that can secure reliable, scalable power, navigate grid access and interconnection timelines, and structure financing around durable demand signals. If AI meaningfully reshapes products and workflows, the practical implication is sustained investment not just in compute, but in the physical and energy foundations that make that compute usable at scale.

“When you think about all the different things society will be able to do with AI over the next five years, 10 years, 20 years, it’s really hard to comprehend what that’s going to ultimately require,” said Wilcoxen. “We’re building the physical underpinnings of the future technological progress of mankind.”