US AI Data Centers Face a Three-Way Squeeze on Power, Financing and Construction

An AI data center can have a tenant, a building permit and an equipment order without having a dependable date for its first paying workload. The missing piece may be a utility connection. It may be a transformer. It may be a financing agreement that assumes the other two arrive on time. That is why the American AI infrastructure expansion is better understood as a coordination problem than as a simple race to build more server space.

Power, financing and construction are often discussed as separate constraints. In practice, they reinforce one another. A delayed substation postpones revenue. Postponed revenue increases the money needed to finish the project. A lender facing that uncertainty may demand more equity or stronger customer commitments. A redesign to accommodate higher-density equipment can then change the electrical and cooling requirements, reopening assumptions that appeared settled.

This is not an argument that demand for AI computing will disappear. It is an examination of why demand alone is insufficient to make every proposed facility deliverable. The evidence below comes from government research and regulatory documents. The financing discussion is an analysis of project economics, not a claim that a particular developer is insolvent or that a nationwide credit crisis has begun.

A demand forecast is not a grid connection

The scale is substantial, but the distinction between measured consumption and modeled demand matters. The Department of Energy's December 2024 summary of Lawrence Berkeley National Laboratory research put US data-center electricity consumption at 176 terawatt-hours in 2023, approximately 4.4 percent of national electricity use. Its scenarios projected 325 to 580 terawatt-hours in 2028. Those figures covered data centers broadly, not AI facilities alone 1.

Berkeley Lab's 2025 Update, whose publication page is dated June 2026, extends the modeling to 2030. It gives a reference case of 649 terawatt-hours, or 11.8 percent of US electricity use, with sensitivity scenarios spanning 9.5 to 15.3 percent. These are forecasts built from assumptions about equipment shipments, device power use, utilization and cooling. They are not readings from future electricity meters, and they are not a count of projects certain to open 2.

A national total also conceals the local problem. Electricity available in one region does not make a substation in another region capable of serving an additional campus. A developer needs deliverable power at a particular site, with a defined capacity, reliability standard, price and energization schedule. Buying annual renewable-energy certificates or signing a power-purchase agreement can address procurement and emissions objectives without resolving that physical delivery problem.

The useful question is therefore not simply how much electricity America can generate. It is how much a particular utility can deliver to a particular project, by a particular date, without compromising its obligations to other customers.

Power arrives through several separate decisions

A DOE advisory working group, presenting recommendations in July 2024, described connection requests for hyperscale facilities of 300 to 1,000 megawatts or more, with requested lead times of one to three years, as stretching local grids' ability to supply power at that pace. That is a description of requests and system pressure, not a universal timetable for completed connections 3.

Serving a large new load can require generation, transmission, distribution upgrades and on-site equipment. Each follows a different process. The building can advance while a transmission route remains unresolved. A generation contract can be signed while the substation design is still changing. A utility's preliminary indication that capacity may be available is not necessarily equivalent to a binding service commitment with an enforceable delivery date.

The bottleneck also differs by location. One area may lack generation during peak demand. Another may have sufficient generation but inadequate transmission into the site. Another may need distribution reinforcement or equipment that has not yet been manufactured. Treating all three as a generic power shortage obscures which intervention could actually help.

FERC's September 2024 discussion of interconnection reform illustrates pressure on the supply side: it described more than 2,000 gigawatts of generation resources waiting to connect. That queue concerns generation projects, not data-center load applications. It cannot responsibly be used as a statistic for how long every data center waits. It does, however, show why assuming that new generation can quickly follow new demand is risky 4.

Co-location changes the problem rather than removing it

Locating computing capacity beside a power plant can reduce dependence on some new transmission construction. But proximity does not settle questions about backup supply, network services, reliability obligations or who pays when electricity must come from the wider grid.

In February 2025, FERC directed action on co-location arrangements in PJM, saying the existing tariff appeared insufficiently clear about the rates, terms and conditions governing arrangements involving large loads such as AI data centers. The significance is not that co-location was categorically prohibited. It is that a potentially attractive engineering arrangement still required a workable regulatory and commercial structure 5.

On-site generation presents another set of trade-offs. It may improve control over supply, but the developer inherits fuel arrangements, air permits, maintenance and reliability responsibilities. Batteries can bridge interruptions and shift consumption, but their duration and charging requirements matter. Future nuclear or geothermal capacity may support a long-term strategy without being available for a campus scheduled to open much sooner. The DOE advisory report explicitly warned about uncertainty in the deployment timing and performance of emerging energy technologies 3.

Construction risk is concentrated in systems, not just buildings

The visible structure is only part of a data center. The revenue-producing asset includes electrical distribution, backup systems, cooling, controls, networking and tested interactions between them. A completed shell is not the same as an operational computing facility.

Electrical-equipment supply is a documented constraint. A DOE-supported NREL report published in February 2024 found that utilities were experiencing distribution-transformer lead times of up to two years, around four times pre-2022 levels. It identified workforce and retention problems, component supply issues and material shortages among the causes. These findings concern the wider distribution-transformer market; they should not be presented as a current quotation for every AI campus 6.

Large power transformers introduce another procurement and logistics challenge. GAO's August 2023 assessment recorded utilities' concerns about manufacturing capacity, labor, materials and long lead times. It also described the expense and difficulty of moving these very large assets. Again, that is evidence about grid equipment, not a project-specific construction-cost estimate 7.

The schedule consequence is straightforward. If an essential electrical component arrives late, additional progress elsewhere may not bring the opening date forward. A developer can spend more to accelerate civil work and still wait for the same transformer. Substituting another component can require design changes and testing rather than a simple purchase from a different supplier.

Cooling makes the design harder to freeze

AI clusters complicate construction because the equipment configuration affects the facility around it. Higher rack densities can change the cooling architecture, power distribution and maintenance arrangements. Some installations use direct-to-chip liquid cooling; others combine liquid and air systems. Neither approach removes the need to understand heat rejection, water requirements where applicable, operating conditions and failure behavior.

DOE's July 2024 best-practices guide treats cooling and electrical design as central to data-center efficiency. Berkeley Lab's consumption modeling also incorporates cooling performance. Together, these sources support an important distinction: the chips determine only part of the energy and engineering problem 8.

For a developer, the difficulty is deciding when to stop redesigning. Equipment can improve while the project is still under construction. Adopting the latest configuration may make the facility more competitive, but it can also alter pipework, electrical specifications, controls and commissioning requirements. A nominally better machine can become a worse project decision if accommodating it shifts the opening date far enough.

This is also why commissioning deserves its own milestone. The operator needs evidence that cooling, electrical protection, backup power and controls work together under the intended load and failure conditions. Installing components is not the same as demonstrating that the complete system can support contracted workloads.

Financing depends on which asset is being financed

Calling the entire expansion data-center investment can hide several different exposures. Land and a building shell have one economic life. Power and cooling infrastructure have another. Servers and accelerators have a shorter commercial cycle, and their value depends on performance, utilization and what buyers can obtain elsewhere.

A company financing the building is not necessarily taking the same risk as a lender financing GPUs or an investor funding an operator's cloud business. Some developers have a large, creditworthy tenant under a long contract. Others depend on a smaller group of customers, shorter reservations or future sales of computing time. Those differences matter more to repayment than an impressive campus announcement.

The central financing question is what turns planned capacity into dependable cash flow. A customer reservation may differ from a binding lease. A lease may contain conditions tied to delivery. A contract for electrical capacity may not guarantee that every megawatt is energized on the assumed schedule. Investors need to know which promises remain conditional and who bears the cost when one of those conditions fails.

Delays convert engineering uncertainty into financial exposure

Before opening, a project may already incur land costs, design fees, equipment deposits and financing expenses. If revenue starts later than planned, those obligations do not necessarily move with it. The resulting carrying-cost pressure is an economic mechanism, not a measured national loss estimate.

Consider an illustrative project with a customer commitment that begins only after successful commissioning. If its grid connection slips, it can face a longer period of spending without the expected receipts. If some computing equipment has already been delivered, warranties and the commercial technology cycle may continue moving while the equipment is idle. The precise outcome depends on contracts and accounting, but the timing mismatch is real.

Financing that appears comfortable under one opening date may therefore look different under a delayed one. A credible downside case should test later energization, staged capacity delivery, construction overruns, lower utilization and less favorable refinancing terms. It should also separate risks that the sponsor can control from risks held by utilities, contractors and customers.

This does not mean debt is inherently unsuitable. It means long-lived infrastructure and rapidly evolving computing equipment should not be treated as interchangeable collateral, and that an assumed utilization rate deserves as much scrutiny as the interest rate.

Utilities face a stranded-cost question too

The financial risk does not stop at the campus boundary. Utilities may need to build assets for a customer whose ultimate demand is uncertain. If that customer cancels, scales down or takes longer to ramp up, someone still has to pay for the investment.

The DOE advisory group recommended exploring tariffs that make data centers share or fully pay for required grid upgrades to reduce the impact on other ratepayers. That recommendation identifies the cost-allocation issue; it does not establish a single rule already adopted across the United States 3.

For developers, clearer allocation can make a project more financeable even when it makes the upfront bill larger. Uncertain obligations are difficult to price. For households and smaller businesses, the question is whether they are being asked to underwrite infrastructure built for customers with very different demand and credit characteristics.

The better projects coordinate the clocks

The most useful development discipline is to track several clocks separately: customer commitment, utility delivery, long-lead equipment, construction completion and the commercial life of the computing hardware. A credible schedule explains where those clocks depend on one another and where there is genuine room to recover from a delay.

Phased delivery can help. A first phase with confirmed power and defined customer demand may be easier to finance than a full campus whose later stages depend on uncommitted upgrades. Flexibility can also help where workloads permit it. Some training tasks can be scheduled or moved more readily than latency-sensitive inference services. That flexibility must be demonstrated against customer requirements, not assumed because the workload is labeled AI.

For buyers and investors, useful evidence includes the capacity actually energized, the status of binding utility agreements, equipment already secured, completed commissioning tests and enforceable customer contracts. Planned megawatts, announcement dates and total investment headlines do not answer those questions.

The US AI data-center challenge is ultimately not a lack of enthusiasm. It is the difficulty of turning enthusiasm into infrastructure whose electrical, construction and financial commitments all hold at the same time. The scarce asset is not merely land or capital. It is a project in which the dependencies have been resolved well enough that computing capacity can become reliable revenue.

Limitations

This article is a source-based structural analysis, not an investigation of individual projects or a live census of construction delays. The national electricity scenarios cover all data centers and depend on uncertain hardware and utilization assumptions. Transformer evidence comes from sector-wide reports published in 2023 and 2024, so its historical lead times should not be mistaken for current supplier quotations. Generation interconnection queues are distinct from data-center load queues, and regional tariff conditions differ.

The financing examples are explicitly illustrative. No project failure rate, funding shortfall, lender loss or private contract term has been inferred from the cited documents. Site-level claims require utility records, permits, contracts and project disclosures that are outside this article's evidence base. The sources support the existence of constraints, not a prediction that every US AI data center will encounter all of them.