AI build‑out stalled by power, water, permits

# AI build-out stalled by power, water, permits The artificial intelligence boom was supposed to be a race for chips. Instead, one of the biggest slowdowns is now happening outside the server room: at substations, in water systems, and in permitting offices. Across the United States, developers still want to build large data centers for artificial intelligence training and cloud services. But getting enough electricity to the site, enough cooling capacity to keep machines from overheating, and enough permits to start construction is taking longer than many companies expected. New reporting and industry research show that power availability has become the main constraint on growth, even while spending on data center construction remains enormous. That shift is visible in the numbers. Construction Dive, citing research from CBRE and Wood Mackenzie, reported on March 26, 2026 that North American data center construction slowed in the second half of 2025 for the first time in six years. Capacity under construction fell nearly 6 percent year over year, even though demand for artificial intelligence and cloud computing stayed strong and vacancy rates hit a record low 1.4 percent. The same report said United States data center construction added 25 gigawatts in the fourth quarter of 2025, down 50 percent from the third quarter. The reason is simple: a data center is only useful if it can turn on. A modern hyperscale campus can need as much electricity as a small city, and that electricity has to arrive reliably every hour of the day. If a project requires a new transmission line, a substation upgrade, or added generation capacity, the construction schedule can slip by months or years before the first server is installed. Construction Dive quoted CBRE data center research director Gordon Dolven saying that “power and electrical equipment is still the main driver of construction delays.” Federal analysis points in the same direction. A Congressional Research Service report published on December 12, 2025 said a 2024 Department of Energy study projected United States data center electricity consumption would rise from 176 terawatt-hours in 2023 to between 325 and 580 terawatt-hours by 2028. That is an extraordinary increase in a short period, and it means the bottleneck is no longer just whether companies can buy enough graphics processing units, but whether the electric system can absorb the new load. Once that demand hits the grid, the next choke point is interconnection. Interconnection is the process that decides how a new large customer connects to the power system and what upgrades are needed to do it safely. GridLab’s 2025 report on large-load interconnections says data center developers are aggressively seeking transmission access because speed to power matters more than almost anything else. Utilities then have to study the request, estimate the impact on reliability, and determine who pays for upgrades. That process can be long, technical, and highly uncertain. The Idaho National Laboratory described the same problem in a January 2026 white paper for the Department of Energy Grid Deployment Office. It said the rapid spread of large loads, especially hyperscale data centers, is straining existing grid infrastructure and complicating planning, interconnection, and deployment. The paper lists interconnection delays, siting constraints, reliability concerns, and cost allocation as core obstacles. In other words, even when money is available and local officials want the jobs, the physical network may still not be ready. Water is becoming a second strategic constraint. Many large data centers use water-based cooling systems to remove heat from densely packed servers. In places already worried about drought, aquifer stress, or municipal water limits, that can turn a straightforward industrial project into a politically sensitive one. A February 18, 2026 alert from Nixon Peabody said developers are facing growing permitting risk around water conservation, backup generator emissions, and environmental review of cooling systems, chemicals, and spill controls. It also warned that errors in those areas can trigger redesigns, offsets, penalties, or prolonged appeals. Permitting itself has become part of infrastructure strategy. The Congressional Research Service notes that data centers need permits not only for the buildings but also for associated energy infrastructure such as grid connections, on-site generation, and backup power systems. That means a project may depend on several overlapping approval paths at the federal, state, and local level. If a site needs diesel or gas backup generation in an area with air-quality restrictions, or if cooling plans raise water concerns, the timeline can stretch well beyond the original construction plan. This is already changing where companies build. Bloom Energy’s 2026 Data Center Power Report says capital is concentrating in “power-advantaged” regions and that power availability is driving a geographic reallocation of growth. It says Texas is poised to become the nation’s leading data center market within three years, a sign that developers are increasingly following available electricity rather than traditional data center geography alone. The same report says on-site generation is moving from a temporary bridge solution into long-term planning. That is where behind-the-meter systems come in. “Behind the meter” means power resources located on the customer’s side of the utility connection, such as batteries or on-site generation that serve the data center directly. Developers are considering these systems because waiting for a full grid upgrade can take too long. The Bloom report says interest in alternative power strategies has risen as interconnection timelines widen, and industry coverage from Utility Dive notes a growing willingness to deploy dedicated on-site power when the queue or traditional procurement process cannot meet data center deadlines. The pressure is also changing how operators think about computing itself. Carnegie Mellon University said on April 7, 2026 that its researchers are working with Bosch Research on ways to coordinate artificial intelligence workloads with energy availability. Instead of running every job the moment servers are free, the idea is to schedule some workloads around periods when renewable electricity is abundant or when grid stress is lower. The researchers say large training runs can cause sharp swings in demand that affect grid stability and, in some cases, local electricity costs. That points to the next phase of the artificial intelligence infrastructure race. For years, the dominant question was who could buy the most computing hardware. Now the harder question is who can secure power, cooling, permits, and grid access fast enough to use that hardware at scale. The result is a quieter but more consequential competition. Companies are no longer just shopping for land and chips; they are shopping for substations, water rights, air permits, transmission capacity, and communities willing to host the facilities. Unless utilities, regulators, and developers find faster ways to expand infrastructure and process approvals, the build-out of artificial intelligence will be limited not by software ambition but by the physical systems underneath it.