AI Hits Power Limits
- Rapid AI expansion is colliding with limited local electricity supplies and slow interconnection approvals, creating buildout bottlenecks. (cnn.com) - Interconnection queues now stretch for years, pushing data-centre operators toward distributed generation and real-estate strategies to secure power. (datacenterknowledge.com) - Industry voices say utility coordination on transmission, cooling and permitting is becoming a gating factor for AI deployment. (powermag.com)
AI companies can buy chips and land faster than they can get electricity, and that is starting to slow where new data centers can open. (cnn.com) A data center is a warehouse of servers, and AI servers draw far more power than older cloud machines because they pack in dense clusters of graphics processors and need heavier cooling. Operators told Data Center World this week that utility service, not building design, is now the first question in site selection. (techrepublic.com) Interconnection is the approval process that lets a new power plant or large customer plug into the grid, and those lines now move slowly enough to shape AI construction schedules. Data Center Knowledge reported on April 21 that waits can stretch for years, pushing developers toward on-site generation and real-estate deals built around secured power. (datacenterknowledge.com) The queue behind those delays is enormous. Lawrence Berkeley National Laboratory said nearly 1,570 gigawatts of generation and 1,030 gigawatts of storage were actively seeking interconnection in the United States at the end of 2023, or about 2,600 gigawatts combined. (lbl.gov) The power strain is growing as AI usage rises. The International Energy Agency said on April 10 that global electricity demand from data centers is set to more than double to about 945 terawatt-hours by 2030, and demand from AI-optimized facilities is projected to more than quadruple. (iea.org) That pressure is showing up first in the Mid-Atlantic, where the nation’s biggest grid operator is rewriting rules around large loads. PJM said on March 2 that it filed proposals at the Federal Energy Regulatory Commission for expedited interconnection and other steps to address a supply-demand imbalance “largely driven by data center growth.” (pjm.com) By April 10, PJM had gone further and proposed a one-time backstop procurement for 14.9 gigawatts of new resources to serve data centers and other large loads. Utility Dive reported the process would begin with bilateral contracting between suppliers and large customers from September through March. (utilitydive.com) Operators are also redesigning the buildings around the power problem. TechRepublic reported from Data Center World that AI racks are moving toward 100-kilowatt and even megawatt-scale densities, which forces changes in power delivery, cooling equipment, and site engineering. (techrepublic.com) POWER Magazine said today that utilities and data-center developers now have to coordinate earlier on transmission, substations, cooling, and permits because those pieces can delay projects as much as the servers themselves. The bottleneck has shifted from finding demand for AI to finding enough electricity, in the right place, on the right timeline. (powermag.com)
