US data centers clash with grid

U.S. data-center growth is colliding with the electric grid and with local politics as AI companies race to add capacity. Industry reports published on May 22 said operators are running into aging transmission systems, long interconnection queues and neighborhood opposition as projects get larger and move closer to population centers. The pressure is showing up both in rural hyperscale campuses that need vast blocks of power and in urban colocation sites built for lower-latency AI inference. The result is a buildout that now depends as much on power procurement, permitting and siting as on servers and chips. ### Why are data centers suddenly a grid story? The U.S. Department of Energy said data centers consumed about 4.4% of total U.S. electricity in 2023 and could reach roughly 6.7% to 12% by 2028, based on a Lawrence Berkeley National Laboratory report released in late 2024. The U.S. Energy Information Administration said on January 13, 2026 that it expected the strongest four-year growth in electricity demand since 2000, driven in part by data centers, and said commercial power sales — a category that includes them — would keep rising in 2026 and 2027. (theregister.com) SingularityHub, citing the International Data Center Authority, reported on May 22 that U.S. data centers account for 29.2 gigawatts of load, about 43% of global data-center electricity use, and that total U.S. consumption has reached roughly 6% of national electricity demand. Mehdi Paryavi, IDCA’s chief executive, said “many very large AI factories” are coming into operation and lifting total consumption. (energy.gov) ### What is breaking first — the hardware plan or the power plan? The Register reported on May 22 that grid operators and developers are “in a bind” because gigawatt-scale AI campuses need firm power without disrupting uptime guarantees, while the U.S. grid is already strained and in need of upgrades. The publication said energy costs have risen in the nation’s largest power market and that utilities and analysts are increasingly pressing data-center developers to secure dedicated generation if they want reliable supply outside grid constraints. (singularityhub.com) NERC escalated those concerns on May 4, when it issued a Level 3 Essential Action alert covering “computational loads” connected to the bulk power system. The reliability group said it had observed large customer-initiated load reductions and oscillations occurring within seconds, leaving little or no time for real-time response, and set an August 3, 2026 deadline for registered entities to respond. NERC has also opened a standards project focused on large electrical loads such as AI compute clusters and data centers. (theregister.com) ### Why are metro data centers back in the conversation? Data Center Knowledge reported on May 22 that AI inference workloads are pulling infrastructure back into metro areas after two years in which the industry favored large rural campuses in Northern Virginia, Texas, Utah and Louisiana. The publication cited a Brooklyn deployment by DataVerge and Mathpix, which said it was installing Nvidia B300 GPU systems in New York to support both training and real-time inference for document-processing workloads. (nerc.com) Brooklyn matters because inference work is more sensitive to latency than centralized model training. That is pushing some operators toward urban colocation even as city projects face tighter land, permitting and utility constraints, according to Data Center Knowledge. ### Where does the backlash show up? (datacenterknowledge.com) SingularityHub reported that resistance tends to emerge when data-center demand moves above about 5% of a country’s electricity supply, citing IDCA analysis. The Register separately described local protests and political opposition as developers search for sites with enough power and water access. Those fights are no longer limited to one region, according to the reports, because AI demand is broadening the map of viable sites while also raising the size of each project. (datacenterknowledge.com) EIA added on March 25 that it had launched pilot surveys in Texas, Washington state and the Northern Virginia-Washington region to measure data-center energy use more precisely. The agency identified 196 companies for those field studies, underscoring how central those markets have become to the next phase of electricity planning. ### What kinds of work become more important when power is the bottleneck? (singularityhub.com) NERC’s May filings and industry coverage point toward a shift in who gets hired around AI infrastructure. The work now includes grid studies, interconnection analysis, power contracting, environmental review, local permitting, site selection and urban colocation strategy — functions that sit between utilities, developers and AI operators rather than inside chip design alone. (eia.gov) That is an inference from the types of constraints named in NERC’s alert, EIA’s demand forecasts and the metro deployment examples now surfacing in New York. The next formal checkpoint is August 3, 2026, when entities covered by NERC’s Level 3 alert must submit responses on how they will address reliability risks tied to large computational loads. EIA’s pilot studies in Texas, Washington and Northern Virginia are also underway, and those results will add detail to how utilities and regulators measure data-center demand. (nerc.com)