Data center pushback threatens AI chip demand cycle as utilities fall 100 GW short
Grid constraints and community opposition could break the AI power demand boom that chipmakers and gas suppliers are banking on.
Bank of America analysts now project the United States will need more than 230 GW of new generating capacity over the next five years, yet regulated utilities are expected to add only about 93 GW of accredited supply. That leaves a gap of more than 100 GW between what AI data centers are demanding and what the grid can plausibly deliver.3
Power availability, equipment lead times and interconnection delays are now the binding constraints on AI infrastructure deployment, not GPU supply or capital. The entire AI chip demand cycle rests on an assumption that power will be available when and where hyperscalers need it.2
Data centers alone could add roughly 125 GW of US electric load over the period, BofA said. Industry estimates put global data center electricity consumption at approximately 565 TWh in 2026, up from 447 TWh in 2025, with longer-term projections moving beyond 1,000 TWh by 2030 and approaching 1,300 TWh by 2035.2,3
The scale is visible in Virginia, where commercial electricity sales increased by nearly 30 million megawatt-hours between 2019 and 2025. The EIA attributes much of that increase to the state's concentration of data centers, and the US now consumes nearly 40% of the world's data center electricity.5,6
But the pushback is building. Higher electricity prices are turning AI into a problem for communities hosting the facilities, and the technology's returns have not lived up to the hype that justified the buildout. What began as a promising growth story has become a source of cost pressure for ratepayers and political friction for developers.1
The uncertainty shows up plainly in forecast spreads. BloombergNEF's two scenarios for US data center electricity demand by 2030 differ by 42 GW, highlighting how little clarity exists on the buildout's ultimate scale even as construction accelerates. BNEF tracks roughly 100 GW of project capacity added across the US in the last year alone.4
Almost all US regions ended 2025 with more data center capacity than BNEF had anticipated, with Texas accounting for the largest difference between forecast and actual buildout. The base case, modeled on development BNEF considers likely, points to 207 GW of US data center demand by 2033.4
The mismatch between physical constraints and financial forecasts is sharp. DataM Intelligence projects the global AI data center market will grow from US$120.74 billion in 2025 to US$1,020.83 billion by 2035, a compound annual growth rate of 22.8%. That trajectory assumes power materializes on schedule.2
The IEA reported that electricity demand from data centers jumped 17% last year, with AI-hosting facilities rising even faster. Academic analysis of 403 US hyperscale facilities operating between May 2024 and April 2025 estimated electricity use of roughly 68 TWh to 99 TWh, about 1.8% of total US consumption under the central scenario.2,1
The IEA's April report underscored how quickly the load base is compounding. Yet regulated utilities move slowly, and the accreditation process for new supply is not keeping pace with the data center queue.1
As utilities struggle to bring capacity online fast enough, BofA expects more data center developers to turn to behind-the-meter generation. That shift would bypass the grid bottleneck but carries its own problems: gas pipeline connections, permitting and financing for on-site plants all take time.3
Natural gas suppliers and pipeline companies are betting on data center demand as a new source of fuel consumption, and turbine and transformer manufacturers see a fresh order book. The bet works only if the buildout survives local opposition and utility rate cases.7,6
The 42 GW spread between BNEF scenarios is widening as communities push back, and a hyperscaler publicly walking back capacity commitments on power grounds — or a state utility commission rejecting a data center tariff — would be the clearest signal yet that the gap between forecast and deliverable supply is becoming a hard ceiling rather than a planning assumption.4