U.S. data centers could consume about 20% of the country’s electricity by 2035, according to a new BloombergNEF forecast reported by Bloomberg. That is up from 5.9% today, and it gives utilities, regulators, and ratepayers a fairly blunt problem: the AI buildout is asking the grid for power faster than the grid is used to delivering it.
BloombergNEF raised its 2035 projection for U.S. data center demand to 194 GW. The research firm’s previous forecast, issued in December, was 106 GW. That December estimate was already 36% above the figure BloombergNEF published in April 2025, which was about 78 GW.
In other words, BloombergNEF’s 2035 number has risen by roughly two and a half times in 15 months. The latest increase, according to the reporting, is tied to a swelling pipeline of proposed AI data centers rather than sites already being built. That distinction matters. A project in a utility queue is not the same thing as a powered building full of GPUs, no matter how aggressively the slide deck implies otherwise.
BloombergNEF estimates data centers will represent around 12% of U.S. electricity use in 2030, then reach about 20% by 2035. Virginia and Texas are expected to sit above the national average, reflecting where large data center clusters and AI-related development are concentrated.
Nearly half of the projected capacity is expected to support AI training and inference, according to BloombergNEF. The firm also estimates that, measured by power demand, the U.S. will host 64% of the world’s AI chips in 2033.
The grid math is ugly
The largest amount of data center capacity ever connected to the U.S. grid in one year is 7.1 GW, according to BloombergNEF. Even if the country repeated that record every year through 2035, the firm’s base case still leaves a 19 GW gap after including on-site gas generation.
Lloyd Arnold, a BloombergNEF analyst and one of the report’s authors, told Bloomberg that one-fifth of the power produced by U.S. coal plants, gas plants, and solar farms would go to data centers under the forecast. That is not a marginal load. It is an industrial-scale reallocation of electricity toward server farms.
Other forecasters are also revising upward. The electrical industry nonprofit EPRI has more than doubled its 2024 estimate, while S&P raised its projection by more than a third between October and April, according to the reporting. Gartner separately forecast that global data center electricity use will rise 26% this year and reach 565 TWh in 2026.
Regional grids face different pressure
TechCrunch reported that PJM Interconnection, the 13-state grid operator that includes Northern Virginia’s data center corridor, could send 34% of its electricity to data centers by 2035. ERCOT, the Texas grid operator, is projected at 22%.
PJM’s independent market monitor has already tied a 75.5% increase in regional power costs directly to data center demand. That is where the abstract AI infrastructure boom becomes a bill. Data center operators may sign the power agreements, but grid upgrades and scarcity can show up across the market.
The pipeline also has a reliability problem of its own. Close to half of planned U.S. data center builds this year are projected to be delayed or canceled, according to the reporting. High-power transformer lead times have stretched as long as five years.
BloombergNEF’s December revision was driven largely by early-stage projects entering utility queues to reserve power before construction started. The same caveat hangs over the new 194 GW forecast: some of the demand is a serious build plan, and some of it is developers trying to claim a place in line before the substation fills up.
This story draws on original reporting from Tom's Hardware.
