Data Centers May Use 20% of US Power by 2035

What's happening

A wave of new reporting is converging on a single, uncomfortable theme: the American electric grid is being pushed harder, faster, and further than it was built to handle, and artificial intelligence is a major reason why. Projections now suggest data centers could account for roughly 20% of total U.S. electricity consumption by 2035, with states like Virginia and Texas — already home to dense clusters of server farms — expected to see the effects show up directly on residential power bills 1. That long-range forecast is landing at the same time grid operators across the country are reporting record-breaking demand in real time, turning an abstract future concern into a present-day operational headache.

In Texas, the grid operator ERCOT confirmed an unofficial new record for peak power demand in a single day, surpassing the previous official mark of 85,508 megawatts set in August 2023 2. A few states over, PJM Interconnection — the grid operator serving a huge swath of the mid-Atlantic and Midwest, including data-center-heavy Virginia — reported preliminary numbers showing an all-time peak load of 168,158 megawatts on July 2, eclipsing its prior record of 165,563 megawatts 4. Meanwhile, in Colorado, utility strain has taken a more localized form: Colorado Springs Utilities twice asked customers to voluntarily cut back on electricity use as high demand pressured the local grid 35.

Layered on top of these demand spikes is a broader scientific and industry conversation about whether grid planning itself is using the wrong assumptions. Researchers highlighted in one report argue that conventional models for forecasting cooling demand fail to properly account for humidity, meaning grids may be designed around outdated or incomplete climate data — a mismatch that could leave systems unprepared for the kind of heat-driven demand surges now becoming routine 6. And on the supply-chain side of the AI buildout, coverage of Texas Instruments frames the company's analog chips as a crucial link between the physical grid and the power-hungry processors driving AI workloads, underscoring how deeply the semiconductor industry is now entangled with electricity infrastructure 7.

Where the reporting agrees

Across these accounts, one point is consistent: electricity demand is climbing faster than the grid's ability to comfortably absorb it. The Yahoo-sourced item on data centers states plainly that demand growth is now outpacing supply growth 1, and the record-setting figures from ERCOT 2 and PJM 4 are concrete, real-world evidence of that same dynamic playing out in two of the country's largest grid territories within a similar window. The Colorado Springs reporting, though describing a smaller regional utility rather than a massive interconnection, tells the same basic story at a different scale — grid strain severe enough to prompt a formal public appeal for conservation 35. Multiple outlets, in other words, are documenting the identical phenomenon — demand outrunning capacity — through different geographic lenses, whether that's a statewide record in Texas, a multi-state peak in the PJM footprint, or a single utility in Colorado asking residents to ease up.

Where it doesn't

The sources diverge mainly in scope and specificity rather than in direct contradiction. The 20%-by-2035 data center figure and its Virginia/Texas bill impact come from a single account 1, and no other source in this set independently verifies that particular projection or timeline — it stands as a claim attributed to that reporting alone. Similarly, the humidity-based critique of grid planning models 6 is a distinct, standalone thread; none of the grid-operator stories about ERCOT, PJM, or Colorado Springs reference or confirm that methodology, so it reads as parallel context rather than corroborated fact. The Texas Instruments piece 7 also operates in its own lane, focused on corporate and investment implications rather than grid strain itself, making it more of an adjacent data point about the AI-chip-to-power pipeline than a direct confirmation of the demand figures reported elsewhere.

The bottom line

Taken together, the evidence supports treating the demand-outpacing-supply narrative as solidly established — the record peaks in Texas and the PJM region are hard operational data points that back up the broader warning about data centers and AI reshaping electricity consumption. The specific 20%-by-2035 forecast and the humidity-modeling critique are best read as important but not yet independently corroborated pieces of the picture, worth watching rather than treating as settled consensus.