BloombergNEF: U.S. Data Centers to Account for 20% of National Power Consumption by 2035
Claire Weston
BloombergNEF projects US data centers will consume 20% of national electricity by 2035, up from 5.9% today — meaning AI infrastructure is now competing directly with homes and EVs for the same grid capacity.
From 6% to 20% — how big is the jump?
US data centers currently draw 5.9% of national power; that rises to 12% by 2030 and 20% by 2035.
Absolute demand hits 194 GW — one gigawatt equals roughly one conventional nuclear plant.
This means → in ten years, one out of every five units of US electricity goes to data centers, drawn from the same pool that powers cities and charges EVs.
Why was the forecast revised up 83%?
BloombergNEF's 2035 power-demand estimate is 83% higher than its December 2025 projection.
This reflects a rapid expansion of the AI facility construction pipeline over just the past few months.
In plain terms = what looked like an aggressive build plan half a year ago now appears to have been an undercount.
Which regions face the most pressure?
Virginia and Texas host the densest clusters; their data-center share will exceed the national average.
PJM Interconnection (grid operator spanning 13 states, including Northern Virginia's "data center alley") and Texas's ERCOT grid are projected to surpass the national mean by 2035.
This means → grid-expansion stress in these two corridors is far more concentrated than headline national figures suggest.
How large is the supply gap?
The historical record for data-center grid interconnection in a single year is 7.1 GW.
Even sustaining that record pace, BloombergNEF still projects a 19 GW supply shortfall by 2035.
In plain terms = at the fastest build rate ever achieved, the grid will still fall short by the equivalent of 19 nuclear plants.
What happens if the gap isn't closed?
Operators are already pursuing on-site power generation as an alternative to relying solely on the public grid.
They simultaneously face local permitting freezes and political pushback over energy use and environmental impact.
This reflects a core validation point for this AI infrastructure cycle: if the grid can't scale before demand materializes, compute capacity simply cannot land.
Content is for reference only, not financial advice.