US data centers could use one fifth of national electricity by 2035

Data centers could consume about one-fifth of US electricity by 2035, according to a BloombergNEF estimate cited by TechCrunch, up roughly fourfold from today as AI training and inference move from novelty to baseline infrastructure. The report projects data center capacity approaching 200 gigawatts over the next decade, with nearly half devoted to AI workloads. The forecast itself has been revised sharply upward: BloombergNEF’s 2035 demand estimate is 83% higher than what it expected as recently as December.

The numbers matter less as a single prediction than as a description of a development pipeline that is already colliding with grid rules. BloombergNEF expects most new data centers to connect to electrical systems that are already strained, and it points to PJM — the vast interconnection spanning Virginia to Illinois — as the clearest example. In PJM, data centers are projected to account for 34% of electricity used on the system, while in Texas’s ERCOT they are expected to require 22% of generating capacity.

Grid operators are responding with the tools they have: queues, delays, and rationing by procedure. TechCrunch notes that PJM paused applications for new sources to connect to the grid for four years, reopening its queue only in April, after connection requests from both large generators and large loads overwhelmed the process. The bottleneck shows up in prices. In PJM, the supply-demand imbalance has pushed electricity prices up 76% over the past year, and data centers represented 38% of charges in the grid’s most recent capacity auction — a sign that the cost of guaranteeing power is increasingly being set by a single buyer class.

For companies building data centers, the economics are straightforward: secure power first, then build. For everyone else on the same grid — households, factories, and smaller businesses — the risk is that the bill is negotiated through regulatory filings and capacity auctions rather than through a direct contract with the new load. The report also implies a geographic sorting effect: regions with faster permitting, spare transmission, or more tolerant regulators become magnets for investment, even if they were not designed to host energy-intensive industry.

The buildout is not confined to the US, but the concentration remains lopsided. BloombergNEF expects the US to host 64% of AI chips by power demand by 2033, even as global data centers add an estimated 1,935 terawatt-hours of new electricity demand under aggressive AI adoption — nearly as much as India uses in a year.

PJM’s queue reopened in April, and electricity prices on the system are already up sharply. Data centers are still filing to connect.