Data centers are projected to consume one-fifth of the electricity generated in the U.S. by 2035, a fourfold increase from current levels, according to a new report from BloombergNEF. The report predicts that surging demand for AI compute will propel data center capacity to nearly 200 gigawatts over the next decade, with nearly 50% allocated for training and inference, primarily in the U.S.
By 2033, the report estimates that the U.S. will host 64% of AI chips by power demand. BloombergNEF’s latest electricity demand estimate for 2035 is 83% higher than its previous prediction made in December. Other organizations, including EPRI and S&P, have also revised their electricity demand forecasts upward, reflecting rapid developments in data center infrastructure across the country.
In the coming years, most new data centers are expected to connect to electrical grids already under strain. The PJM Interconnection, which serves Virginia to Illinois, will allocate 34% of its electricity to data centers, while ERCOT, covering much of Texas, will allocate 22% of its generating capacity. PJM has faced challenges with connection requests, pausing new applications for four years, which has created precarious conditions amid rising demand.
Although PJM reopened its queue for new generating sources in April, the situation remains critical, leading American Electric Power to threaten to withdraw from the interconnection process. This supply-demand imbalance has contributed to a 76% rise in electricity prices over the past year.
Data centers represented 38% of charges in PJM’s latest capacity auction, indicating continued interest in connecting to the grid despite congestion issues. If aggressive AI adoption trends persist, worldwide data centers could contribute 1,935 terawatt-hours of new electricity demand by 2033, approaching India’s annual usage.
