Key facts
- Data centers are projected to consume 20% of U.S. electricity by 2035, a fourfold increase.
- AI compute demand is expected to drive data center capacity to nearly 200 gigawatts in the next decade.
- The U.S. will host the majority of AI chips by power demand.
- Strained electrical grids, like PJM, face significant data center load.
- Electricity prices in some strained grids have increased substantially.
Data centers are poised to dramatically increase their electricity consumption, with U.S. usage expected to quadruple by 2035 to one-fifth of the nation's total power generation, according to a new report from BloombergNEF. This surge is primarily driven by the escalating demand for AI compute power, which is projected to push data center capacity to nearly 200 gigawatts over the next decade.
Nearly half of this new capacity will be dedicated to AI training and inference, with the U.S. expected to host the majority of AI chips by power demand, accounting for 64% by 2033. The report's revised estimate for 2035 electricity demand is 83% higher than previous predictions, reflecting a broader trend of upward revisions by other organizations like EPRI and S&P.
The expansion of data centers is placing significant strain on existing electrical grids. The PJM Interconnection, a vast grid spanning from Virginia to Illinois, anticipates data centers will consume 34% of its electricity, while ERCOT in Texas expects them to use 22% of its capacity. PJM has already faced challenges managing connection requests, leading to a pause in new applications and a substantial increase in electricity prices over the past year.
Despite these grid constraints, data centers continue to seek connections, representing a significant portion of capacity auction charges. Globally, the demand for electricity from data centers could reach 1,935 terawatt-hours by 2033 if AI adoption follows an aggressive path, a figure comparable to India's current annual electricity usage.