Nature Energy study estimates AI data centers could add up to 140 TWh to US electricity use by 2030, highlights regional grid strains
Priya Nair ·
A team led by Lawrence Berkeley National Laboratory reports in Nature Energy today that AI workloads could add 80 to 140 terawatt hours to annual US electricity consumption by 2030, depending on efficiency gains and deployment pace. The peer reviewed analysis maps likely build outs to specific grid regions in the United States and Europe and finds that without transmission upgrades and demand management, nodes in PJM and ERCOT face the steepest constraints.
The paper matters for utilities and regulators that are processing a surge of siting applications from hyperscalers and colocation providers. It provides a transparent method to test how chip roadmaps, cooling choices, and training schedules translate into grid impacts and water use.
What the researchers modeled
The study, led by Arman Shehabi at Lawrence Berkeley National Laboratory with co authors from Northwestern University, the University of California Berkeley, and Imperial College London, builds a bottom up inventory of AI accelerators shipped by Nvidia, AMD, and custom silicon vendors. The team combines nameplate power, measured utilization factors for training and inference, and facility efficiency metrics to estimate total load and peak demand.
Researchers coupled these estimates to nodal power system data from the US Energy Information Administration and ENTSO E to allocate demand to likely regions based on announced campuses and land transactions. They ran three scenarios, conservative, reference, and accelerated adoption, and tested sensitivity to improvements such as liquid cooling, lower precision arithmetic, higher temperature operation, and optical interconnects.
Under the reference case, the authors estimate total US data center electricity use rising to roughly 300 to 340 terawatt hours in 2030, with AI contributing about one third of that total. Facilities that rely on evaporative cooling exhibit water use intensities near one to two liters per kilowatt hour, which implies tens of billions of liters of annual water withdrawals unless operators shift to water free systems.
Limits and implications
The authors note that they did not measure facility level consumption, and that results hinge on uncertain variables such as accelerator utilization, model size trends, and the pace of efficiency improvements. They report that demand response and flexible training could shift 15 to 25 percent of load to off peak hours, which eases local capacity constraints in several nodes.
For technology builders, the study reinforces that energy per token and water free thermal management are becoming design constraints alongside performance. For utilities and grid planners, it signals that interconnection queues, time of use rates, and transmission upgrades will shape where and how the next wave of AI capacity is built.