.jpg)
A new scenario study estimates that data centers operated by six large technology companies could use 239 to 295 terawatt-hours of electricity in 2030, up from about 118 terawatt-hours in 2024. That upper range is close to 1 percent of projected global electricity demand, but it is a modelled trajectory rather than a measurement of committed facilities.
The researchers examined Amazon, Microsoft, Google, Meta, Oracle and Apple. They built a retrieval system that searches corporate disclosures, policy documents and other public sources, then used a large language model to classify likely AI-oriented sites and extract signals about expansion plans. Those inferred locations were combined with firm-level operating estimates and regional electricity data.
The projection uses three rates for annual growth in new AI load: 15, 25 and 35 percent. Existing data-center consumption grows at 10 percent a year in every scenario, while assumed efficiency improvements and the share of AI-intensive workloads vary. The study also defines an electricity demand pressure index, or EDPI, as projected data-center demand divided by the electricity available in a region from generation and positive net imports.
More than 90 percent of the projected computing capacity is concentrated in North America, Western Europe and the Asia-Pacific, and fewer than ten regions account for nearly two-thirds of projected demand. Oregon, Ireland and Iowa rank high on the EDPI measure because the estimated loads are large relative to their electricity systems. Texas has more absolute demand but a lower ratio because its available supply is much larger.
The estimates depend on several uncertain layers. Site classification and expansion intent are inferred from uneven public documents; smaller cloud, colocation, telecom and regional operators are excluded; and the three growth paths are assumptions, not observed orders. Regional electricity supply is extrapolated from 2019 to 2024 trends, a short period affected by the pandemic. EDPI measures annual demand concentration, so it does not simulate transmission bottlenecks, hourly peaks, prices or demand response.
The authors have released their processed data and code, which makes the scenarios inspectable.