Data centers' energy consumption set to double by 2030

Data centers currently consume 414 TWh annually, accounting for roughly 1.5% of global electricity demand. With growth rates hitting 17% in 2025, projections suggest power requirements could more than double by 2030. This surge is tied to the physical demands of Large Language Models (LLMs), which rely on transformer architectures and power-hungry GPUs to execute billions of calculations for every word generated. The energy cost of a single interaction is driven by the computational complexity required to predict the next token in a sequence. While the input prompt initiates the process, the volume of the output-the text the AI writes back-is the primary driver of energy flux. This suggests that the ecological footprint of AI is not just a function of the model's size, but of the specific way users interact with it. Reducing the Computational Load Analysis by Ivan Drobnyak from University College London identifies specific shifts in usage that alter the resource baseline. The most significant reduction comes from switching to small, specialized models for narrow tasks like translation or summarization. These models can consume 15 to 50 times less energy than general-purpose giants like Llama 3.1, in some cases reducing total energy expenditure by 90% while maintaining or improving quality.