How to Tame AI's Voracious Appetite for Energy (nautil.us)

A recent analysis highlights the escalating energy demands of AI systems, particularly within data centers that power generative AI applications like Google's Gemini and ChatGPT. As generative AI becomes increasingly popular, the energy consumption of U.S. data centers has surged, consuming over 224 terawatt-hours in 2025 alone, which is a significant increase from prior years. Currently, processing a median-length prompt with Gemini requires about 0.24 watt-hours, which may seem minimal but accumulates across billions of daily queries. This growing demand fears a reliance on fossil fuels, underscoring the urgent need for sustainable energy solutions amidst tech companies’ investments in AI. Experts are advocating for energy-saving strategies, such as developing task-specific language models that consume significantly less power than larger models, and innovations like long short-term memory (LSTM) architectures, which reduce energy costs during processing. Furthermore, advancements in chip design, including wafer-scale chips, aim to enhance efficiency significantly. The AI/ML community faces a pivotal moment to reimagine energy use in AI, where software and hardware innovations could redefine the sustainability trajectory of future models. Encouragingly, discussions around energy-efficient designs point to a marked shift toward reducing AI's carbon footprint while maintaining computational capabilities.