energy-consumption

This paper explores how prompt properties like cognitive load and phrasing pattern influence energy usage in on-device LLM inference, showing that cognitive load affects energy per token while phrasing impacts token usage, highlighting the need for model-aware prompt design for energy efficiency.

Elon Musk discusses humanity's minimal energy usage on the Kardashev scale compared to the sun's output.

This paper presents a systematic empirical study of energy consumption in large language model inference, analyzing how attention architectures like Multi-Head Attention, Grouped Query Attention, and Sliding Window Attention affect energy scaling across context lengths and workloads.

The GOP's Senate campaign arm has warned AI companies about the growing public opposition to data centers in Ohio, which is affecting a Senate election. Polls show high concerns over energy costs and rapid construction, prompting calls for better communication on benefits and impacts.

Inner Mongolia's Ulanqab has become a major hub for AI data centers in China, with companies like DeepSeek and ByteDance investing heavily in infrastructure, though water scarcity poses challenges.

This paper introduces MÖVE, a holistic evaluation framework for LLMs in the German public sector, examining governance dimensions like energy consumption, provider transparency, and knowledge of German-party positions, revealing trade-offs that necessitate context-specific model selection.

The article examines and debunks exaggerated claims about AI's environmental impact, using data from the IEA to show that while AI consumes energy and water, its effects are often overstated and context-dependent.

A discussion on the unsustainable energy and infrastructure demands of current AI systems, exploring potential alternatives like new mathematics, materials, or quantum computing.

The article argues that the massive AI data center buildout is a speculative bubble driven by subsidized pricing rather than real demand, and that the future of AI lies in smaller open-source models at the edge. It highlights negative impacts on energy grids and climate goals, warning that a bubble burst could cause a recession but not the end of AI.

A user inquires about the average water consumption per 100 tokens for frontier AI models, highlighting the environmental impact of AI.

A new report from BloombergNEF predicts data centers will quadruple their electricity use by 2035, consuming one-fifth of U.S. electricity, driven by surging AI compute demands and straining already burdened power grids.

A critical reflection on the AI industry's environmental costs and overhyped claims about chatbot capabilities, arguing that many harms are ignored and that not using AI remains an easy choice.

Irish datacenters consumed 23% of the country's electricity in 2025, up 10% from 2024, despite a moratorium on new grid connections. New regulations require large datacenter operators to provide backup generation and feed power back to the grid.

The article discusses the growing local opposition to AI data centers due to environmental and energy concerns, highlighting a historic case in Ireland and current widespread resistance.

Google's electricity consumption surged 12 TWh from 2024 to 2025, reaching 43 TWh, driven by generative AI infrastructure, causing exponential growth in emissions and undermining climate goals.

The author argues that current AI scaling methods, despite being the pinnacle of engineering, are woefully inefficient and will be viewed as primitive in hindsight, similar to how we now see 1960s mainframes.

Henrico County, Virginia, home to 37 data centers, is asking schools and government employees to conserve electricity after a 25% rate increase driven by data center power demands, highlighting the growing conflict between tech infrastructure and local communities.

An exploration of the water consumption associated with training and running AI models, often overlooked in discussions of AI's environmental footprint.

A UN report reveals that by 2030, AI data centers could consume water equivalent to the basic annual needs of 1.3 billion people, with 80-90% of energy used for daily operations rather than training. Generating a single AI image uses over 1000 times more energy than a basic text task.

Presents a systematic study of parameter-efficient fine-tuning using LoRA on Qwen2.5-3B for telecommunications customer support, comparing 16 LoRA configurations with both traditional metrics and energy consumption analysis. Finds divergence between quantitative and qualitative performance.