AI Energy
Power Usage Effectiveness (PUE)
PUE is the dominant efficiency metric for data centers: the ratio of a facility's total electrical energy to the energy delivered to IT equipment. Defined by {{The Green Grid}} in 2007 and codified as {{ISO/IEC 30134-2}}, it has an ideal value of 1.0. Industry surveys put the global average around 1.5-1.6, while hyperscalers such as Google report fleet-wide PUE near 1.09-1.10. Critics note that PUE ignores water consumption, on-site generation losses, and whether the IT load itself is doing useful work.
Water Usage Effectiveness (WUE)
WUE is The Green Grid's companion metric to PUE, measuring annual site water consumption per kilowatt-hour of IT energy. It captures evaporative cooling and humidification draws that PUE ignores, and is increasingly tracked as data centers expand in arid regions.
Power Hungry Processing (Luccioni et al. 2024)
ACM FAccT 2024 paper by Luccioni, Jernite, and Strubell that benchmarked inference-time energy use across 88 models and 10 task types. Found that energy per request varies by more than four orders of magnitude, with general-purpose generative models typically far more costly than task-specific fine-tuned models.
Energy Cost Per LLM Query
Current best estimates put the electricity cost of a standard ChatGPT-class text query at roughly 0.3 to 3 watt-hours, versus about 0.3 Wh for a Google search. Early 2023-2024 figures (de Vries, EPRI, IEA) clustered near the high end at ~2.9 Wh, while 2025 measurements from Epoch AI, OpenAI, and Google converged at 0.24-0.34 Wh as hardware and models improved. Variation by task type and reasoning depth can still span two orders of magnitude.
Training vs Inference Emissions in Large Language Models
Training a frontier large language model emits hundreds of tonnes of CO2-equivalent — {{GPT-3}} at roughly 552 tCO2eq, [[BLOOM]] at about 25 tCO2eq operational plus ~25 tCO2eq {{embodied carbon}} — but for widely deployed models, cumulative {{inference}} emissions from millions of daily queries surpass one-time training within weeks to months.
Embodied Energy of AI Training Hardware
Manufacturing an AI accelerator like the NVIDIA H100 or A100 consumes energy and emits carbon long before the chip is ever powered on. Wafer fabrication at advanced nodes, rare-mineral extraction, packaging, and global logistics together produce the chip's embodied carbon, which must be amortized over a typical 3-to-5-year service lifetime and compared against operational electricity use to judge real-world impact.
Jevons Paradox in AI Inference
William Stanley Jevons observed in 1865 that more efficient coal use increased rather than decreased total coal consumption, because lower effective prices unlocked new applications. The same dynamic appears in AI: falling per-token inference costs — driven by distillation, MoE routing, and hardware gains — enable agent loops, always-on classifiers, and chat-as-search to replace cheaper substitutes, pushing aggregate compute and energy use up despite per-query efficiency wins.