Energy Cost Per LLM Query

Estimating how much electricity a single large-language-model query consumes has become one of the most contested numbers in AI sustainability. Public estimates for a standard ChatGPT-class text query span roughly 0.3 to 3 watt-hours (Wh), with the range driven by model size, prompt and output length, hardware generation, and accounting choices about what to include (e.g., idle GPU time, cooling overhead, embodied hardware emissions). The high end of the range comes from Alex de Vries's 2023 commentary in the journal Joule, which extrapolated from NVIDIA server shipments and OpenAI usage to estimate roughly 2.9 Wh per ChatGPT request. The EPRI 2024 report "Powering Intelligence" cited a similar figure, around 2.9 Wh, and contrasted it with about 0.3 Wh for a typical Google Search query — the basis for the widely-repeated "ten times a Google search" claim. The IEA Electricity 2024 report used the same 0.3 Wh search baseline and 2.9 Wh AI-query figure while projecting that data-center electricity demand could double between 2022 and 2026. More recent measurements suggest those early numbers were too high. A February 2025 Epoch AI analysis estimated a typical GPT-4o query at about 0.3 Wh, roughly ten times lower than de Vries's figure, attributing the gap to denser models, the move from A100 to H100 GPUs, and more realistic prompt-length assumptions. OpenAI's Sam Altman publicly stated 0.34 Wh as the average ChatGPT query in mid-2025, and Google's own 2025 study reported a median Gemini text query at 0.24 Wh. Variation within any single model can still be large. The Luccioni et al. "Power Hungry Processing" study (2024) showed that inference energy varies by orders of magnitude across task types: simple classification can take under 0.01 Wh, while open-ended text generation runs around 0.05 Wh and image generation averages near 3 Wh. Reasoning model queries, which produce long internal chain-of-thought tokens, can consume 50 to 100 times more energy than a plain chat response. Important caveats apply. None of the headline per-query figures include the one-time cost of model training, the embodied energy of GPU manufacturing, or water used for cooling. Providers do not publish their own metering data, so external estimates depend on assumed utilization rates and PUE (power usage effectiveness). The "average query" is itself ill-defined: a one-word reply and a 4,000-token essay differ by more than an order of magnitude, and provider-side optimisations such as Prompt Caching in LLMs: How Reusing Context Cuts Cost and Latency can change the marginal cost of a request dramatically.