A deep-research report on the environmental cost of an AI query
For a single LLM query the best-supported 2025 numbers come from Google's August 2025 technical report on Gemini. A median Gemini Apps text prompt consumes:
| Metric | "Comprehensive" | Narrow "existing" |
|---|---|---|
| Energy | 0.24 Wh | 0.10 Wh |
| Water | 0.26 mL (~5 drops) | 0.15 mL |
| CO2 | 0.03 gCO₂e | 0.01 gCO₂e |
These align in order of magnitude with independent estimates: Epoch AI (~0.3 Wh for a typical GPT-4o query, Feb 2025) and Sam Altman (0.34 Wh and ~0.32 mL water, June 2025).
Per-token figures are not published anywhere — they must be derived. Epoch assumes ~500 output tokens/query, which implies ~0.0006 Wh/token at 0.3 Wh.
For comparison baselines: EPA gives ~400 g CO₂/mile for an average gas car and 5.3 kg CO₂/therm of natural gas. The Chicago→Austin flight and per-winter-day heating figures were not independently verified.
Key caveat: the AI per-query numbers are largely vendor self-reports using favorable boundaries (median not mean, text-only, market-based carbon, training excluded). Treat them as best-case lower bounds.
confidence: high · vote 3-0
Google's primary technical report and Cloud blog (Aug 21, 2025) state verbatim 0.24 Wh for the median text prompt under the comprehensive methodology; Table 1 lists 0.10 Wh (10,000 prompts/kWh) under the existing approach. Corroborated by MIT Technology Review, TechRepublic, and Towards Data Science. Google's own analogy: running a standard microwave for about one second.
confidence: high · vote 3-0
Google states verbatim "0.26 mL (about five drops)" comprehensive and "0.15 mL" existing/narrow, based on a standard 0.05 mL drop. Critics dispute scope (excludes electricity-generation and indirect water) but not the on-site number itself.
confidence: high · vote 3-0
Google states 0.03 gCO₂e comprehensive and 0.01 gCO₂e existing. The paper uses market-based accounting (net 94 gCO₂e/kWh) vs a 2024 location-based factor of 345 gCO₂e/kWh; applying that ~3.67× ratio implies ~0.09 gCO₂e under location-based accounting (a derivation by Towards Data Science, not Google).
confidence: high · vote 3-0
Epoch AI (Feb 7, 2025, Josh You): "typical ChatGPT queries using GPT-4o likely consume roughly 0.3 watt-hours" — ten times less than the older ~3 Wh estimate. Methodology: ~200B-param MoE, ~100B active, 500 output tokens/query, H100s at ~1500W. Notes 10k-token inputs ~2.5 Wh and 100k-token queries ~40 Wh. Implied per-token energy: ~0.0006 Wh/token.
confidence: low · vote 2-1
Verified verbatim on Altman's blog (The Gentle Singularity, June 2025): "the average query uses about 0.34 watt-hours" and "0.000085 gallons of water." Accurate as an attribution, but an undisclosed-methodology CEO blog statement — not independently verified.
confidence: high · vote 3-0
arXiv:2304.03271 (accepted by CACM): a medium conversation = ≤800 words in / 150-300 words out; ~0.004 kWh server energy per GPT-3 request; GPT-3 consumes a 500 mL bottle of water per ~10-50 such responses — and "actual water consumption could be several times higher." These are GPT-3-era (2020-2023) figures, much higher-water than current models.
confidence: high · vote 3-0
EPA: 8,887 g CO₂/gallon at 22.8 mpg = 393 g/mile, rounded to "about 400 grams of CO₂ per mile" (tailpipe-only; upstream fuel production not included).
confidence: high · vote 3-0
EPA: 0.0053 metric tons CO₂/therm (combustion-only). An average US home: 39,319 cubic feet/year → ~2.16 metric tons CO₂/year → ~5.9 kg/day if spread evenly. Caveat: that annual figure averages across all US homes (incl. non-gas), so ~5.9 kg/day understates real winter daily use for an actual gas-heated home.
The strongest AI numbers (0.24 Wh / 0.26 mL / 0.03 gCO₂e) are Google first-party self-reports, not independently audited, using favorable boundary choices:
Per-token figures are not directly published and must be derived from assumed token counts. All AI figures are highly time-sensitive (efficiency improving fast; Google data is May 2025).
"GPT-3 consumes ~500 mL per 10-50 responses, implying ~10-50 mL per query." Killed because the 500 mL / 10-50-responses ratio does not cleanly invert to a single per-query value.