{
  "name": "How much water does AI use? Water, energy and carbon per prompt",
  "description": "How much water does AI use? Sourced per-prompt water, energy and carbon figures for Gemini, ChatGPT, Mistral and others, plus big tech data center water totals.",
  "url": "https://agitracker.io/data/ai-water-and-energy-use/",
  "license": "https://creativecommons.org/licenses/by/4.0/",
  "publisher": "agitracker.io",
  "updated": "2026-10-29",
  "columns": [
    {
      "key": "subject",
      "label": "Model or product"
    },
    {
      "key": "organization",
      "label": "Model developer"
    },
    {
      "key": "publisher",
      "label": "Published by"
    },
    {
      "key": "water_ml",
      "label": "Water per prompt (mL)",
      "type": "number",
      "note": "Dash means not published"
    },
    {
      "key": "energy_wh",
      "label": "Energy per prompt (Wh)",
      "type": "number"
    },
    {
      "key": "carbon_g",
      "label": "Carbon (gCO2e)",
      "type": "number"
    },
    {
      "key": "water_scope",
      "label": "What the water figure covers"
    },
    {
      "key": "prompt_type",
      "label": "Prompt measured"
    },
    {
      "key": "figure_type",
      "label": "Figure type"
    },
    {
      "key": "date",
      "label": "Published",
      "type": "date"
    }
  ],
  "rows": [
    {
      "subject": "Gemini Apps (comprehensive method)",
      "organization": "Google",
      "publisher": "Google",
      "water_ml": 0.26,
      "energy_wh": 0.24,
      "carbon_g": 0.03,
      "water_scope": "On-site data center cooling water consumed only (fleet WUE of 1.15 L/kWh); excludes water used to generate electricity. Carbon is market-based.",
      "prompt_type": "Median Gemini Apps text prompt, May 2025; includes idle capacity, host CPU/RAM and data center overhead",
      "figure_type": "Company disclosure (with technical paper, not third-party verified)",
      "date": "2025-08-21",
      "source": {
        "title": "Google Cloud blog: Measuring the environmental impact of AI inference",
        "url": "https://cloud.google.com/blog/products/infrastructure/measuring-the-environmental-impact-of-ai-inference"
      }
    },
    {
      "subject": "Gemini Apps (accelerator-only method)",
      "organization": "Google",
      "publisher": "Google",
      "water_ml": 0.12,
      "energy_wh": 0.1,
      "carbon_g": 0.02,
      "water_scope": "On-site cooling water only, counting active TPU/GPU energy alone; Google calls this narrower method an underestimate",
      "prompt_type": "Median Gemini Apps text prompt, May 2025",
      "figure_type": "Company disclosure",
      "date": "2025-08-21",
      "source": {
        "title": "Google: Measuring the energy, emissions, and water impact of Gemini prompts (arXiv 2508.15734)",
        "url": "https://arxiv.org/abs/2508.15734"
      }
    },
    {
      "subject": "ChatGPT",
      "organization": "OpenAI",
      "publisher": "Sam Altman (OpenAI CEO)",
      "water_ml": 0.32,
      "energy_wh": 0.34,
      "carbon_g": null,
      "water_scope": "Not stated. Published as 0.000085 gallons (about 1/15 teaspoon), shown here converted to mL",
      "prompt_type": "\"Average query\"; model and boundary not defined",
      "figure_type": "Company disclosure (CEO blog post, no methodology)",
      "date": "2025-06-10",
      "source": {
        "title": "Sam Altman: The Gentle Singularity",
        "url": "https://blog.samaltman.com/the-gentle-singularity"
      }
    },
    {
      "subject": "Le Chat (Mistral Large 2)",
      "organization": "Mistral AI",
      "publisher": "Mistral AI, with Carbone 4 and ADEME",
      "water_ml": 45,
      "energy_wh": null,
      "carbon_g": 1.14,
      "water_scope": "Lifecycle assessment (water consumption potential), marginal impact of inference; excludes users' devices",
      "prompt_type": "One 400-token response",
      "figure_type": "Company disclosure (lifecycle analysis peer-reviewed by Resilio and Hubblo)",
      "date": "2025-07-22",
      "source": {
        "title": "Mistral AI: Our contribution to a global environmental standard for AI",
        "url": "https://mistral.ai/news/our-contribution-to-a-global-environmental-standard-for-ai"
      }
    },
    {
      "subject": "Frontier-scale LLMs (>200B parameters) on H100",
      "organization": "Various (open models modeled by Microsoft)",
      "publisher": "Microsoft researchers (Oviedo et al.)",
      "water_ml": null,
      "energy_wh": 0.31,
      "carbon_g": null,
      "water_scope": "No point value. Microsoft's summary estimates 0.0 to 0.067 mL of cooling water per typical query",
      "prompt_type": "Typical query (500 input, about 300 output tokens), median; interquartile range 0.16 to 0.60 Wh",
      "figure_type": "Peer-reviewed estimate (Joule)",
      "date": "2026-04-22",
      "source": {
        "title": "Oviedo et al., Energy use of AI inference, efficiency pathways, and test-time scaling, Joule",
        "url": "https://www.cell.com/joule/fulltext/S2542-4351(26)00114-5"
      }
    },
    {
      "subject": "Frontier-scale LLMs, long reasoning query",
      "organization": "Various (open models modeled by Microsoft)",
      "publisher": "Microsoft researchers (Oviedo et al.)",
      "water_ml": null,
      "energy_wh": 3.91,
      "carbon_g": null,
      "water_scope": "Not estimated",
      "prompt_type": "Test-time scaling query (about 5,000 output tokens), median; interquartile range 2.15 to 7.05 Wh",
      "figure_type": "Peer-reviewed estimate (Joule)",
      "date": "2026-04-22",
      "source": {
        "title": "Oviedo et al., Energy use of AI inference, efficiency pathways, and test-time scaling, Joule",
        "url": "https://www.cell.com/joule/fulltext/S2542-4351(26)00114-5"
      }
    },
    {
      "subject": "GPT-3 (on-site water only)",
      "organization": "OpenAI",
      "publisher": "Li, Yang, Islam, Ren (UC Riverside, UT Arlington)",
      "water_ml": 2.2,
      "energy_wh": 4,
      "carbon_g": null,
      "water_scope": "On-site cooling water consumed (scope 1), U.S. average Microsoft data center",
      "prompt_type": "Medium-length request (up to about 800 words in, 150 to 300 words out), assumed 4 Wh per request",
      "figure_type": "Peer-reviewed estimate (Communications of the ACM, 2025; first posted 2023)",
      "date": "2025-03-26",
      "source": {
        "title": "Li et al., Making AI Less \"Thirsty\" (arXiv 2304.03271), Table 1",
        "url": "https://arxiv.org/abs/2304.03271"
      }
    },
    {
      "subject": "GPT-3 (on-site plus power plant water)",
      "organization": "OpenAI",
      "publisher": "Li, Yang, Islam, Ren (UC Riverside, UT Arlington)",
      "water_ml": 16.904,
      "energy_wh": 4,
      "carbon_g": null,
      "water_scope": "On-site cooling (2.2 mL) plus off-site water consumed to generate electricity (14.7 mL), U.S. average; source of the \"500 mL per 10 to 50 responses\" figure",
      "prompt_type": "Medium-length request, assumed 4 Wh per request",
      "figure_type": "Peer-reviewed estimate (Communications of the ACM, 2025; first posted 2023)",
      "date": "2025-03-26",
      "source": {
        "title": "Li et al., Making AI Less \"Thirsty\" (arXiv 2304.03271), Table 1",
        "url": "https://arxiv.org/abs/2304.03271"
      }
    },
    {
      "subject": "ChatGPT (GPT-4), 100-word email",
      "organization": "OpenAI",
      "publisher": "The Washington Post with UC Riverside researchers",
      "water_ml": 519,
      "energy_wh": 140,
      "carbon_g": null,
      "water_scope": "Not fully stated in the article; varies by data center location, for example 235 mL in Texas and 1,468 mL in Washington state",
      "prompt_type": "Writing one 100-word email with GPT-4",
      "figure_type": "Press estimate",
      "date": "2024-09-18",
      "source": {
        "title": "The Washington Post: A bottle of water per email",
        "url": "https://www.washingtonpost.com/technology/2024/09/18/energy-ai-use-electricity-water-data-centers/"
      }
    },
    {
      "subject": "ChatGPT request",
      "organization": "OpenAI",
      "publisher": "EPRI",
      "water_ml": null,
      "energy_wh": 2.9,
      "carbon_g": null,
      "water_scope": "Not estimated",
      "prompt_type": "One ChatGPT request; EPRI cites earlier third-party estimates and compares it with 0.3 Wh for a Google search",
      "figure_type": "Analyst estimate",
      "date": "2024-05",
      "source": {
        "title": "EPRI: Powering Intelligence (May 2024)",
        "url": "https://www.epri.com/research/products/000000003002028905"
      }
    },
    {
      "subject": "GPT-4o (March 2025)",
      "organization": "OpenAI",
      "publisher": "Jegham et al. (\"How Hungry is AI?\")",
      "water_ml": null,
      "energy_wh": 0.423,
      "carbon_g": null,
      "water_scope": "Paper's water figures count on-site cooling plus power plant water consumed; per-model values shown only in charts",
      "prompt_type": "Short prompt (100 input, 300 output tokens), mean",
      "figure_type": "Preprint estimate",
      "date": "2025-11-24",
      "source": {
        "title": "Jegham et al., How Hungry is AI? (arXiv 2505.09598, v6), Table 4",
        "url": "https://arxiv.org/abs/2505.09598"
      }
    },
    {
      "subject": "GPT-4.1 nano",
      "organization": "OpenAI",
      "publisher": "Jegham et al. (\"How Hungry is AI?\")",
      "water_ml": null,
      "energy_wh": 0.207,
      "carbon_g": null,
      "water_scope": "Paper reports under 4 mL even for long prompts for this and other small models",
      "prompt_type": "Short prompt (100 input, 300 output tokens), mean",
      "figure_type": "Preprint estimate",
      "date": "2025-11-24",
      "source": {
        "title": "Jegham et al., How Hungry is AI? (arXiv 2505.09598, v6), Table 4",
        "url": "https://arxiv.org/abs/2505.09598"
      }
    },
    {
      "subject": "Claude 3.7 Sonnet",
      "organization": "Anthropic",
      "publisher": "Jegham et al. (\"How Hungry is AI?\")",
      "water_ml": null,
      "energy_wh": 0.95,
      "carbon_g": null,
      "water_scope": "Not given as a number for this model",
      "prompt_type": "Short prompt (100 input, 300 output tokens), mean",
      "figure_type": "Preprint estimate",
      "date": "2025-11-24",
      "source": {
        "title": "Jegham et al., How Hungry is AI? (arXiv 2505.09598, v6), Table 4",
        "url": "https://arxiv.org/abs/2505.09598"
      }
    },
    {
      "subject": "o3",
      "organization": "OpenAI",
      "publisher": "Jegham et al. (\"How Hungry is AI?\")",
      "water_ml": null,
      "energy_wh": 12.222,
      "carbon_g": null,
      "water_scope": "Not given as a number for this model",
      "prompt_type": "Long prompt (10,000 input, 1,500 output tokens), mean",
      "figure_type": "Preprint estimate",
      "date": "2025-11-24",
      "source": {
        "title": "Jegham et al., How Hungry is AI? (arXiv 2505.09598, v6), Table 4",
        "url": "https://arxiv.org/abs/2505.09598"
      }
    },
    {
      "subject": "DeepSeek-R1 (DeepSeek's own servers)",
      "organization": "DeepSeek",
      "publisher": "Jegham et al. (\"How Hungry is AI?\")",
      "water_ml": null,
      "energy_wh": 29.078,
      "carbon_g": 17,
      "water_scope": "Paper reports \"over 200 mL\" (on-site plus power plant water consumed) without an exact value",
      "prompt_type": "Long prompt (10,000 input, 1,500 output tokens), mean",
      "figure_type": "Preprint estimate",
      "date": "2025-11-24",
      "source": {
        "title": "Jegham et al., How Hungry is AI? (arXiv 2505.09598, v6)",
        "url": "https://arxiv.org/abs/2505.09598"
      }
    },
    {
      "subject": "DeepSeek-R1 (hosted on Microsoft Azure)",
      "organization": "DeepSeek",
      "publisher": "Jegham et al. (\"How Hungry is AI?\")",
      "water_ml": 34,
      "energy_wh": 7.41,
      "carbon_g": 2.5,
      "water_scope": "On-site cooling plus power plant water consumed",
      "prompt_type": "Long prompt (10,000 input, 1,500 output tokens), mean",
      "figure_type": "Preprint estimate",
      "date": "2025-11-24",
      "source": {
        "title": "Jegham et al., How Hungry is AI? (arXiv 2505.09598, v6)",
        "url": "https://arxiv.org/abs/2505.09598"
      }
    },
    {
      "subject": "Llama 3.1 8B",
      "organization": "Meta",
      "publisher": "Jegham et al. (\"How Hungry is AI?\")",
      "water_ml": null,
      "energy_wh": 0.052,
      "carbon_g": null,
      "water_scope": "Paper reports under 4 mL even for long prompts for this and other small models",
      "prompt_type": "Short prompt (100 input, 300 output tokens), mean",
      "figure_type": "Preprint estimate",
      "date": "2025-11-24",
      "source": {
        "title": "Jegham et al., How Hungry is AI? (arXiv 2505.09598, v6), Table 4",
        "url": "https://arxiv.org/abs/2505.09598"
      }
    }
  ]
}