[
  {
    "organization": "Google",
    "model": "Gemini Apps, median text prompt (comprehensive method)",
    "energy_wh": 0.24,
    "water_ml": 0.26,
    "emissions_g": 0.03,
    "covers": "Inference only. Includes active accelerators, host CPU and RAM, idle machines and data center overhead. Water is on-site consumption, calculated as energy excluding data center overhead multiplied by the Google WUE of 1.15 L/kWh",
    "source_type": "Company disclosure (technical paper)",
    "date": "2025-08-21 (data from May 2025)",
    "source": "https://arxiv.org/pdf/2508.15734v1#page=7",
    "notes": "Also in the Google Cloud blog of August 22, 2025 and the Google 2026 Environmental Report. Excludes training, networking and user devices. Google says energy per median prompt fell 33x over the prior 12 months."
  },
  {
    "organization": "Google",
    "model": "Gemini Apps, median text prompt (narrow method, active accelerators only)",
    "energy_wh": 0.1,
    "water_ml": 0.12,
    "emissions_g": 0.02,
    "covers": "Inference only. Active TPU/GPU energy only, comparable to many earlier public estimates",
    "source_type": "Company disclosure (technical paper)",
    "date": "2025-08-21 (data from May 2025)",
    "source": "https://arxiv.org/pdf/2508.15734v1#page=7",
    "notes": "Table 2 of the paper and the Google Cloud blog give 0.10 Wh, 0.02 gCO2e and 0.12 mL. One sentence in the paper text instead gives 0.01 gCO2e and 0.15 mL for this method; this table uses the Table 2 values."
  },
  {
    "organization": "Hugging Face (Delavande, Luccioni, Pierrard)",
    "model": "WAN2.1-T2V-14B, one short video clip",
    "energy_wh": 109.47,
    "water_ml": "",
    "emissions_g": "",
    "covers": "Inference only, measured on one H100 with CodeCarbon; GPU 93.83 + CPU 10.47 + RAM 5.17 Wh",
    "source_type": "Research blog (not peer-reviewed)",
    "date": "2025-07-02",
    "source": "https://huggingface.co/blog/jdelavande/text-to-video-energy-cost#:~:text=WAN2.1%2DT2V%2D14B%20burns%20almost%2094%20Wh",
    "notes": "Clip of 81 frames at 480x832. Excludes data center overhead."
  },
  {
    "organization": "Hugging Face (Delavande, Luccioni, Pierrard)",
    "model": "AnimateDiff, one short video clip",
    "energy_wh": 0.14,
    "water_ml": "",
    "emissions_g": "",
    "covers": "Inference only, measured on one H100; GPU 0.11 + CPU 0.02 + RAM 0.01 Wh",
    "source_type": "Research blog (not peer-reviewed)",
    "date": "2025-07-02",
    "source": "https://huggingface.co/blog/jdelavande/text-to-video-energy-cost#:~:text=AnimateDiff%20uses%20barely%200.11%20Wh",
    "notes": "Clip of 16 frames at 512x512. Shows the roughly 800x spread between open video models."
  },
  {
    "organization": "Jegham et al.",
    "model": "DeepSeek-R1 on DeepSeek servers, long prompt (10k input, 1.5k output tokens)",
    "energy_wh": 29.08,
    "water_ml": "over 200",
    "emissions_g": "about 17",
    "covers": "Inference only, estimated. Water includes on-site and off-site (electricity generation) components",
    "source_type": "Preprint (arXiv v6)",
    "date": "2025-11-24",
    "source": "https://arxiv.org/pdf/2505.09598v6#page=9",
    "notes": "The same model served on Azure was estimated at 7.41 Wh, about 34 mL and 2.5 gCO2e per long query."
  },
  {
    "organization": "Jegham et al.",
    "model": "GPT-5, long query with high reasoning effort",
    "energy_wh": 33.8,
    "water_ml": "",
    "emissions_g": "",
    "covers": "Inference only, estimated average",
    "source_type": "Preprint (arXiv v6)",
    "date": "2025-11-24",
    "source": "https://arxiv.org/pdf/2505.09598v6#page=11",
    "notes": "A short query with minimal reasoning was estimated at 0.67 Wh."
  },
  {
    "organization": "Jegham et al. (University of Rhode Island and others)",
    "model": "GPT-4o, short prompt (100 input, 300 output tokens)",
    "energy_wh": 0.42,
    "water_ml": "",
    "emissions_g": "",
    "covers": "Inference only, estimated from public API throughput plus assumed hardware, PUE, WUE and grid factors",
    "source_type": "Preprint (arXiv v6)",
    "date": "2025-11-24",
    "source": "https://arxiv.org/pdf/2505.09598v6#page=8",
    "notes": "Table 4 gives 0.423 plus or minus 0.085 Wh; text gives 0.42 Wh plus or minus 0.13 Wh, or 0.37 Wh without data center overhead. Title: How Hungry is AI?"
  },
  {
    "organization": "Luccioni, Jernite and Strubell",
    "model": "Open text generation models, mean per query",
    "energy_wh": 0.047,
    "water_ml": "",
    "emissions_g": "",
    "covers": "Inference only, measured on A100 GPUs; reported as 0.047 kWh per 1,000 inferences",
    "source_type": "Peer-reviewed (ACM FAccT 2024)",
    "date": "2024-06",
    "source": "https://arxiv.org/pdf/2311.16863v3#page=6",
    "notes": "Models tested are far smaller than current frontier chat models."
  },
  {
    "organization": "Luccioni, Jernite and Strubell (Hugging Face, Carnegie Mellon)",
    "model": "Open image generation models, mean per image",
    "energy_wh": 2.907,
    "water_ml": "",
    "emissions_g": "",
    "covers": "Inference only, measured on A100 GPUs; reported as 2.907 kWh per 1,000 inferences (standard deviation 3.31)",
    "source_type": "Peer-reviewed (ACM FAccT 2024)",
    "date": "2024-06",
    "source": "https://arxiv.org/pdf/2311.16863v3#page=6",
    "notes": "The least efficient model tested used 11.49 kWh per 1,000 images (about 11.5 Wh each). Small open models, not production services."
  },
  {
    "organization": "Microsoft",
    "model": "Large production LLM, typical query (company blog on the Joule study)",
    "energy_wh": "0.16 to 0.60",
    "water_ml": "0.0 to 0.067",
    "emissions_g": "",
    "covers": "Inference only. On-site cooling water per query under conservative assumptions",
    "source_type": "Company disclosure (blog)",
    "date": "2026-06-15",
    "source": "https://www.microsoft.com/en-us/microsoft-cloud/blog/2026/06/15/scaling-ai-with-8-to-20x-energy-efficiency/#:~:text=0.0%20to%200.067%20mL%20of%20water",
    "notes": "Microsoft describes the median water use as about one hundredth of a teaspoon."
  },
  {
    "organization": "Microsoft Research",
    "model": "Frontier-scale LLM, typical query under optimized production serving",
    "energy_wh": 0.31,
    "water_ml": "",
    "emissions_g": "",
    "covers": "Inference only. Bottom-up model of energy per query; interquartile range 0.16 to 0.60 Wh",
    "source_type": "Peer-reviewed (Joule)",
    "date": "2026-04",
    "source": "https://www.microsoft.com/en-us/research/publication/energy-use-of-ai-inference-efficiency-pathways-and-test-time-scaling/#:~:text=median%200.31%20Wh",
    "notes": "Oviedo et al., Joule, doi 10.1016/j.joule.2026.102430. Summarized in a Nature Energy research highlight on May 27, 2026 (https://www.nature.com/articles/s41560-026-02084-9). Longer reasoning or agentic queries use more."
  },
  {
    "organization": "Mistral AI",
    "model": "Le Chat, one 400-token response (Mistral Large 2)",
    "energy_wh": "",
    "water_ml": 45,
    "emissions_g": 1.14,
    "covers": "Marginal inference impact from a life cycle assessment, including upstream server manufacturing, excluding user devices",
    "source_type": "Company disclosure (LCA with Carbone 4 and ADEME, peer-reviewed by Resilio and Hubblo)",
    "date": "2025-07-22",
    "source": "https://mistral.ai/news/our-contribution-to-a-global-environmental-standard-for-ai/#:~:text=45%20mL%20of%20water",
    "notes": "No energy per prompt disclosed. The same study puts training plus 18 months of use of Mistral Large 2 at 20.4 ktCO2e and 281,000 m3 of water (as of January 2025). Water figure includes water embodied in hardware, so it is not comparable with on-site cooling figures."
  },
  {
    "organization": "OpenAI",
    "model": "ChatGPT, average query",
    "energy_wh": 0.34,
    "water_ml": 0.32,
    "emissions_g": "",
    "covers": "Not stated. CEO blog statement without published methodology",
    "source_type": "CEO blog statement (no methodology)",
    "date": "2025-06",
    "source": "https://blog.samaltman.com/the-gentle-singularity#:~:text=the%20average%20query%20uses%20about%200.34%20watt%2Dhours",
    "notes": "Stated as about 0.000085 gallons of water, roughly one fifteenth of a teaspoon; 0.32 mL is our unit conversion. Model, scope (on-site or including power generation water) and date range are not specified."
  }
]