{
  "dataset": "training-footprints",
  "title": "Published training footprints",
  "index": "Environmental Disclosure Index",
  "url": "https://aisafetytracker.org/environment/",
  "citation": "AI Safety Tracker, \"Environmental Disclosure Index\". https://aisafetytracker.org/environment/. Retrieved 23 October 2026.",
  "licences": [
    {
      "source": "environmental-disclosure-index",
      "name": "Environmental Disclosure Index",
      "publisher": "This site",
      "licence": "MIT",
      "licence_url": "https://opensource.org/licenses/MIT",
      "attribution": "AI Safety Tracker, \"Environmental Disclosure Index\". Licensed MIT."
    }
  ],
  "generated": "2026-10-23T13:22:31+00:00",
  "retrieved": "2026-10-23T00:00:00+00:00",
  "unit": "tCO2e",
  "notes": "Our own reading of primary publications, reviewed 2026-10-23. Every published training energy or emissions figure, with its basis. Energy for Meta models is derived from GPU hours and stated power.",
  "count": 9,
  "records": [
    {
      "model": "GPT-3 175B",
      "developer": "OpenAI",
      "year": 2020,
      "epoch_model": "GPT-3 175B (davinci)",
      "hardware": "NVIDIA V100",
      "gpu_hours": null,
      "device_w": null,
      "energy_mwh": 1287,
      "energy_derived": false,
      "tco2e": 552,
      "tco2e_basis": "location-based",
      "market_tco2e": null,
      "who": "third-party estimate",
      "scope": "Final training run, estimated by Google researchers from published details. OpenAI has published no figure.",
      "url": "https://arxiv.org/abs/2104.10350",
      "reviewed": "2026-10-23",
      "source_id": "environmental-disclosure-index"
    },
    {
      "model": "BLOOM 176B",
      "developer": "BigScience",
      "year": 2022,
      "epoch_model": "BLOOM-176B",
      "hardware": "NVIDIA A100",
      "gpu_hours": null,
      "device_w": null,
      "energy_mwh": 433,
      "energy_derived": false,
      "tco2e": 24.7,
      "tco2e_basis": "location-based",
      "market_tco2e": null,
      "who": "developer study",
      "scope": "Final training run, dynamic power only, on the Jean Zay supercomputer in France. 50.5 tCO2e including idle power and hardware manufacture.",
      "url": "https://arxiv.org/abs/2211.02001",
      "reviewed": "2026-10-23",
      "source_id": "environmental-disclosure-index"
    },
    {
      "model": "Llama 2 70B",
      "developer": "Meta",
      "year": 2023,
      "epoch_model": "Llama 2-70B",
      "hardware": "NVIDIA A100 80GB",
      "gpu_hours": 1720320,
      "device_w": 400,
      "energy_mwh": 688.1,
      "energy_derived": true,
      "tco2e": 291.42,
      "tco2e_basis": "location-based",
      "market_tco2e": null,
      "who": "developer model card",
      "scope": "Pre-training. The card states the emissions were fully offset.",
      "url": "https://github.com/meta-llama/llama/blob/main/MODEL_CARD.md",
      "reviewed": "2026-10-23",
      "source_id": "environmental-disclosure-index"
    },
    {
      "model": "DeepSeek-V3",
      "developer": "DeepSeek",
      "year": 2024,
      "epoch_model": "DeepSeek-V3",
      "hardware": "NVIDIA H800",
      "gpu_hours": 2788000,
      "device_w": null,
      "energy_mwh": null,
      "energy_derived": false,
      "tco2e": null,
      "tco2e_basis": null,
      "market_tco2e": null,
      "who": "developer technical report",
      "scope": "GPU hours only. No power, energy, location or emissions figure.",
      "url": "https://arxiv.org/abs/2412.19437",
      "reviewed": "2026-10-23",
      "source_id": "environmental-disclosure-index"
    },
    {
      "model": "Llama 3 70B",
      "developer": "Meta",
      "year": 2024,
      "epoch_model": "Llama 3-70B",
      "hardware": "NVIDIA H100 80GB",
      "gpu_hours": 6400000,
      "device_w": 700,
      "energy_mwh": 4480,
      "energy_derived": true,
      "tco2e": 1900,
      "tco2e_basis": "location-based",
      "market_tco2e": null,
      "who": "developer model card",
      "scope": "Pre-training. The card states the emissions were fully offset.",
      "url": "https://github.com/meta-llama/llama-models/blob/main/models/llama3/MODEL_CARD.md",
      "reviewed": "2026-10-23",
      "source_id": "environmental-disclosure-index"
    },
    {
      "model": "Llama 3.1 405B",
      "developer": "Meta",
      "year": 2024,
      "epoch_model": "Llama 3.1-405B",
      "hardware": "NVIDIA H100 80GB",
      "gpu_hours": 30840000,
      "device_w": 700,
      "energy_mwh": 21588,
      "energy_derived": true,
      "tco2e": 8930,
      "tco2e_basis": "location-based",
      "market_tco2e": 0,
      "who": "developer model card",
      "scope": "Training. The largest training run with a published emissions figure.",
      "url": "https://github.com/meta-llama/llama-models/blob/main/models/llama3_1/MODEL_CARD.md",
      "reviewed": "2026-10-23",
      "source_id": "environmental-disclosure-index"
    },
    {
      "model": "Mistral Large 2",
      "developer": "Mistral AI",
      "year": 2024,
      "epoch_model": null,
      "hardware": null,
      "gpu_hours": null,
      "device_w": null,
      "energy_mwh": null,
      "energy_derived": false,
      "tco2e": 20400,
      "tco2e_basis": "lifecycle",
      "market_tco2e": null,
      "who": "developer lifecycle analysis",
      "scope": "Training and 18 months of use to January 2025 together, on a lifecycle basis, so not comparable with a training-only figure.",
      "url": "https://mistral.ai/news/our-contribution-to-a-global-environmental-standard-for-ai",
      "reviewed": "2026-10-23",
      "source_id": "environmental-disclosure-index"
    },
    {
      "model": "Llama 4 Maverick",
      "developer": "Meta",
      "year": 2025,
      "epoch_model": "Llama 4 Maverick",
      "hardware": "NVIDIA H100 80GB",
      "gpu_hours": 2380000,
      "device_w": 700,
      "energy_mwh": 1666,
      "energy_derived": true,
      "tco2e": 645,
      "tco2e_basis": "location-based",
      "market_tco2e": 0,
      "who": "developer model card",
      "scope": "Pre-training.",
      "url": "https://github.com/meta-llama/llama-models/blob/main/models/llama4/MODEL_CARD.md",
      "reviewed": "2026-10-23",
      "source_id": "environmental-disclosure-index"
    },
    {
      "model": "Llama 4 Scout",
      "developer": "Meta",
      "year": 2025,
      "epoch_model": "Llama 4 Scout",
      "hardware": "NVIDIA H100 80GB",
      "gpu_hours": 5000000,
      "device_w": 700,
      "energy_mwh": 3500,
      "energy_derived": true,
      "tco2e": 1354,
      "tco2e_basis": "location-based",
      "market_tco2e": 0,
      "who": "developer model card",
      "scope": "Pre-training. No figure is published for Llama 4 Behemoth, the largest Llama 4 run.",
      "url": "https://github.com/meta-llama/llama-models/blob/main/models/llama4/MODEL_CARD.md",
      "reviewed": "2026-10-23",
      "source_id": "environmental-disclosure-index"
    }
  ]
}