ecocost
Copyright 2026 Gooey.AI

This product is licensed under the Apache License, Version 2.0 (see LICENSE).

The knowledge base in ecocost/data/ contains figures taken or derived from the
third-party sources below. Their licences apply to those figures; the Apache
License covers only this project's own code and compilation. Every value names
its source in the YAML, and ecocost/data/sources.yaml records each source's
licence and terms.


Grid carbon intensity and generation mix (regions.yaml, countries)
------------------------------------------------------------------
Ember, Yearly Electricity Data, 2025.
  https://ember-energy.org/data/yearly-electricity-data/
  Licence: CC BY 4.0, https://creativecommons.org/licenses/by/4.0/
Our World in Data, "Carbon intensity of electricity generation" and "Share of
electricity production by source", based on Ember and the Energy Institute.
  https://ourworldindata.org/grapher/carbon-intensity-electricity
  https://ourworldindata.org/grapher/share-elec-by-source
  Licence: CC BY 4.0
Changes: values transcribed per country; intensity used as a lifecycle factor
with a tier-based range added.


Grid carbon intensity and generation mix (regions.yaml, US subregions)
----------------------------------------------------------------------
U.S. Environmental Protection Agency, eGRID2023 Summary Tables (rev. 2).
  https://www.epa.gov/egrid
  A work of the U.S. Government, not subject to copyright in the United States.
Changes: direct emission rates multiplied by 1.124 (the 2023 Ember/eGRID
national ratio) to put them on a lifecycle basis.


Water consumed in electricity generation (regions.yaml)
-------------------------------------------------------
Reig, P., Luo, T., Christensen, E. and Sinistore, J. 2020. "Guidance for
Calculating Water Use Embedded in Purchased Electricity." World Resources
Institute. Appendices 1 and 2.
  https://www.wri.org/research/guidance-calculating-water-use-embedded-purchased-electricity
  Licence: the report is CC BY 4.0. Its appendix tables are based on data from
  thinkstep AG (now Sphera), which the report makes available under separate,
  more restrictive terms. ecocost uses a small number of aggregate national and
  subregional factors (17 values) to support and explain its own calculated
  results.
Changes: converted from gal/kWh to L/kWh; ranges added.


Primary energy factors (regions.yaml); closed-model size estimates (models.yaml)
------------------------------------------------------------------------------
EcoLogits, GenAI Impact. electricity_mixes.json and models.json.
  https://github.com/genai-impact/ecologits
  Licence: Mozilla Public License 2.0, https://mozilla.org/MPL/2.0/
  Primary energy factors in electricity_mixes.json originate from ADEME Base
  Empreinte, https://base-empreinte.ademe.fr/
ecocost does not include EcoLogits source code or its data files; individual
values were transcribed, and for primary energy divided by 3.6 (MJ/kWh to
MJ/MJ).


Measured inference energy (models.yaml, hardware.yaml)
------------------------------------------------------
The ML.ENERGY Leaderboard, University of Michigan.
  https://ml.energy/leaderboard/
  https://github.com/ml-energy/leaderboard
Chung, J.-W. et al. "The ML.ENERGY Benchmark: Toward Automated Inference
Energy Measurement and Optimization." 2025. https://arxiv.org/abs/2505.06371
Changes: per-model Wh per 1k output tokens on H100, and paired H100/B200
runs used to derive chip energy ratios.


Hardware embodied carbon (hardware.yaml)
----------------------------------------
NVIDIA, HGX H100 Product Carbon Footprint Summary.
ADEME / Lees-Perasso et al., GPU life-cycle assessment.
Boavizta, BoaviztAPI server impact model, https://github.com/Boavizta/boaviztapi
EcoLogits, LLM inference methodology, https://ecologits.ai/
Figures cited, no code or data files included.


Other factual figures
---------------------
Model parameter counts come from each model's published model card. Chip
specifications come from vendor data sheets. PUE, WUE and site figures come
from the operators' own sustainability reports and documentation. Epoch AI
data is used under CC BY 4.0 (https://epoch.ai). Each value's source is linked
in the YAML.
