SOMA AI排出係数（SAEF）—企業GHGインベントリ向けAI推論の排出・資源係数、データセットv3.1

SOMA AI Emission Factors (SAEF) – AI Inference Emission and Resource Factors for Corporate GHG Inventories, dataset v3.1 (原題)

Llopis-Garcia, Guillermo

Zenodoデータセット2026-11-12#AI×ESGOrigin: EU経営インパクト: 調達リスク対象セクター: cross_sector

DOI: 10.5281/zenodo.24160409

原典: https://zenodo.org/records/24160409

🤖 gxceed AI 要約

日本語

商用・自社ホストの大規模言語モデルについて、100万トークンあたりのCO2eと水使用係数をモデルクラス・地域別に提供するデータセット。電力・ハードウェア・建設・学習の4項ライフサイクル境界を採用し、210カ国・27 eGRID区分・143クラウドリージョンをカバー。v3.1ではケベック州の電力係数修正など誤り訂正のみを行った。

English

SAEF v3.1 provides kg CO2e and water factors per million tokens for commercial and self-hosted LLMs by model class and region. It uses a four-term lifecycle boundary (electricity, embodied hardware, construction, training) across 210 countries, 27 US eGRID subregions and 143 cloud regions, with Monte Carlo uncertainty propagation. v3.1 is a corrections-only release fixing regional grid factors.

Unofficial AI-generated summary based on the public title and abstract. Not an official translation.

📝 gxceed 編集解説 — Why this matters

日本のGX文脈において

Scope 3カテゴリ1・上流排出としてAI利用を算定する動きは日本企業でも実務課題化しつつあり、SSBJ開示やサプライヤーへの排出データ要求に対応する基礎資料となる。トークン単位の係数は自社AIガバナンスと統合報告書の算定根拠に直結する。

In the global GX context

As AI adoption scales, purchased inference becomes a material Scope 3 line item that current GHG Protocol guidance does not address. This dataset operationalizes token-based accounting, offering a reproducible factor set that complements ISSB/CSRD disclosure of upstream value-chain emissions and gives disclosure teams a defensible methodology for AI-related emissions.

👥 読者別の含意

🔬研究者:AI推論の排出をトークン単位で定量化する手法と係数セットを、再現可能な形で検証・拡張できる。

🏢実務担当者:自社の生成AI利用をScope 3上流排出として算定する際の係数と計算ツールとして直接利用できる。

🏛政策担当者:GHGプロトコル等の算定ガイドラインにAI推論という新たな排出源をどう位置づけるかの議論材料になる。

📄 Abstract（原文）

SOMA AI Emission Factors (SAEF), version 3.1.0, is the machine-readable companion to Llopis-Garcia, G. (2026), Accounting for Purchased Artificial Intelligence Inference in Corporate Greenhouse Gas Inventories: A Token-Based Method (arXiv:2606.10660, accepted for publication in Sustainable Production and Consumption). It gives estimated kg CO2e and water factors per million tokens for commercial and self-hosted large language models, by model class and region, with a standard-library Python calculator, end-to-end checks and a data dictionary. Released under CC BY 4.0. Changes from v3.0.0 (10.5281/zenodo.23371447). Corrections only; CHANGELOG.md gives before and after values and RELEASE_NOTES_v3.1.0.md the summary. (1) Azure canadaeast and Google Cloud northamerica-northeast1, which the providers place in Quebec, now take the Quebec provincial grid (0.0265 kg CO2e/kWh) instead of the Canadian national average (0.1744), as AWS ca-central-1 already did; their class B factor per million billed tokens falls from 0.0168 to 0.0063 kg. (2) The Central African Republic grid row, which Ember publishes as 0 g/kWh because its emissions round to 0.00 MtCO2, now takes a documented zero-intensity fallback of 0.0238 kg/kWh (100% hydro). (3) The README quick-start example now states the value the calculator returns, 0.0871 kg. No other row changes, and no number printed in arXiv:2606.10660 changes. Changes in v3.0.0 from v2. (1) The boundary is a four-term lifecycle: electricity, embodied hardware, construction and training, each reported separately. (2) Factors are on an effective-token basis, with the prefix-cache share restated per benchmark run to a reference mix. (3) The three model classes are the means of their measured members (18 open-weight models on H100 and B200, ML.ENERGY Leaderboard v3). (4) Coverage is 210 countries and economies, 27 US eGRID subregions and 143 public cloud regions; water, with AWARE 2.0 scarcity, is given for 17 regions. (5) Monte Carlo propagation of the input ranges is added. (6) Region-range and grid-year-2024 sensitivities are added. The raw third-party corpora (Azure traces, WildChat and Arena samples, OpenRouter rankings, Economic Index extracts) are excluded because they carry their own licences; they can be re-downloaded as described in THIRD_PARTY_DATA.md, which gives source, licence, attribution and script for each. The derived statistics are included. The folder v4_inputs/ is uploaded as v4_inputs.zip. See MANIFEST.md for every file. Licence CC BY 4.0.

🔗 Provenance — このレコードを発見したソース

Zenodo https://zenodo.org/records/24160409first seen 2026-11-13 04:15:56 · last seen 2026-11-14 04:13:09

🔔 こうした論文の新着を逃したくない方はキーワードアラートに登録（無料・3キーワードまで）。

gxceed は公開メタデータに基づく研究支援データセットです。要約・翻訳・解説は AI 支援で生成されています。最終的な解釈・検証は利用者が原典資料に基づいて行うことを前提とします。