「では、AIを使う炭素コストは？」限定的開示下で大規模言語モデル推論のトークン当たり炭素強度を算出する有界推定フレームワーク：OpenAI GPT-5モデルファミリーを事例として

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

🔬研究者:限定的開示下でAI推論の炭素強度を推定する方法論と不確実性の扱いを学べる。

🏢実務担当者:生成AI利用の炭素コストを把握し、モデル選択・努力設定による削減策を検討できる。

🏛政策担当者:AI事業者への推論エネルギー・炭素開示義務の必要性を検討する根拠となる。

So, what is the carbon cost of using AI? No leading provider tells us: OpenAI, Anthropic and others publish no per-model, per-query energy or carbon figures. This means that nothing is disclosed about the operational carbon footprint of inference at the one unit where a user decides – a particular prompt, model and effort setting. We present a bounded estimation framework that answers the question from public and collected data alone. It triangulates an anchor (a published energy figure for one model) across three classes of public source, scales between models by observed throughput, applies query-length-conditional reasoning multipliers measured by our token-level campaign, corrects for the rise in per-GPU power between hardware generations, reports location-based carbon (what the local grid emitted) and market-based carbon (net of the provider's clean-energy contracts) in parallel under the GHG Protocol Scope 2 Guidance, and compounds residual uncertainty into a single multiplicative envelope (a range around each figure). Applied to OpenAI's GPT-5.x family, the framework estimates a fleet intensity of 353 gCO₂e/kWh location-based (319 – 500 across routing scenarios) and 52 gCO₂e/kWh market-based (47 – 65, rising to 159 in a particular scenario.) Per-model central estimates – per 1,000 tokens, about 750 words – span roughly 14×, from 1.07 gCO₂e/1,000 tokens location-based (0.16 market-based) for GPT-5.4-nano with reasoning off to 14.80 (2.16) for GPT-5 at high effort – within envelopes of roughly 10 – 19× top-to-bottom, median ~11×. All headline figures carry Low confidence (IPCC AR5 – the evidence is limited): a single line of audited per-model disclosure would collapse most of the envelope, and we invite it. What should a user do? A general user need not worry – twenty short prompts a day for a year is roughly 4 kg CO₂e, about 0.04% of the UK's average per-person footprint – however framing the task clearly and asking for shorter answers costs little. A power user, at about 122 kg a year (a short-haul plane flight), should choose the smallest capable model and the lowest effort setting the task allows – on GPT-5 to 5.4 the effort setting is the larger lever, from GPT-5.5 onward model choice is. An application designer holds the largest lever: at a million queries a month, routing from a flagship model at high effort to a mini model at medium could save on the order of 38 tCO₂e a year, the equivalent of just under four business-class return flights from Dallas to Delhi.

• Zenodo https://zenodo.org/records/23084616first seen 2026-10-02 04:14:17