Per ADEME/Arcep (2022 data, published January 2025), digital technology accounts for 4.4% of France's carbon footprint. Within that total, device manufacturing accounts for roughly 50% — versus 46% for data centers (usage). The Shift Project's thesis ("Lean ICT", 2018) therefore still held in 2022: it's hardware manufacturing, not computation's electricity, that dominates digital's impact.
This is not a stable gap. The data-center share nearly tripled between 2020 and 2022, before generative AI's data-center boom even took off. This study does not claim "manufacturing dominates forever" — it's true today, and the trend is reversing fast.
Sources: ADEME/Arcep, notre-environnement.gouv.fr (01/2025); The Shift Project, "Lean ICT" (2018).
An EQΦS score (Λ, φ(S)) is a weighted or geometric mean over 9 to 144 parameters — a few thousand floating-point operations. An equivalent analysis produced by an LLM is a language-model inference: an energetically different object, by several orders of magnitude.
9 to 144 parameters (dimension grid, coupling matrices) × ~20 FLOP/parameter (a wide margin: multiplication, log, exp, pow, summation) = 180 to 2,880 FLOP. At 1 nanojoule/FLOP — a deliberately pessimistic assumption, an order of magnitude above the picojoule regime measured on dedicated accelerators (Koomey's law) — the calculation's energy sits between 5×10⁻¹¹ and 8×10⁻¹⁰ Wh.
Sourced range for a short prompt: 0.24 Wh (Google Gemini, official disclosure, August 2025 median) to 2.9 Wh (EPRI). Intermediate academic figure: 0.42 Wh (Jegham et al. 2025, arXiv:2505.09598, infrastructure-aware benchmark across 30 models).
| Comparison | Gap |
|---|---|
| LLM low bound (0.24 Wh) vs. most expensive EQΦS calc (144 param.) | ×3×10⁸ |
| LLM median (0.42 Wh, academic) vs. most expensive EQΦS calc | ×5×10⁸ |
| LLM high bound (2.9 Wh, EPRI) vs. lightest EQΦS calc (9 param.) | ×6×10¹⁰ |
Two other facts, often absent from AI-vs-classical-computation comparisons: EQΦS trains no model (training GPT-3 alone is estimated at ~1,287 MWh, Patterson et al. 2021, arXiv:2104.10350 — a cost entirely absent from our approach), and the engine currently runs on the two founders' own laptops, not on dedicated server infrastructure provisioned for the occasion.
Our own internal methodology (the "Digital Responsibility & Green IT" library) forbids displaying "green" without proving it. Here's what stays true only with nuance, without discarding it.
The verdict calculation is pure arithmetic — not the research, drafting, and formatting of our studies, which are AI-assisted (systematic disclosure on our deliverables). The sobriety claim applies to the calculation engine, not the entire production process.
True in 2022 at the macro level (ADEME: 50% manufacturing vs. 46% data centers), but the data-center share nearly tripled in two years, driven by AI. This is not a fixed fact.
Making a structural analysis 500× cheaper (US$390 vs. €200,000 at a consulting firm) can mechanically increase the total volume of audits performed — the Jevons paradox applied to AI is documented (arXiv:2501.16548). Each individual calculation still costs orders of magnitude less; total impact depends on volume, which we don't claim is negligible without real data.
The site and questionnaires are hosted in France/EU (Amen) — compliant with PDPA and digital sovereignty, but with no identified environmental certification to date. We don't include it in the sobriety claim.
Google (official Gemini disclosure, 08/2025) · Jegham et al. 2025, "How Hungry is AI?", arXiv:2505.09598 · EPRI, "Powering Intelligence" (2024/2025)
Patterson et al. 2021, "Carbon Emissions and Large Neural Network Training", arXiv:2104.10350 · Meta, Llama 3/3.1 model cards
Koomey's law (Koomey et al., revisited 2024, Cluster Computing)
The Shift Project, "Lean ICT" (2018) · ADEME/Arcep, 2022 data (published 01/2025, notre-environnement.gouv.fr)
"From Efficiency Gains to Rebound Effects: The Problem of Jevons' Paradox in AI's Polarized Environmental Debate", ACM FAccT 2025, arXiv:2501.16548
The Shift Project · GreenIT.fr · ADEME — internal EQΦS methodological framework
Per-request LLM energy remains a disputed topic (a 10x uncertainty acknowledged by the research community itself) — we cited sourced ranges, not a single figure chosen to fit the demonstration.
You just read the calculation, not a promise.
Computed on your data, your case, your framing. Not an opinion — a reproducible mathematical verdict.