Methodology

Spine's carbon and water footprint

Version: 1.0
Date: October 28, 2026
Tools and data: EcoLogits 0.11.3; Scaleway's Environmental Footprint API; ADEME's Impact CO₂, retrieved on October 28, 2026

In brief

Spine shows each customer company the footprint of its usage over its billing cycle, in CO₂ equivalent (CO₂e) and in water.

Spine has been hosted by Scaleway since September 2026, in part because Scaleway is particularly transparent about the environmental footprint of its services: it publishes that footprint for each project, product by product.

Consumption figures come from the providers: Scaleway for hosting, the AI model APIs for each call (tokens, pages).
Conversion to CO₂e and water comes from Scaleway for hosting and from EcoLogits for AI models.
The equivalents (km by car, showers…) come from ADEME, the French Agency for Ecological Transition; Cieau, the French water information center; and France's National Consumer Institute.

The figures are orders of magnitude and should be read as such.

1. What we estimate

The unit: providing Spine to a company for one billing cycle. A user's individual share is derived from it (§ 5).

Two indicators:

greenhouse gas (GHG) emissions, in kg of CO₂e;
water consumed, in liters.

Two shares:

the fixed share, hosting: the machine, its disks, and the company's file storage;
the variable share, usage: each call to an AI model.

2. Where the consumption figures come from

Item	What we receive	From whom
Hosting	CO₂e per day for a machine of each type, its disk, and file storage	Scaleway, through its Environmental Footprint API, queried once per version of the method
Conversations, automations, background tasks	The model called and its output tokens, including reasoning	The model provider's API, for each call
Scanned page reading (OCR)	The number of pages read	The reading service, for each reading
Image generation	The output tokens	The image service, for each image
Web search	The number of searches	The search service, for each search

A company's hosting is derived from its machine type (§ 4).

3. How we convert them into CO₂e and water

Hosting: Scaleway

Scaleway calculates the footprint of its resources itself, following ADEME's methodological framework for data centers and the cloud. Its calculation includes:

the manufacturing of the servers and the data center;
the electricity consumed, based on the national average mix, with no deduction for renewable electricity contracts;
cooling.

Spine uses these figures as is, for each machine type.

AI models: EcoLogits

EcoLogits is an open-source library (MPL-2.0 license), published by the nonprofit CodeCarbon and cited by ADEME as an order-of-magnitude tool. For a call, it starts from the model and the number of output tokens, then estimates:

the computation time, from the model's known throughput;
the energy consumed by the GPUs and the server, based on the size of the model;
the emissions and water associated with that energy, based on the electricity mix of the data center's country;
the share of hardware manufacturing, pro rata to the time of use.

The result is a range, not a single figure. Spine uses its midpoint.

Cases EcoLogits does not cover

They are mapped to the closest match in EcoLogits.

Case	What we use
Model not in EcoLogits	The closest version in the same model family, from the same provider
OCR	750 output tokens per page (one page of text), calculated on a similar model, from the OCR provider
Images	The actual output tokens, calculated on a similar model, from the image provider
Web search	ADEME's value for one web search, per search; no water
Models served by the hosting provider	Their equivalent in EcoLogits, otherwise a similar model, based on the electricity mix of the hosting country

A run that calls several models is calculated model by model.

4. The calculation

Company footprint over the cycle = hosting + sum of calls

hosting = daily machine footprint × days in the cycle
machine = instance of its type + disk + file storage
call    = EcoLogits(provider, model, output tokens)

A machine's footprint is the same every month. Scaleway calculated it for each instance type, for the disk, and for file storage, and Spine keeps that value until the next version of the method. The instance type follows the company's machine tier; a company that shares its machine is allocated a portion of it. Instances set up before September 2026, on a shared machine at Hetzner, in Germany, are calculated with these same Scaleway values.

During a cycle, Spine projects the end of the cycle: the machine's daily footprint multiplied by the length of the cycle, plus the calls already made, scaled to the length of the cycle. A call's footprint is calculated when the call ends. The projection starts from the date usage measurement went live, or from the start of the cycle if that is later: a cycle that ended before usage measurement went live has no figure. Calls made between October 15, 2026, and the date usage measurement went live are estimated from their credits, using the model served by default at the time.

user's share = company footprint × user's credits ÷ company's credits

Credits follow the actual cost of the models. That cost grows with their size, and therefore with their footprint: a call to a large model weighs more, in credits and in CO₂e alike. Every credit attached to a user is counted toward that user, including credits used by an automation or a background task. The shares therefore add up to the company's total.

6. What is counted, what is not

Counted

Manufacturing and use of the servers that host the company
Computation of AI model answers, hardware manufacturing included
Water used by data centers and electricity generation, for AI models

Not counted

Embeddings (indexing for search): neither Scaleway nor EcoLogits quantifies them, and their weight is small
Model training
Input tokens and the cache: EcoLogits does not count them
The network, and users' computers or phones
Water for hosting, which is negligible
The vendor's internal tools and its own operations

7. Uncertainty and rounding

The figures are orders of magnitude. Depending on the model, the top of the EcoLogits range is up to seven times the bottom, and the size of closed models is estimated.

The display takes this into account:

the unit is chosen so that the number is at least 1 (g, kg; mL, L);
below 1 g, it reads “< 1 g” (“< 1 mL” for water); nothing at all reads “0 g”;
the number is rounded to a whole number, with at most two significant figures: 3.5 kg is shown as “4 kg,” 123 kg as “120 kg”;
never any decimals;
the figure for a cycle in progress, which is projected, is preceded by “≈”; the figure for a cycle that has ended, and whose usage was measured from its first day, is shown without it.

8. What we compare it to

Each figure comes with the four comparisons on its list. The number for each one follows the rounding in § 7; below 1, it becomes a fraction, 1/2, 1/3… down to 1/100, then “less than 1/100.”

CO₂e, according to ADEME's Impact CO₂, over the full life cycle: 1 km in a gas or diesel car, 1 km on a short-haul flight, one day of gas heating for a 100 m² home, one meal with beef. Transportation is counted per person.

Water, according to Cieau and France's National Consumer Institute: one toilet flush, one minute with the faucet running, one 5-minute shower, one car wash with a hose.

The equivalents illustrate a figure; they are not a result.

9. Using these figures in carbon accounting

For the customer company, Spine's footprint falls under Scope 3, Category 1 (purchased goods and services) of the GHG Protocol. It replaces the spend-based emission factor for IT services applied to the Spine invoice; it is not added to it.

The figure is to be used with its boundary (§ 6): excluding model training, the network, and users' devices.

10. Versions

Each version of Spine sets the version of this method, the version of EcoLogits, and the date the equivalents were retrieved; a change to any of the three produces a new version of this document, stating what changed. The footprint of a call keeps the measurement that produced it (the model and the output tokens), so it can be recalculated with a later version of the method. Hosting for a past month, however, follows the table of the version currently in use.

Framework and sources

The method draws on three texts:

ADEME's methodological framework for data center hosting and cloud services (2025): national electricity mix, manufacturing included;
ADEME's opinion on generative AI (July 2026): several indicators, a per-service and per-user approach;
the Software Carbon Intensity (SCI) standard (ISO/IEC 21031:2024): the footprint of software is the sum of the emissions from the energy consumed and from hardware manufacturing, expressed per unit.

Data and tools: