Your photos1,000 pages

Reading by the AI44.8 gonce

Translation by the AI20.8 gonce

Processing on our server1 gonce

Storage, 3 copies5.1 gevery year

Over five years92 g

On average per year18 g

Reading and translation by the AI, like processing on our server, happen only once, in the first year. After that, only keeping the photos remains: over five years, 92 g, or 18 g per year on average.

Not counted:your computeryour camerayour connection

A year of wondarchive

1,000 archive pages

18 gof CO₂e per year, on average over five years

170 mby car

11 minof a 2,000 W electric oven, switched on once in the year

35 hof a 10 W LED bulb

The calculation for 1,000 pages

1,000 pages read, translated into one language and kept for five years

The 1st year

72 g

to read, translate and keep the pages

Each following year

5.1 g

to keep the pages

Per year, on average

18 g

170 m by car or 11 min of electric oven

Over the whole period

92 g

the first year, then four years of keeping

The same work done differently, per year on average · 1,000 pages

Approach	CO₂e per year	What is counted
wondarchive	18 g	Reading, translation, server and storage.
The same work with an AI hosted in the United States	50 g	Reading, translation, server and storage, with American electricity.

The detail for wondarchive, over the whole period · 1,000 pages

Item	CO₂e
Reading by the AI	44.8 g
Translation	20.8 g
Server	1 g
Photo storage	25.5 g
Over the whole period	92.1 g

Estimate, not measurement: each factor and its source are given below. Every figure is for 1,000 pages; the average per year divides the five-year total by five, first year included.

What we do, and what you can check

once

A page is read only once

Each photo carries a fingerprint: sent twice, it is neither uploaded nor read again. A page already read is never re-read automatically; a new reading is something you ask for, and anything corrected by hand is spared by default.

light

A light model first

Printed and typewritten pages are read by a light AI model. A heavier model re-reads only handwriting and doubtful pages: 12% of pages.

useful

No needless call

A page already written in your language, or without text, is not translated.

image

A light reading image

The reading view shows a 2,400-pixel WebP copy, the gallery 320-pixel thumbnails. The original, at full definition, is transferred only in full screen and on export. Images stay for a year in your browser’s cache.

cache

Calculate once, read afterwards

Dates, names, places and indexes are calculated once, at processing. A search keeps its result for a year: running it again re-reads only the pages changed since.

site

A site without extras

Static pages, no framework, no advertising tracker or third-party content delivery network; the fonts are hosted by us.

What still costs, and why we keep it

The original at full resolution is sent to the model. We measured that a reduced image degrades the reading of double pages: we kept the quality.
Handwriting goes through a heavier model. It is the one that reads it correctly; it is used only for those pages.
Your originals are kept in three copies, in three data centres in Paris. A loss would be permanent: we triple the storage rather than risk it.

And on your side

Sort before importing: bursts, blurred photos and duplicates cost the same as the others.
One offline copy rather than four online. Keep a backup of your own, on a drive, and avoid syncing the same photos to several clouds.
Do not rerun a reading without reason: correcting a line by hand costs less than a re-reading.

Never delete your only copy: our originals are not versioned, a deleted page is deleted permanently.

The method

The AI’s CO₂e is calculated by multiplying two lines: the energy of one reading (Wh) by the CO₂e per kWh of the electricity that powers the AI. They are not two items that add up.

Item	Value used	Source
Energy of one AI reading or translation (Wh)	0.25 Wh per reading, 0.13 Wh per translation (0.1 to 0.4 Wh)	Measured on our side: ≈ 3,700 tokens read and 1,600 written for a reading, 1,800 and 870 for a translation. Energy of a small model per token, increased by half for the data centre; same order as the estimates by Epoch AI (2025) and as the measurements published by AI providers (≈ 0.24 Wh for a median query).
CO₂e per kWh of the electricity that powers the AI	160 g CO₂e/kWh	Electricity of the provider’s European data centres.
Our server	≈ 1 mg per page processed; ≈ 15 kg CO₂e per year of operation, whatever the volume	Measured on our server: image preparation and reading orientation, 9 seconds of processor time per page. Shared virtual server at OVHcloud, PUE 1.26 (OVHcloud, 2025), French electricity at 52 g/kWh (ADEME). The fixed cost (share of a machine and manufacturing) is not allocated to the pages.
Our storage	1.5 kg CO₂e per TB per year (three copies)	Data-centre drives (Seagate Exos X20, ISO 14040/14044), French electricity.
Electricity in the United States	370 g CO₂e/kWh	National average of the American grid.
Cloud in the United States	5.4 kg CO₂e per TB per year	Same data-centre drives, American electricity, three copies.
Car	108 g CO₂/km	2026 malus threshold.
Electric oven	104 g CO₂e per hour	2 kWh (2,000 W) at 52 g/kWh.
10 W LED bulb	0.52 g CO₂e per hour	10 Wh at 52 g/kWh.

Measured on our side over thirty days: 30,403 pages, 2.85 MB per original. We offset nothing and we do not call the service “neutral”: it emits, here is how much.

See it on your own pages

A hundred pages transcribed and translated, for a year, no card required.

Create my free account See pricing

How much do 1,000 archive pages on wondarchive emit