Energy & carbon, measured honestly.

Our own tool

Measured by Argand Energy

Most companies that publish a carbon figure either guess, or paste their traffic into someone else's calculator. We built our own. Argand Energy is a small in-house tool (written in Rust, like the rest of Argand) that produces every number on this page — and it's built to be honest by construction, not by promise:

It runs continuously (recording the grid) and on demand (measuring a search). It isn't public yet — Argand is pre-launch — but we'll open it up alongside the search engine so anyone can check our maths.

The honest surprise

It's barely the servers

Argand runs in a Finnish data centre on a very clean, hydro-powered grid. So the electricity to actually answer your search is a tiny sliver — well under a thousandth of a gram of CO₂.

Almost all of the footprint is two things added on top: the weight of the web page we send to your device, and the one-time carbon of manufacturing the server, spread across every search it will ever serve. So the honest way to make Argand greener is to keep pages tiny and hardware long-lived — not to obsess over server wattage. (Keeping pages tiny is the whole point of "runs on a potato".)

Scope-2, both ways

Why we publish two numbers

There are two honest ways to count the carbon of electricity, and blending them would flatter the result — so we report both, separately, and never net them against each other:

It's a range, not a single bragging number, because two of the pieces are shared costs spread over how many searches the hardware serves in its lifetime — and pre-launch, that total is genuinely unknown. The more Argand is used, the smaller each search's share becomes.

Breakdown

Where the carbon comes from

Per search, assuming the hardware serves a billion searches in its life:

Per-search carbon by source: Answering your search (server electricity) 12%, The page we send to your device 59%, Making the server hardware (shared across every search) 29%, Building the search index (shared, one-time) <1%.

• Answering your search (server electricity) Measured 12%
• The page we send to your device Modeled 59%
• Making the server hardware (shared across every search) Fetched 29%
• Building the search index (shared, one-time) Measured <1%

The "making the hardware" figure is a deliberately cautious stand-in: we use a big dual-socket reference server (1,200 kg to manufacture, looked up live from the Boavizta database) as a stand-in for our smaller single box, so we over-state rather than under-state it.

How the shared costs shrink with use

The hardware's manufacturing carbon is split across every search it ever serves, so the per-search share falls the more Argand is used. We quote the headline at the 1 billion-search point.

In context

How that compares

* independent estimate; the company publishes energy, not a carbon method. The real point: Argand publishes a complete, measured, current number. Google's search figure is 15 years old, the AI assistants publish serving energy only, and some publish nothing. We also deliberately keep large language models out of the search path — which is exactly why a search costs a fraction of an AI answer.

Intuition

What these numbers actually feel like

A gram of CO₂ is hard to picture, so here's the same carbon as everyday things. These are rough third-party estimates — for intuition, not precision.

So one kilometre of driving is roughly forty thousand Argand searches, and a banana is about twenty thousand. Digital and food figures are approximate, from Mike Berners-Lee's How Bad Are Bananas?; driving via the US EPA; search and AI figures as cited in Sources.

Method

What "measured", "modeled" and "fetched" mean

Every number in our own breakdown is labelled with how we got it. There are three honesty levels — strongest first:

• Measured Read straight off the hardware's own energy chips while real work runs (Intel/AMD RAPL for the processor, NVML for the graphics card). This is the strongest kind.
• Modeled A peer-reviewed formula applied to a real measurement. The page-carbon figure uses CO2.js (the Sustainable Web Design model) over our actual measured 20 KB page size.
• Fetched Looked up live from a public, reproducible database. The hardware-manufacturing carbon comes straight from the Boavizta API.

Right now the weakest link is the page-carbon model (it's an estimate, not a direct measurement) — and Argand Energy says so plainly rather than hide it behind the stronger numbers.

No fine print

Our honesty rules

• If we can't measure something, it's shown as "pending" — never as a fake zero.
• Location-based and market-based carbon are reported separately and never blended.
• Carbon offsets and green-energy certificates are reported separately and never subtracted from the figure.
• Shared costs (hardware, index build) are shown as a range over lifetime searches, never collapsed into one flattering number.
• Every figure says how we got it — measured, modeled, or fetched.
• Other companies' figures are shown with their source, date, and method — never restated as if they were ours or directly comparable.

The boundary

Limits & what's not included

The honest boundary of the number — what it leaves out, and where it's least certain:

• Web-crawl & indexing beyond embedding aren't in the per-search figure — the index-build energy we do count is document embedding only.
• Hardware end-of-life (recycling/disposal) isn't included — the Boavizta embedded figure is manufacturing only.
• Network transit beyond the modeled page bytes (backbone, your ISP) isn't counted.
• Embodied is a deliberately over-stating proxy — a big dual-socket reference server stands in for our smaller box, so we round up, never down.
• Lifetime search volume is the dominant unknown — the shared costs are a curve, and pre-launch the real total is a genuine guess.
• Competitor figures are not directly comparable — different products, boundaries, methods and grids (see "How that compares").

For the skeptic

Questions a skeptic should ask

Don't trust — check

Verify it yourself

The parts you can reproduce today, without us:

1. Hardware carbon — ask the public Boavizta API for the reference server:

   curl "https://api.boavizta.org/v1/server/?archetype=dellR740&verbose=false&criteria=gwp"

   → impacts.gwp.embedded.value = 1200 kgCO₂e.
2. Grid carbon — Finland's live emission factor is Fingrid dataset 396 (~20 g/kWh).
3. Page carbon — run the open-source CO2.js on a 20 KB green-hosted page → ≈ 0.00248 g.
4. Put it together with the SCI formula:

   SCI = (E·I + M) / R   — energy × grid intensity, plus embodied, per search

The live serving measurement and the full Argand Energy tool are reproducible by anyone once we open-source them at launch. The machine-readable figures are at /energy/data.json.

Receipts

Sources

Our measurement

• The standard: Software Carbon Intensity (Green Software Foundation), published as ISO/IEC 21031:2024; two-method Scope-2 accounting follows the GHG Protocol.
• Grid carbon (Finland): Fingrid open data, dataset 396.
• Renewable contract: Hetzner sustainability — its Finnish data centre runs on hydropower.
• Page carbon: CO2.js (The Green Web Foundation), using the Sustainable Web Design model.
• Hardware carbon: the Boavizta open hardware-impact API.

Comparison + everyday figures (third-party)

• Google Search (≈0.2 g, 2009): “Powering a Google search”, Google (2009).
• Google Gemini (≈0.03 g, 0.24 Wh, 2025): Measuring the environmental impact of AI inference, Google (2025) (paper).
• ChatGPT (≈0.34 Wh, 2025): Sam Altman, OpenAI (2025); context + carbon estimate via IEEE Spectrum and Hannah Ritchie.
• Anthropic / Bing (no per-query figure): disclosure comparison (2025–26).
• Everyday equivalents: Mike Berners-Lee, How Bad Are Bananas?; driving via the US EPA equivalencies calculator.