Methodology
How we calculate
The chain is the same for every tool. Only the first step differs — and that is exactly where it is decided how much a number is worth.
The chain
Four calculation steps, four uncertainties. The decisive difference sits right at the front: how precise is the usage signal?
- 01 · Usage signalmostly counted
tokens · interactions · requests · seconds
Comes from admin interfaces, exports, or by hand. The origin is stated for every tool.
- 02 · × energy factorfactor of 2.3
→ kWh
This is where the large uncertainty sits. No provider publishes what cached context costs — hence two limits instead of one number.
- 03 · × grid mixofficial
→ CO₂
By serving region. Sources Ember and IEA, the version is printed on every record.
- 04 · × water efficiencyofficial
→ litres
Cooling per data centre region. No range, because these factors are published.
Sources per step Jegham et al. · Epoch AI · IEA · Ember
The energy factor is the shaky part. No provider publishes how much electricity a single request actually draws. What we have is research and vendor statements — we cite both with source and date.
Two statements, not one
Every row carries two independent statements. They are often confused, and that confusion is why so many numbers in this field are worthless.
Quality
How reliable is the energy factor?
The three values appear on every row — including when the answer is uncomfortable.
Origin
How did the usage figure arrive?
Whoever enters a figure by hand is named underneath. Self-reported figures stay attributable.
A hand-typed credit count is estimated and manual. Those are two different caveats about the same row, and we show both.
Example · our own measurement
May 2025 to September 2026
200 – 457 kWh
384.2 kWh
model value · cache share 94.4 %
No provider publishes what cached context costs in energy. That is why we state both limits. One tidy number here would be invented.
Three ways in
Automatic
Through the providers official admin APIs. One approval by your administration is enough; nothing gets installed.
From a file
Where there is no API but there is an export: upload it, done.
By hand
Where there is neither. Whoever enters a figure is named beneath it — self-reporting stays attributable.
Where our model ends
- Cached requests
- A large share of modern AI requests reuses cached context. No provider publishes what such a lookup costs in electricity. We bound it from three directions: our own measurement (eight cold/warm pairs against a local model, September 2026) gives the lower bound — it measures saved compute time and therefore less than actually occurs. Systems literature and the providers’ own pricing both put the saving at roughly 90 per cent; that is our upper bound. We report both.
- Training
- We account for usage, not for training the models. The training share per request cannot be apportioned credibly from the outside.
- End devices
- The electricity drawn by the machine in front of the person belongs in the building account, not here. Counting it twice would be convenient and wrong.
What we do not collect
- No request content. Only usage signals — how much, when, which tool.
- No breakdown by person unless a company explicitly asks for it. The default is a daily total per tool.
Why this sounds laborious
Because in this field a clean number is almost always an invented one. We would rather give you a range that holds than a point value that does not. Anyone who has to stand behind it — in front of customers, auditors, or their own staff — needs the version that survives questions.