Here is what AI consumes in a company like yours, over one year. The calculation runs in your browser: no address to give, nothing is sent.
Typical company · 25 people
511 kWhper year
Electricity consumption of AI usage, over one year. Between 264 kWh and 990 kWh depending on the assumptions retained.
This figure is that of a typical SME, not yours yet. Correct the headcount and the reading follows.
Intensity · per employee
20.4 kWhper year
This intensity is what compares from one company to another, not the volume: volume mostly follows your headcount.
What this represents
186kettles
brought to the boil, per employee per year.
These settings describe a single query, to show where the gap comes from. The region and the type of data center also apply to your reading above; the four others describe usage and are set in the "My company" step.
The query
Cost paid once per conversation thanks to the cache, not on every message.
The hosting
x9.4
times the most frugal configuration
Factor breakdown
Model
x1
Length
x1
Reasoning
x1
Agentic
x1
PUE
x1.54
Region
CO2 only
x22
For identical energy, the region only moves the CO2 emitted to produce that energy. It changes neither the watts consumed nor the PUE.
Logarithmic scale. The vertical mark is x1: to the left the variable reduces, to the right it amplifies.
Energy
per query
0.4 Wh
(range 0.21 to 0.77)
CO2
0.17 g CO2
(range 0.09 to 0.33)
37.4 mg CO2
(range 19.3 to 72.4)
Market-based: after purchasing green certificates. Location-based: what the grid physically emits. At Google the gap is x3.7 (94 vs 345 g/kWh); at Microsoft, FY2025 scope 2 varies by a factor of 4.4 depending on the method (2.71 vs 12.03 Mt).
Water
0.46 mL
(range 0.24 to 0.89)
1.69 mL
(range 0.87 to 3.27)
Site only: what the data center evaporates. Full scope: plus the water used by power plants. Between the two official disclosures, Google (0.26 mL) and Mistral (45 mL, full LCA), there is a factor of 173, and neither is lying.
That is, for one query0.8 s of microwave12 s of television110 s of LED bulb