"AI is destroying the planet." That's what you read everywhere — often with 2023 figures applied to 2026 models.
The studies in circulation contradict each other because they don't measure the same thing: not the same models, not the same years, not the same scope (electricity only, or full lifecycle). So we redid the math, for real usage, with dated, verifiable sources.
The starting point: 0.34 Wh per query
In June 2025, Sam Altman published OpenAI's first official figure: an average ChatGPT query consumes about 0.34 Wh of electricity and 0.000085 gallons of water (one-fifteenth of a teaspoon).
Intellectual honesty requires two caveats:
- It is not audited. OpenAI published no methodology, and independent researchers treat it as a plausible floor: it covers inference, not training or hardware manufacturing.
- It applies to an "average" query. Long reasoning queries consume significantly more. Academic estimates put recent reasoning-model responses at 18 Wh and above.
That said, the order of magnitude is corroborated by independent work (Epoch AI estimated ~0.3 Wh for GPT-4o), and it marks a major shift: the 2023 figures — "one ChatGPT query = 10 Google searches" — no longer describe 2026 reality. Models have become smaller and more efficient; the alarmist figures were simply never updated.
The math for real professional usage
Take a heavy user: 40 queries a day, every day, for a year. Roughly one hour of daily usage.
40 × 365 × 0.34 Wh ≈ 5 kWh per year.
To put 5 kWh in perspective:
- ≈ 33 km in an electric car,
- ≈ 5 to 6 laundry loads,
- ≈ 4 hours of oven use,
- less than a month of an internet router left on.
The variable that changes everything: where the server runs
5 kWh of electricity does not emit the same CO₂ everywhere. This is where the debate gets interesting — and political.
According to RTE's 2025 electricity report, the carbon intensity of French electricity production is 19.6 gCO₂e/kWh — among the lowest in Europe, thanks to nuclear and renewables (95% low-carbon production). The European average is 178 g. Germany is estimated at around 330 g, Poland above 500 g.
Our 5 annual kWh therefore become:
| Datacenter country | Carbon intensity | Annual emissions |
|---|---|---|
| France | 19.6 g/kWh | ≈ 100 g CO₂e |
| EU average | 178 g/kWh | ≈ 0.9 kg CO₂e |
| Germany | ~330 g/kWh | ≈ 1.7 kg CO₂e |
| Poland | ~500 g/kWh | ≈ 2.5 kg CO₂e |
A cheeseburger is roughly 2.5 kg CO₂e according to common estimates. In other words: a year of intensive AI on a server in France emits less than a single cheeseburger. The same usage on a coal-heavy grid emits the equivalent of a whole one.
The limits of the calculation (because they exist)
Embodied carbon in hardware. Our 100 g only covers inference electricity. NVIDIA published the manufacturing carbon footprint of an 8-GPU HGX H100 chassis: about 1.3 tonnes CO₂e. That cost is real, but amortized over billions of queries across several years — marginal per use, never zero. Including the full lifecycle of the electricity system, RTE puts French intensity at 29 g/kWh rather than 19.6: our 100 g becomes ~145 g. The order of magnitude holds.
Model training. It is energy-intensive, but also shared across all usage of the model. And using already-trained open-source models — which is what we do — adds no additional training.
Water. The problem is real, but localized. In the United States, datacenters are documented in conflict with local water resources in several states. In Europe, recent datacenters compliant with regulations mostly run closed-loop cooling. Once again: the question is not "does AI consume?", it's "where and how does the server run?".
What this means for a company
The environmental argument converges with the sovereignty argument: infrastructure location is the decision that matters most. The same AI usage can emit 25 times more CO₂ depending on the datacenter's country — exactly as it can fall under the Cloud Act or not depending on the operator.
At Eridia, models run on your infrastructure or on servers in France: the most decarbonized electricity mix among large European countries, and your data never leaves your perimeter. Both arguments hold in the same architecture.
If your CSR policy now covers AI — increasingly the case in RFPs — we can help you document the real footprint of your usage, with numbers and sources.