The full cost of a prompt: 1.13 Wh, screen included
A text prompt costs roughly 1.13 Wh of electricity once you count the reader's screen. Compute in the datacentre accounts for 0.3 Wh of that. The two minutes spent typing the question and reading the answer account for 0.83 Wh. Inference is therefore 26% of the total. Providers publish their half of the sum. Nobody publishes the other half, so we work it out here.
Figures as of June 30, 2026. No site value was changed for this page: both terms come from data already published here.
The arithmetic, term by term
Two terms, two sources, one assumption. Nothing else goes into the total.
Google measures 0.24 Wh as the median for Gemini Apps (21 August 2025), Epoch AI estimates 0.3 Wh, OpenAI states 0.34 Wh as an average. This site uses 0.3 Wh, with a range of 0.24 to 0.4 Wh.
25 Wh per hour of use, the ADEME figure this site already applies to its “Laptop for 1 h” action. Over 2 minutes: 25 × 2 ÷ 60 = 0.83 Wh. This is the shakiest assumption on the page, and the next section tests it.
Electricity in use, and nothing else. No model training, no hardware manufacturing, no network: the exclusions are listed further down.
How long at the screen before it costs more than the AI?
43 seconds. Divide the compute cost by the laptop's draw: 0.3 Wh ÷ 25 Wh per hour = 0.012 hours, or 43 seconds. Past that point, the machine on the desk has pulled more electricity than the servers that produced the answer.
The threshold shifts with the inference figure you pick, but it stays the same size. Google's measured 0.24 Wh brings it down to 35 seconds. OpenAI's stated 0.34 Wh pushes it up to 49 seconds. At the top of this site's range, 0.4 Wh, it reaches 58 seconds. The whole range therefore sits under a minute.
A useful exchange rarely takes under a minute, if only to type the question. That is what makes the result hold: the threshold is crossed in almost every real session, whichever inference figure you pick.
What if you spend less time reading?
The datacentre's share climbs back up, but it only regains the majority below 43 seconds. Here is the same sum for three plausible screen times.
| Screen time | Screen | Total | Compute share |
|---|---|---|---|
| 30 seconds | 0.21 Wh | 0.51 Wh | 59% |
| 2 minutes | 0.83 Wh | 1.13 Wh | 26% |
| 5 minutes | 2.08 Wh | 2.38 Wh | 13% |
Thirty seconds is a short question with a skimmed answer. Five minutes is a working exchange with a reread. Between those two, the compute share falls from 59% to 13%.
The starting assumption is not what decides this. The gap in scale is. Compute is fixed and tiny. The terminal grows with every second of reading, and reading takes tens of seconds.
Why is there no matching total for water and CO₂?
Because adding two different boundaries together would manufacture a fake number. Carbon looks easy: at 480 g per kWh, the operational factor this site uses, 1.13 Wh comes to 0.54 g CO₂e. That sum assumes the datacentre and the reader's desk draw from the same grid.
They almost never do. Google reports 0.03 g CO₂e for its 0.24 Wh, which works out to roughly 125 g per kWh: the mix of an operator buying low-carbon power at scale. The laptop runs on whatever grid its owner lives on. A prompt's carbon total therefore depends on two geographies, and the provider knows only one of them.
Water has the same problem in another form. A prompt's 0.3 mL is cooling water drawn on the datacentre site. The laptop's electricity carries a water footprint too, but upstream, in the plants that generate it, and it varies with the generating mix. Without a dated factor to pin it down, we leave it out.
What we did not count
- The network. A text prompt round trip weighs a few kilobytes, where an hour of streaming weighs millions. Published estimates of network energy intensity diverge sharply depending on what they include, and we have no dated figure we would defend here. It stays out rather than dressed up.
- Model training. Spread over billions of queries, it does not attach cleanly to a single prompt. Our training versus inference comparison handles it separately.
- Hardware manufacturing. The site already puts a laptop at 250 kg CO₂e and a smartphone at 50 kg. Amortising them would mean picking a lifetime and a daily usage figure. That choice would weigh on the result more than the prompt's compute does.
- The providers' own boundaries. Google's 0.24 Wh covers idle machines and datacentre overhead, cooling included. Other published figures do not always say what they cover. That is the first reason they cannot be compared term for term.
Sources
Citing this figure
The 43-second threshold and the 1.13 Wh total are derived from published values, all linked above. Free to reuse with a link to this page. The arithmetic is reproducible: one division, one multiplication, one addition.
Knowing the screen outweighs the model changes what you do with the number: the room to move is in how AI gets used, not in which provider you pick. That is the ground the GHIS studio works on, advising small and mid-sized companies on their AI practices. Here, we only measure.