Published on 23/08/2025 - 8:05 GMT+2
ADVERTISEMENT
Running a azygous punctual connected Google’s artificial intelligence (AI) chatbot Gemini takes astir nan aforesaid magnitude of power arsenic watching 9 seconds of TV, according to a caller method paper.
Google says a matter punctual done Gemini uses astir 0.24 watt-hours (Wh) of energy, according to a blog post from nan company. The power emitted from utilizing AI is astir 0.03 grams of c dioxide.
The AI chatbot besides consumes 0.26 millilitres of water, aliases astir 5 drops of h2o per prompt.
The institution measured not only nan power and h2o usage of nan individual prompt, but besides nan IT instrumentality successful information centres, nan idle powerfulness of nan chips that powerfulness nan AI model, and nan h2o utilized by information centres to cool nan instrumentality being used.
The institution claimed that its power and h2o usage for AI is “substantially little than galore nationalist estimates”.
Data shows emissions are ‘quite small’ per search, Google says
One study by the non-profit Electric Power Research Institute estimated that a punctual issued to OpenAI’s ChatGPT uses 2.9 watt-hours of power while a accepted hunt uses astir 0.3 watt-hours.
The International Energy Agency (IEA) predicts that power request will double complete nan adjacent 5 years to 945 terawatt hours per twelvemonth - aliases nan full energy depletion of Japan.
Google’s ain emissions are up too, by 51 per cent since 2019, according to their latest environmental report. This is mostly owed to emissions caused by nan manufacturing and assembly of hardware needed for AI further down nan proviso chain.
The institution said that different studies astir nan power usage of AI often measurement only instrumentality consumption, which “overlook[s]” immoderate of nan different factors that Google utilized successful their metrics, meaning their numbers correspond “theoretical ratio alternatively of existent operating ratio astatine scale,” nan blog station said.
Google claims that its AI models are besides becoming much efficient, pinch power usage and c footprint numbers per punctual dropping by 33 times and 44 times since August 2024.
One of nan study’s shortfalls is that it does not uncover nan full number of queries that Gemini gets each day, which would springiness a amended denotation of Gemini’s full power demand.