How green is your AI prompt?

Have you ever wondered how much electricity is burned every time you ask an AI chatbot to summarise an article, draft an email, or generate an image of a cute orange cat?

Most of us take these instant answers for granted. But behind the scenes, every query costs energy, consumes water and leaves a carbon trace.

A few months ago, OpenAI's Sam Altman admitted that typing "please" and "thank you" in ChatGPT is not just polite but expensive. Those extra words, he said, waste millions of dollars in computing power. Now, Google has stepped into this conversation with a first-of-its-kind report that tries to put numbers on the hidden cost of AI.

Google's new technical study, titled 'Measuring the Environmental Impact of Delivering AI at Google Scale' reveals that the "median" Gemini text prompt consumes 0.24 watt-hours of electricity. That is the equivalent of running a microwave oven for about one second.

On the water side, each prompt takes 0.26 millilitres, roughly five drops, while carbon emissions per query come to 0.03 grams of CO₂.

For academics who have long tried to track AI's growing footprint, the report is a breakthrough. "This will be a keystone piece in the AI energy field," said Jae-Won Chung, a PhD candidate at the University of Michigan who helps run ML.Energy, a leaderboard tracking AI's energy costs.

What makes the report striking is its scope. It not only looks at the custom chips powering the AI model, but also the wider infrastructure of a data centre. Google's chief scientist Jeff Dean told MIT Technology Review that the team wanted to be comprehensive.

According to the breakdown, the AI chips themselves account for 58% of the total energy use, with CPUs and memory taking another 25%. Backup machines absorb 10%, and cooling plus other overheads cover the rest.

For academics who have long tried to track AI's growing footprint, the report is a breakthrough. "This will be a keystone piece in the AI energy field," said Jae-Won Chung, a PhD candidate at the University of Michigan who helps run ML.Energy, a leaderboard tracking AI's energy costs. His colleague, Professor Mosharaf Chowdhury, added that companies like Google are the only ones who can provide figures at this scale.

The study also shows that Google's energy use per Gemini query has dropped sharply. Between May 2024 and May 2025, the company claims it achieved a 33-fold reduction in electricity consumption. 

Its blog credits software optimisation and hardware efficiency gains for the improvement. Dean framed the findings in relatable terms, saying that using Gemini is no more taxing than watching a few seconds of television or pouring five drops of water.

Although the report was welcomed for its openness, as soon as it was released, critics began pointing out the gaps. Experts speaking to The Verge argued that Google has left out important details, which makes its claims less reassuring than they first appear.

"They are just hiding the critical information," said Shaolei Ren, associate professor at the University of California, Riverside. He argued that the water figure is misleading because it counts only the liquid used to cool servers directly inside data centres. The larger share of water is tied to electricity generation. Power plants, whether gas, coal, or nuclear, also use water to produce steam and prevent overheating. By ignoring this indirect use, Ren believes Google is showing only "the tip of the iceberg".

Alex de Vries-Gao, a researcher at Vrije Universiteit Amsterdam and founder of Digiconomist, echoed this view. He said that location matters. Google uses what it calls a "market-based" emissions figure, which accounts for its investments in renewable energy. But that does not reflect the actual grid mix in the regions where its data centres run. A "location-based" figure, de Vries-Gao stressed, would give the ground truth, and it is usually higher.

There are other omissions too. The report does not reveal how many prompts Gemini handles daily, which makes it impossible to know the total footprint. Nor does it disclose details like word counts or token lengths behind the "median" query. Without these, critics say the results cannot be compared fairly with other studies.

The bigger concern is that efficiency alone does not guarantee sustainability. The phenomenon known as Jevons paradox warns that as technology becomes more efficient, overall use often rises, wiping out the gains. Google's own sustainability report shows that its company-wide carbon emissions grew by 11% last year and by over 50% since 2019.

"If you look at the total numbers that Google is posting, it is actually really bad," de Vries-Gao told The Verge. "This is not telling the complete story."