Google claims its Gemini AI assistant is significantly more efficient than earlier research suggested, requiring only about 0.26 millilitres of water (five drops) and electricity equivalent to running a TV for under nine seconds per prompt. This translates to roughly 0.03 grams of CO₂ emissions. According to Google, recent efficiency improvements have cut energy use per prompt by 33 times and reduced its carbon footprint by 44 times, challenging the notion that generative AI is a heavy resource consumer.
Per new paper, Google says a median Gemini text prompt uses ~0.24 Wh (9 seconds of TV), 0.26 mL water (~5 drops).
~58% of the energy is used toward the accelerator, 25% is the host CPU/DRAM, 10% is “idle” capacity kept ready for reliability, and 8% is datacenter overhead… pic.twitter.com/17JNfPzVG6
— Shanu Mathew (@ShanuMathew93) August 21, 2025
Independent researchers have raised doubts about Google’s claims on Gemini AI’s efficiency, reported The Verge. Shaolei Ren, a professor at the University of California, Riverside, whose research Google cited, argued that the company’s study leaves out critical data. He noted that while Google only measured the water used directly in cooling its data centres, it ignored the much larger amount of water consumed in generating the electricity that powers those centres.
ALSO SEE: Instagram Rolls Out Feature To Link Reels In A Series: Here’s How It Works
Experts add that power plants, whether gas or nuclear, depend heavily on water for cooling and steam production. With AI driving up electricity demand, this indirect water use becomes a significant factor in the environmental footprint. Alex de Vries-Gao of Digiconomist stressed that Google’s numbers reveal only “the tip of the iceberg,” masking the broader impact.
Google says a typical AI text prompt only uses 5 drops of water — experts say that’s misleading https://t.co/OGPaKAHft4
— The Verge (@verge) August 21, 2025
The study’s treatment of emissions data has also sparked criticism. Google reported “market-based” figures that reflect renewable energy purchases, but critics say this overlooks “location-based” data, which accounts for the actual energy mix, fossil fuels or renewables, where data centres run. Researchers argue that location-based measures give a more realistic picture of emissions, which tend to be higher than Google’s estimates.
Another concern is methodology: Google used a median prompt to calculate water and energy use, while most academic studies rely on averages. Without disclosing details like token length, experts say comparisons are flawed. Google further claimed Gemini consumes 0.26ml of water per prompt, far below earlier studies suggesting up to 50ml. Ren countered that his research included both direct and indirect water use, unlike Google’s narrower scope. Google has not submitted the study for peer review but says it is open to doing so, pledging continued efficiency improvements.
ALSO SEE: iPhone 17 Series Launch Event Date Accidentally Confirmed By Apple
Great Job Priya Singh & the Team @ Mashable India tech Source link for sharing this story.
