Tiny energy cost per AI prompt could still carry global impact

- Google’s first comprehensive study shows a Gemini chat consumes 0.24 Wh of energy, 0.03 grams of CO2, and five drops of water

Google’s first comprehensive study of its Gemini AI shows that each prompt consumes just 0.24 watt-hours of energy—about nine seconds of TV time—but the company cautions that billions of daily interactions worldwide could add up to a significant environmental footprint.

According to the report, a single Gemini prompt generates 0.03 grams of carbon emissions and uses just 0.26 milliliters of water, which the company likened to "five drops of water."

The report underlined that the footprint of AI chats is far lower than previously thought. Between May 2024 and May 2025, the average energy use per Gemini prompt dropped by a factor of 33, while carbon emissions fell 44 times.

Software improvements and model optimization accounted for a 23-fold efficiency gain, with machine utilization adding 1.4-fold. Meanwhile, Google's clean energy procurement reduced the carbon intensity of electricity use by another 1.4 times.

Despite the relatively small environmental impact of a single chat, Google cautioned that the global scale of billions of interactions means the cumulative effect must be monitored.

- Narrow estimates fall short

Google emphasized that many earlier studies produced misleading results by considering only the power of active AI processors. Its methodology instead factored in idle machines, CPU and memory use, and data center cooling overhead.

The difference was significant: the same Gemini prompt measured at 0.10 Wh under the narrow existing approach rose to 0.24 Wh when all relevant factors were included.

Each Gemini prompt was linked to 0.26 milliliters of water consumption, far below earlier estimates of 45–50 milliliters per chat.

The report recalled that since 2023, Google has pledged to use air-cooled systems in new data centers located in water-stressed regions, aiming to reduce AI-related water use in such areas to near zero.

- Efficiency gains at scale

The findings also showed a dramatic reduction in emissions: Scope 2 emissions from electricity purchases dropped 47-fold, while hardware-related Scope 1 and 3 emissions decreased 36-fold in the past year.

"As the capabilities of AI advance, its environmental efficiency must advance at the same pace," Google researchers said.

The report did not disclose figures on daily prompt counts, total usage volumes, or Gemini’s overall user base.

By Murat Temizer

Anadolu Agency

energy@aa.com.tr