Google did the math on AI’s energy footprint

Ever wonder how much energy it takes when you ask an AI to draft an email, plan a vacation, or role-play as a friend? Google just shed some light: The company calculated the energy, water, and carbon emissions tied to text prompts on its Gemini apps.

The energy use for the median text prompt is far less than some previous estimates: Google found that it uses just 0.24 watt-hours, or the equivalent of watching TV for less than nine seconds. Other models might be wildly different: One estimate of OpenAI’s GPT-5 model suggested that it might use as much as 40 watt-hours per prompt. (That estimate, unlike Google’s calculations, wasn’t based on internal data.)

Google laid out a detailed methodology for measuring the environmental impact of getting answers from its AI. For energy, that includes the electricity burned as the system converts your text prompt into a form it can process, calculates all the probabilities for which word could come next, and decodes the result into readable text. It also includes the energy for cooling and other infrastructure needed at a data center, and for keeping idle machines running in case demand spikes.

The carbon footprint of the median prompt was 0.03 grams of CO2, based on the energy mix for each grid and Google’s own investments in renewable energy. Each prompt also uses 0.26 milliliters, or about five drops, of water.

Efficiency is improving rapidly. Over a 12-month period, the company reports that the median energy use per Gemini prompt has fallen 97%, while the carbon footprint has dropped 98%. Google attributes these gains to advances in its language model architecture, more efficient algorithms, custom-designed hardware, and broader system-wide optimizations.

Still, even if the footprint per prompt is relatively small, the cumulative impact could be enormous as AI use scales. The research Google shared focuses only on text prompts, not on more energy-intensive tasks like video generation. Other companies’ footprints may also differ significantly. But as firms race to build more data centers—and utilities respond by constructing new power plants, often powered by fossil fuels—this is at least an initial step toward understanding how much additional energy AI will truly require.