Google shares AI power, water and emissions data for the first time

Google has published data on the power and water consumption of Gemini in an effort to improve transparency on the environmental impacts of AI.

The tech giant estimates that the median Gemini Apps text prompt uses 0.24 watt-hours (Wh) of energy, emits 0.03 grams of CO2 equivalent and consumes 0.26 milliliters (or about five drops) of water.

These figures are substantially lower than previous public estimates, which placed the energy consumption of an AI query around 2.9Wh. Google’s estimation is even lower than OpenAI’s recently publicised numbers: CEO Sam Altman earlier this year said that each query to ChatGPT consumes about 0.34 Wh of electricity.

Google’s methodology to calculate AI footprint

Google admits that “measuring the footprint of AI serving workloads isn’t simple” – a thought echoed by many in the green digital space.

Its calculations include not just the energy and water used by the primary AI model during active computation, but also the actual achieved chip utilisation at production scale, as well as the energy consumption of idle chips, CPU and RAM energy use, data centre overhead power consumption (such as cooling systems, power distribution, etc), and data centre water consumption.

The low number obtained is said to be the result of having “built efficiency into every layer of AI”, including model architecture, hardware and data centres.

Is Google’s calculation reliable?

While many are hailing Google’s transparency effort at a time of growing concern over the environmental impacts of the age of AI, greenops experts are sceptical – just as they were about OpenAI’s estimation.

First, they note that Google’s estimate does not include image or video prompts and is focused on a median prompt, without sharing any underlying data about these prompts. Second, Google excluded certain sources of energy consumption from its study, including networking before the AI prompt is received, as well as end user device consumption and the power needed to train large language models and store data.

Finally, Google used the Greenhouse Gas (GHG) Protocol’s market-based methodology to calculate the emissions of a prompt, meaning renewable energy attribute certificates (EAC), even unbundled and not related to any on-site energy generation, have been included in the calculation. This is likely to lower the overall emissions number presented by Google, even if location-based emissions are higher. 

“Why do companies continue to do this and hide the true impact? It’s great that Google are working on understanding the impact of Gemini AI but there’s a danger that we undersell the true impact in exchange for some favourable sound bites,” commented Steve Whyley, Climate Lead for Technology at NatWest.

Meanwhile, for Mark Butcher, an outspoken critic of Big Tech’s lack of climate transparency, Google’s technical paper is “a PR spin using cherry-picked stats to present a misleading story”.