• Google released a study outlining how much energy and water its AI uses
  • Experts say the figures lack important context
  • Google insists these figures represent the average user experience

A new study from Google claims its Gemini AI model only uses very minimal water and energy for each prompt – with the median usage sitting at around 5 drops (0.26 milliliters) – the equivalent electricity used for 9 seconds of TV watching (roughly 0.24 watt-hours), resulting in around 0.003 grams of CO2 emissions.

Experts have been quick to dispute the claims, however, with The Verge claiming Google omitted key data points in its study, drastically under-reporting the environmental impacts of the model.

Whilst models and data centers have become more efficient, it seems there’s more to the story than Google is letting on.

The tip of the iceberg

One of the authors of a paper cited in the study, Shaolei Ren, associate professor of electrical engineering at the University of California told the publication; “They’re just hiding the critical information. This really spreads the wrong message to the world.”

AI models like Gemini are supported by data centers – huge warehouses full of servers which consume intense amounts of water and energy, straining local resources.

Governments across the globe have been sanctioning the building of these data centers, despite the destruction they could bring to local countryside – and consumers are likely to be the ones paying for the extra energy used.

One of the biggest concerns with Google’s study is that it omits indirect water usage in the estimates, which form the majority of the use related to AI. Whilst the figures are technically correct, the missing context of the extreme energy use paints a misleading picture.

The study only looks at the water used by data centers to cool their servers, but left out is the electricity these data centers demand, which in turn leads to new gas and nuclear plants – which also cool their systems with water, or use steam to turn turbines.

Water isn’t the only metric Google misrepresented though, with the paper only outlining a ‘market based’ carbon emissions measure, which offsets the figure using Google’s promises to use renewable energy to support power grids. Savannah Goodman, Head of Advanced Energy Labs told TechRadar Pro,

“We hope to share environmental metrics that are representative of a typical user’s behavior, and reasonably comparable over time. However, with the rapidly evolving landscape of AI model architectures and AI assistant user behavior, there are outliers either from small subsets of prompts served by models with low utilization or with high token counts.”

“In order to share metrics that represent a typical user’s experience and are robust to this rapidly evolving field, we chose to measure metrics for the median prompt — which is robust to extreme values and provides a more accurate reflection of a typical prompt’s energy impact.”

You might also like