Google claims its Gemini AI model consumes 0.26 milliliters of water per prompt, significantly less than previous estimates. However, the report’s methodology has been criticized for drawing a false equivalence between on-site and total water consumption.
Datacenters use water not only for cooling but also for generating energy. This process, known as evaporative cooling, is more power-efficient than refrigerant. According to Google, 80% of water removed from watersheds near its datacenters is consumed by these cooling towers.
UC Riverside researchers estimate that on-site water consumption in the US averages around 2.2 milliliters per request. In contrast, Google’s 0.26 milliliters per prompt is significantly lower. However, this comparison is flawed, as it only accounts for on-site water usage and not the total amount used by datacenters.
The issue lies with the fact that Google compared its results to a study that included all water consumption, while its report focuses solely on on-site consumption. UC Riverside’s researchers point out that if they had only considered on-site water consumption, their estimates would be lower than Google’s.
Google disputes these claims, stating that the research is flawed and doesn’t reflect the actual data center operations. However, this raises questions about the accuracy of the original report and whether it provides a true representation of AI’s water footprint.
Source: https://www.theregister.com/2025/08/22/googles_gemini_water