Google says that it actually uses little AI resource in its Gemini AI. A single Gemini AI text prompt consumes just 0.24 Wh of energy, emits 0.03 grams of CO2, and consumes approximately five drops of water according to Google. These figures suggest that how much AI compute is used per prompt is extremely small- less than 9 seconds of TV.
With respect to that, Google says AI prompts actually do not even take much of the resources. The report stresses that AI resource consumption is also significantly below prior public estimates – especially for Gemini AI. As AI models such as ChatGPT, Claude, and Gemini become more widely used, Google’s results suggests that AI resource usage might not be so bad after all.
But experts are yelling “Mayday!” Shaolei Ren, associate professor at the University of California, Riverside and co-author of the paper, as well as Alex de Vries-Gao, founder of Digiconomist, both said Google’s statements were inaccurate. But some experts contend Google’s accounting for AI resource use is much lower than it actually is. One big missing factor is indirect water consumption. AI data centers depend on sophisticated cooling systems that use a lot of water, which Google didn’t take into account in its estimates.
Google also employed a “market-based” measure of carbon emissions in its report that is based on company commitments to renewable energy. It’s a better representation of true AI resource usage, according to experts. The power generation mix is different in different places, and overlooking this might underestimate the AI carbon footprint.
In truth the total AI resource usage, if you take into account energy, water and carbon emissions, could be significantly higher than Google states. Experts caution that concentrating solely on direct indicators of AI prompts is also missing the bigger picture. For example, AI models have an environmental cost, not only from training but from their operational usage, in terms of require lots of GPU compute and the accompanying data center cooling.
As the usage of AI grows, it is important to understand AI resource consumption in the field. Although Google Gemini AI may consume little energy per request, the impact integrated over millions of requests and large AI systems can be significant. Policy experts are calling for transparency in how AI’s energy, water, and carbon use are reported to avoid drawing false conclusions about AI’s actual environmental impact.
The corresponding AI resources are also no small potatoes. It also includes energy use, carbon emissions and water use at scale. An accurate estimation of the resource consumption of AI contributes to an awareness among companies, researchers and users on the real environmental costs of AI technologies and supports a sustainable development of AI.
FAQ
A Gemini AI prompt uses 0.24 Wh of an energy, compared to watching TV for fewer than nine seconds, according to Google.
No, Google’s report doesn’t count indirect water use, like what is used in water used to cool server systems, which experts say can be considerable.
Experts say they’re not fully accurate, because Google relies on market-based as opposed to location-based measures, the latter of which more closely corresponds with actual emissions.
Resource use of AI involves energy, water and carbon emissions. Understanding it could be key in assessing AI technologies’ impact on the environment.
No, experts caution that Google’s claims are misleading and do not account for the full environmental toll of AI consumption.
