How Much Electricity Does an AI Query Use? Google Offers New Data

The energy consumption of artificial intelligence is a growing concern, particularly as AI tools become more integrated into daily life. A commonly cited statistic suggests that an AI search consumes ten times more electricity than a traditional search. This figure often stems from a comparison between Google's 2009 estimate of 0.3Wh per search and a 2023 estimate by Dutch economist Alex de Vries, which placed a ChatGPT query at approximately 2.9Wh.

“

The median query for the Gemini app text query is 0.24Wh.

However, Google's more recent data presents a different picture. The company's latest statistics indicate that a text query via its Gemini app consumes a median of 0.24Wh. This amount is comparable to running a household microwave, which typically uses around 1000W, for just one second. Google further detailed that within this 0.24Wh, the AI accelerators themselves account for 58% of the power usage. Other components like memory and CPUs use 25%, backup equipment 10%, and cooling and power conversion systems the remaining 8%.

“

Accelerators account for 58% of the power usage.

Experts believe that advancements in model performance, software optimization, and cooling efficiency can significantly reduce the energy required per query. Yet, these figures are based solely on text-based queries. When considering queries involving images, videos, or more complex, multi-step reasoning processes, the energy consumption is likely to be substantially higher. As Jeff Dean, a former Google senior fellow, noted, requests to input and summarize entire books would naturally demand more energy than simpler queries, underscoring the potential for increased power demands as AI capabilities expand.

“

Requests like inputting dozens of books and summarizing the entire content will require much more energy than median queries.