The company published a report on the environmental footprint of its Gemini chatbot. It explains how the AI works in real-world conditions and what resources it consumes, and these figures turned out to be significantly lower than public projections.
The company published a report on the environmental footprint of its Gemini chatbot. It explains how the AI works in real-world conditions and what resources it consumes, and these figures turned out to be significantly lower than public projections.
ZDNet writes about this, noting that Google has for the first time shared estimates of energy consumption, carbon emissions, and water use during Gemini’s operation.
The company explains that such data is the result of its own methodology, which takes into account not only the work of processors and graphics chips during request processing, but also «background» consumption — the work of RAM, CPU, cooling systems, and even «idle» machines that are kept ready for peak loads.
By comparison, Google estimated a «partial approach» that would have shown lower numbers: just 0.10 Wh of energy and 0.12 ml of water. However, the company emphasizes that such estimates significantly underestimate the real footprint of AI.
Despite the higher numbers than in the simplified calculations, Google notes that Gemini’s overall energy consumption is significantly lower than publicly available estimates. The company emphasized that over the past year, the average footprint of a single text query has been reduced by 33 times in energy consumption and 44 times in CO₂ emissions thanks to improved algorithms and the use of technologies such as «speculative decoding.»
Google’s data release marks the tech giant’s first transparent attempt to show the real resource consumption of AI queries. Previously, OpenAI and Meta had limited themselves to general statements without methodology, which drew criticism from environmentalists and researchers.
According to Reuters/Ipsos estimates, more than 60% of Americans already express concern about the increasing energy consumption of AI. Meanwhile, energy costs for Google’s data centers have doubled in the past four years, despite the company declaring the use of «clean» energy and the intention to reimburse the water consumed.
AI models are increasingly affecting infrastructure: they require powerful data centers that consume enormous amounts of electricity and water. Google is trying to demonstrate that its Gemini has a smaller footprint than expected thanks to optimization and the transition to renewable sources. But experts remind us that with the growing popularity of services like Gemini and ChatGPT, even «a few drops» per request are multiplying into billions of requests, which poses new challenges for energy and ecology.
Previously, dev.ua wrote about how OpenAI CEO Sam Altman published detailed calculations on his blog about electricity and water consumption when processing a single request to ChatGPT.
