Google Reveals The Energy Cost Of An AI Prompt

As concerns grow over the environmental toll of artificial intelligence, Google has provided a concrete figure for the energy used by its Gemini AI service. According to the tech giant, a single text prompt requires the same amount of energy as watching television for nine seconds.

In a recent announcement, Google unveiled a new framework designed to measure the environmental impact of its AI models, tracking key metrics like energy consumption, water usage, and carbon emissions.

The Growing Environmental Concerns of AI

Economists have highlighted the potential for AI to significantly boost economic growth and productivity, with a Goldman Sachs report estimating it could add $7 trillion to the global GDP over ten years. However, this potential comes with a significant environmental cost. Scientists are increasingly flagging the massive energy requirements of the data centers that power AI, an impact that is still not fully understood.

Ben Gomes, Google's senior vice president of learning and sustainability, emphasized the need for transparency in a company blog post. "In order to improve the energy efficiency of AI, a clear and comprehensive understanding of AI's environmental footprint is important," he stated. "To date, comprehensive data on the energy and environmental impact of AI inference has been limited."

Beyond electricity, these data centers consume vast amounts of water for cooling, which can strain local water supplies. Research from MIT highlights that this can disrupt local ecosystems. The demand for high-performance hardware also adds indirect environmental impacts from manufacturing and transportation. A recent GAO report noted that some new data centers can require up to 1000 megawatts of energy, enough to power as many as 800,000 homes. Despite this, there are currently no regulations requiring corporations to disclose the energy or water consumption of their AI technologies.

Google Puts a Number on Gemini's Footprint

Google's new technical paper aims to establish a standard for measuring the environmental impact of AI models. The company hopes that by sharing its methodology, it can help the industry better understand and mitigate the environmental effects of this powerful technology.

According to the report, a typical text query sent to Gemini uses:

• 0.24 watt-hours (Wh) of energy
• 0.03 grams of carbon dioxide equivalent (gCO2e)
• 0.26 milliliters of water, which is about five drops

For comparison, OpenAI CEO Sam Altman has previously written that an average ChatGPT query uses 0.34 Wh of energy and a small amount of water.

A New Benchmark for AI Efficiency

Google also highlighted significant progress in making its Gemini platform more efficient. Over the last year, the company reported that the energy consumption and carbon footprint of a median Gemini text prompt have been reduced by factors of 33x and 44x, respectively. Impressively, Google noted that these efficiency gains were achieved while simultaneously improving the quality of the model's responses. This move towards greater transparency and efficiency could set a new benchmark for the rapidly evolving AI industry.