Artificial intelligence, or AI, is transforming industries, but its environmental impact is often overlooked. Shaolei Ren, an associate professor of electrical and computer engineering at UC Riverside, appeared on the stage of TED talk in Vienna, Austria last fall, discussing the hidden consequences of the rapid expansion of AI.
In his speech, “AI consumes a lot of water, why?” explained how the growing demand for AI-powered computing is driving significant water usage, especially in cloud-based data centers. Such facilities may require millions of liters of water to cool servers that power AI models.
Ren’s speech was held at the Ted AI Vienna Conference in October, where experts from the world of AI gathered to discuss its promises and social and environmental consequences. The video of his speech was recently released by Ted Conferences, a nonprofit organization that hosts several TED talk conferences each year. Ted stands for technology, entertainment and design.
Ren’s research found that AI models consume a large amount of water. For example, training the GPT-3, one of Openai’s powerful models, required around 700,000 liters of water by direct evaporation to manufacture approximately 370 BMW or 320 Tesla electric vehicles.
Many AI data centers rely on water cooling systems that generate heat that the power consumption of AI servers needs to be dissipated. This process often involves evaporation of water, such as a cooling tower that converts freshwater into vapor, and is released into the atmosphere for cooling.
“Every time I ask an AI chatbot, I also consume water without realizing it,” Ren told the Ted Audience. “AI doesn’t just need computing power, it requires cooling, and that cooling comes with costs.” AI’s water consumption is called water from rivers, lakes, and groundwater sources, blue water that can be directly used by humans.
With extended droughts becoming the norm, addressing water consumption is essential to ensure that AI infrastructure is responsibly advanced. Ren emphasized that AI companies must take responsibility for their water footprint. He suggested simple but effectiveness
VE Solution: Timing AI training in cool hours to reduce water evaporation.
“We don’t water our lawns at midday, because it’s inefficient,” explained Len. “Similarly, you shouldn’t train when AI models are at their hottest outside. Scheduling AI workloads in the coolest parts of the day can significantly reduce water waste.”
Click here to see Ren’s full TED talk.
Header Photo: UCR’s Shaolei Ren will deliver TED talk at the TED AI Vienna conference in October.
