![]() |
"Chatting with AI for five minutes consumes approximately 500 milliliters of water." Wu Xuchu, vice chairman of KPMG China, recently revealed this statistic, shedding light on the "other side" of AI.
The entire AI value chain — spanning chip manufacturing, power generation, and computing infrastructure operations — consumes 23 billion cubic meters of fresh water annually.
It turns out that AI is not just an electricity guzzler but also a water drain.
To understand why AI consumes water, one must first understand why it gets "thirsty." Simply put, the "brain" of AI consists of chips — specifically CPUs and GPUs. We can visualize these chips as incredibly intricate "electronic sandboxes." Training an AI or prompting it to answer questions essentially involves issuing countless instructions to these chips. At a microscopic level, this means the transistors inside the chips must switch on and off billions of times per second.
"When electric current flows through a conductor, it generates heat — a thermal effect that is unavoidable in chips. This is especially true for modern large-scale AI models, which often feature hundreds of billions of parameters and require exponentially increasing amounts of computing power," said Wang Libang, a PhD researcher at the Institute of Physics, Chinese Academy of Sciences. In short, the greater the computing power of an AI chip, the higher its power consumption and the more waste heat it produces.
In the early days, data centers relied on massive air-conditioning fans to pump in cool air. However, as the density of AI computing power has increased, air cooling has been pushed to its limits; air has a low heat capacity, meaning the rate at which it carries away heat cannot keep pace with the speed at which the chips generate it. Consequently, engineers devised a more efficient solution — "drinking water," known today as liquid cooling technology.
Wang says the specific heat capacity of water — the amount of heat absorbed or released when the temperature of a unit mass of a substance rises or falls — is more than four times that of air; thus, its ability to dissipate heat far surpasses that of air cooling.
While liquid cooling is effective, it consumes vast quantities of water. In some water-scarce regions, competition between data centers and the agricultural sector for water resources has sparked social controversy.
Faced with these challenges regarding energy and water consumption, the tech industry is already seeking alternative "cooling solutions" for AI.
The first strategy is to leverage the natural environment. Since AI is sensitive to heat, one solution is to relocate to cooler regions. Tech giants like Google and Facebook have built data centers in the Arctic Circle, where the cold air cools the servers naturally. Some Chinese companies have submerged their servers in the ocean to take advantage of the cold seawater.
The second strategy is to put AI on a "diet." Since larger AI models are prone to resource waste, scientists are exploring ways to make AI "smarter" rather than simply "bigger." A research team at Peking University developed a medium-scale inference model that achieves the full performance of mainstream large models using only five percent of the parameter count, thereby fundamentally reducing both water and electricity consumption.
The third strategy is to turn waste into a resource. Innovative projects in Europe have begun channeling waste heat generated by data centers into municipal district heating systems to warm residential buildings. Although this technology faces challenges regarding transmission losses and infrastructure retrofitting, it points toward the future: transforming AI from a mere "major energy consumer" into an integral part of the energy cycle.
As AI technology continues its rapid advance, it is time for us to tally the environmental costs behind every human-AI interaction. As for the future of AI, the competition hinges not only on intelligence but also on an environmentally friendly "emotional intelligence." Only when green becomes the fundamental characteristic of every chip and data center can we secure an intelligent, sustainable future.