Most Americans (75%, to be precise) consume only 2.5 cups (591 ml) of water per day. That’s well below the recommended amount, meaning three-quarters of the people you see every day are probably dehydrated. Artificial intelligence, on the other hand, does not suffer from this problem. We already know how much power a single ChatGPT prompt consumes, but a new finding from the Association for Computing Machinery (ACM) shows that an “average-sized” GPT-3 query alone requires 500 ml of water, which is roughly equivalent to that 16-ounce water bottle sitting on your desk. For the record, an average AI request is considered to be an input of around 800 words and a response of less than 300 words – just think about how many words and responses it takes for the AI to write an email for you.
And this trend is not slowing down. A 2025 report from McKinsey & Company shows that companies are expected to spend $5.2 trillion by 2030 to build more data centers to keep pace with growing global demand for AI. In 2023, the on-site cooling consumption of a single Google-owned data center exceeded 6.07 billion gallons (23 billion liters) of fresh water, and more data centers means even more water consumption in the future. At this rate, the water used by AI data centers will eventually rival the annual water needs of some countries.
How much water does AI need?
Gorodenkoff/Shutterstock
Generative AI models require a lot of energy, and while electricity needs and associated greenhouse gas emissions are often easy to understand, the “thirst” for water is less discussed. How much water exactly? The ACM study shows that training Microsoft’s AI data centers in the United States consumed 185,000 gallons (700,000 liters) of water on-site and represented a total consumption of 1.4 million gallons (5.4 million liters) of water. To help you visualize this better, one million gallons equals over 25,000 bathtubs.
These massive data centers are filled with an equally large number of servers: the largest of these facilities are estimated to be large enough to accommodate 2.6 million servers! These servers need to be cooled and local water reserves are tapped for this purpose. Even when a data center uses dry cooling, water consumption remains high and these systems are less efficient when exposed to hot ambient temperatures or high humidity.
Naturally, GPT-3 is old news at this point, but even if future models become more efficient, researchers warn that water consumption may not decrease, as growing demand for AI will offset any improvements. Although many people look down on AI (in part because of its obvious environmental impact), companies are increasingly using it, meaning that even if AI eliminates white-collar jobs and drives computer prices higher, AI’s growth trajectory is unlikely to slow down any time soon.
Is the problem getting worse?
Richard Newstead/Getty Images
Yes. Despite all these signs and warnings, there has been no substantial change in the way data centers approach the problem of water consumption. Case in point: Two-thirds of new AI facilities are built in the driest regions of the United States. More specifically, more than 500 of them will find their home in regions that have experienced severe droughts. This includes Texas, which will likely see data centers account for nearly 9% of the state’s water consumption by 2040.
The researchers who wrote the remarkable paper on AI water costs suggested several ways to make AI more sustainable. Greater transparency would be helpful. Reporting and tracking water consumption could certainly be beneficial, both for cooling needs and the often overlooked costs of offsite power generation. A similar thing applies to finding more suitable areas for training AI models (such as data centers that have better water efficiency) and training models during more water-efficient hours.
Can you change something? Barely. Not writing that ChatGPT email might help you sleep better at night, but the impact of that choice will be microscopic at best. After all, more than a billion people use AI daily, so even if you abandon AI in one fell swoop, you’re still bringing a knife to a fight involving drones and fighter jets. Think about that the next time someone hurts you for using a plastic straw.
