AI chatbots guzzle enormous amounts of water, study finds

We’re shielded from the environmental impact of our apps, but the effect is very real. Not only are electricity shortages forcing the Norwegian government to choose between ammunition production and TikTok storage, but also the water required by data centres for cooling is truly astounding.

A new paper from the University of Colorado Riverside and the University of Texas Arlington attempts to estimate a water consumption figure for AI chat models such as Google Bard and ChatGPT. And, if the estimates are accurate, the thirst of such advanced AI models is truly astounding.

The paper estimates that training GPT-3 at Microsoft’s state-of-the-art US data centres would consume 700,000 litres of clear freshwater. This is due to the sheer scale of an operation that Microsoft has revealed contains supercomputers with 10,000 graphics cards and more than 285,000 processor cores.

And that’s a conservative estimate, because training could also be done at the company’s less efficient Asian data centres. If that’s the case, water consumption could be tripled to 4.9 million litres.

The upshot of this is that consumers engaging in a 20-50 question conversation with ChatGPT will see the bot “drink” a 500ml bottle of water, the researchers say.

“While a 500ml bottle of water might not seem too much, the total combined water footprint for inference is still extremely large, considering ChatGPT’s billions of users,” the paper adds.

It’s all about cooling. Data centres, whether used to train algorithms or store TikTok videos, have a lot of computer hardware inside. This in turn generates a lot of heat.

If left unchecked, this could damage the equipment, so server rooms are kept between 10 and 27 degrees centigrade. Because of the constant heat emitted by racks and racks of computers, data centres evaporate water to keep things frosty.

Not just any water, either. It has to be fresh water, because using seawater could cause corrosion of the hardware.

That’s not a great look when draughts are very much a thing across the world. And — while it’s not as headline-grabbing as tech’s carbon footprint — big companies are keen to show they’re working on the problem.

The paper on AI chatbots’ water consumption starts with a selection of quotes from Meta, Amazon and Google. These highlight their awareness of the importance of water security. “Water is a finite resource, and every drop matters,” reads Facebook’s 2020 Sustainability Report.

The paper has a few suggestions as to how AI chat models could be more water efficient.

First, when and where AI models are trained makes a huge difference. Cooler ambient temperatures require less water. This means that models could be trained at night and/or in locations with colder climates.

There is an environmental dilemma here, though. Hotter climates obviously have more sun, and solar power certainly helps to reduce tech’s massive carbon footprint. But, in an environmental catch-22, that heat requires more water consumption. “In other words, only focusing on AI models’ carbon footprint alone is far from enough to enable truly sustainable AI,” the researchers say.

What can consumers do? That’s a bit harder to answer. They could theoretically time their ChatGPT requests for “water-efficient hours”, in the same way that people sometimes run laundry loads overnight when electricity is cheaper. But that’s currently difficult for even the most contentious user due to chatbot makers keeping such data quiet.

“We recommend AI model developers and data centre operators be more transparent,” the researchers conclude.

“When and where are the AI models trained? What about the AI models trained and/or deployed in third-party colocation data centres or public clouds?”

“AI models’ water footprint can no longer stay under the radar — water footprint must be addressed as a priority as part of the collective efforts to combat global water challenges,” the paper concludes.