Does ChatGPT ‘drink’ water? Research reveals hidden cost of every conversation with AI chatbot

ChatGPT drinks 500ml of water for every conversation

In their paper “Making AI Less “Thirsty”: Uncovering and Addressing the Secret Water Footprint of AI Models,” which awaits peer review, researchers from the University of Colorado Riverside and the University of Texas Arlington have revealed the staggering amounts of water AI chatbots “drink” to cool themselves down as they respond to your gruelling questions.

The researchers calculated that training GPT-3 in Microsoft’s US data centers can directly consume 700,000 liters of freshwater, which is equivalent to what it takes to produce 370 BMW cars or 320 Tesla electric vehicles. And if this training were to take place in the company’s Asian data centers, the water consumption would have tripled.

But water consumption does not end at training. For every conversation of about 20-50 questions and answers you have with ChatGPT, the chatbot drinks 500ml of water.

That may not seem like a lot. But given that the tool was the fastest piece of software to reach the 1 million users mark in history and now has over 100 million users, the total combined footprint is enough to cause concern.

ChatGPT not the only water guzzler

ChatGPT is a product of OpenAI, a research organisation that relies on Microsoft’s Azure cloud platform to power its AI innovations. But Microsoft is not alone in the race to develop generative AI. Google has Bard, its own AI chatbot, which may have millions of active users, though the exact numbers are not disclosed.

The inside of an Azure data center. (Image: Microsoft)

The Oregonian revealed that Google’s global data centers guzzled 15 billion gallons of water in 2022, after it challenged the city of Dallas, Oregon to disclose how much water a single Google data center in the region was using. The investigation showed that the company consumed about 1.25 billion liters of water in the area in 2021 alone. That is enough water to supply more than a million households for a year, according to some estimates.

Water consumption and carbon emissions inversely related

Unfortunately, as the paper cited above points out, there’s often a trade-off between carbon efficiency and water efficiency. This is because most solutions focus on engineering aspects, such as improving the cooling systems of data centers. These solutions save water, but ignore the demand factors related to when and where AI models are trained and used.

“For example, AI model developers may want to train their models during the noon time when solar energy is more abundant, but this is also the hottest time of the day that leads to the worst water efficiency,” researchers write, using LaMDA’s training in sun-drenched Nevada as an example.

The challenge here is to find a sweet spot between carbon and water efficiency. One potential solution cited in the paper is ‘spatio-temporal diversity.’ In short, by scheduling AI model training and inference in different places and at different times, developers can reduce the water footprint of their AI models.

Another solution being proposed – while not mentioned in the paper – is shifting entire data centers to colder countries or underwater. Regions that are naturally cold can act as the most ideal place to build data centers which will consume a lesser amount of energy as well as lesser water too. These places are the first choice of green data center providers. Microsoft is one of the companies that has experimented with underwater data centers, and claims that they are reliable, practical, and use less energy than conventional ones.

Different approaches

But Google and Microsoft seemingly have radically different approaches to how they will make their data centers to be more sustainable.

Google’s official water consumption numbers for its data centers. (Image: Google)

In a press release, Google said that it uses seawater or recycled water where possible, and that it invests in renewable energy sources and carbon offsets to power its data centers. It also says that it collaborates with local communities and authorities to ensure that its water usage does not affect the availability or quality of water for other users.

Microsoft, on the other hand, whose infrastructure powers OpenAI as mentioned above, aims to cut the amount of water its data centers use by 95 percent by 2024, with the ultimate goal of eliminating water use altogether. To achieve its zero-water footprint in those hot and dry climates, Microsoft plans to adopt a new way to cool servers: submerging them in fluorocarbon-based liquid baths.

As climate change worsens and the world grows warmer, issues like drought and water shortages are likely to become more frequent. In such a scenario, finding sustainable ways to cool data centers is not only important but essential.