The Thirsty AI

You’ve probably heard about the massive energy consumption of artificial intelligence, but there’s a hidden, and arguably more pressing, environmental cost that often goes unmentioned: water. As powerful AI models like ChatGPT become integrated into our daily lives, a new study reveals the staggering amount of water they consume, raising critical questions about the sustainability of the AI revolution.

The Thirsty AI: Why Your Digital Assistant Needs a Drink

When you type a query into a chatbot, you're not just communicating with lines of code. You're activating a vast network of servers housed in massive data centers. These servers are running sophisticated machine learning models, and all that computational power generates an immense amount of heat. To prevent the hardware from melting down, these data centers rely on sophisticated cooling systems—and that's where the water comes in.

Data centers use massive cooling towers that function like giant swamp coolers. They pull in outside air and use water evaporation to cool the equipment. The process of turning water into vapor is what "consumes" the water, as it's released into the atmosphere and isn't recycled back into the local water supply. This isn’t a small amount, either. A single data center can use hundreds of thousands, or even millions, of gallons of water per day.

According to a study from researchers at the University of California, Riverside, and the University of Texas, Arlington, the water footprint is surprisingly specific. They estimated that Microsoft’s data centers, which host OpenAI’s models, consumed an estimated 185,000 gallons of water to train the GPT-3 model alone. If the model had been trained in a larger data center in Asia, that number could have been tripled. To put that into perspective, 185,000 gallons is enough water to produce 370 BMWs or 320 Tesla electric vehicles.

And the water consumption doesn't stop after the initial training. Every single time you use the AI, it "drinks" more water. The researchers estimated that a simple conversation with ChatGPT, consisting of 20-50 questions and answers, consumes about a 500ml bottle of fresh water. While that may not sound like much for a single interaction, consider that people send over a billion messages to AI chatbots every single day. The total amount of water consumed for these conversations is astronomically large and continues to grow as AI becomes more ubiquitous.

The Human and Environmental Cost

The environmental cost of this thirst extends far beyond the data center. The vast majority of data centers are located in areas with cheap land and, more importantly, cheap energy. This often puts them in regions that are already water-stressed or prone to drought.

Take, for example, the case of West Des Moines, Iowa. Microsoft operates several data centers there, and their water consumption has been a source of local controversy. In 2022, Microsoft’s data centers guzzled a staggering 1.7 billion gallons of water globally, with a significant portion of that coming from its Iowa facilities. This occurred during a multi-year drought, raising the ire of local residents and community leaders. Critics pointed out the absurdity of a non-essential technology consuming so much water while local farms and households struggled with water scarcity.

The issue is made worse by the fact that the water used in cooling towers is not easily recycled. It often evaporates or is drained as a concentrated, briny wastewater that requires further treatment before it can re-enter the local water system. This adds another layer of environmental strain to an already taxed resource.

Furthermore, the water consumption for AI is not limited to just data center cooling. The production of electricity itself is a water-intensive process, especially for fossil fuel and nuclear power plants. Since AI requires a massive amount of energy, it also has an indirect "water footprint" through the power grid.

Towards a More Sustainable AI

So, what can be done? The good news is that tech companies are not ignoring the problem entirely, even if they are slow to release specific data. The researchers and other experts have pointed to several key solutions:

  • Public Accountability: The first step is transparency. Tech companies need to publicly report their water usage in the same way they are increasingly reporting their carbon emissions. This would create a powerful incentive for them to find more efficient solutions and would allow communities to make informed decisions about hosting data centers.

  • Location, Location, Location: A simple but effective solution is for companies to build new data centers in cooler, less water-stressed regions. Locating facilities in colder climates reduces the need for water-based cooling and allows for the use of more energy-efficient methods like "free cooling" (using outside air).

  • Technological Innovation: Companies are also exploring new cooling technologies, such as liquid immersion cooling, where servers are submerged in a non-conductive fluid. This method is far more efficient and uses significantly less water than traditional cooling towers.

  • Algorithm Optimization: Researchers are also advocating for more efficient algorithms and hardware. A less resource-intensive model would require less energy, which in turn would reduce both the carbon and water footprints.

Ultimately, this is a conversation we must all have. As we embrace the incredible power and convenience of AI, we must also acknowledge its environmental toll. The "cloud" is not a magical, ethereal place; it is a physical reality with a very real and tangible impact on our planet. Acknowledging this hidden cost is the first step toward building a truly sustainable and responsible AI-powered future.

What our community are reading

Moonshots, Models, IoT and Machine Learning in Agriculture

  What do Google, Climate Corporation, early stage startups in farm robotics, and researchers trying to figure out how to feed the world sustainably have in common? They’re all grappling with one of the toughest challenges of working with natural systems - how do you work with data that is sparse, unevenly distributed and with systems that have so many connections and interactions with other systems?     Before the advent of cheap sensors that are connected to phones, easily accessible satellite data and drones that can fly over fields quickly and inexpensively -     scientists in companies and academia worked on developing plant and crop models that incorporated as many aspects of the farm and as much data as was available so that they could understand and predict what was likely to happen on the field. Understandably, the forecasts took some time to produce and as the models grew more complex, issues about how to estimate model parameters and the uncertainty a...
Read more

What’s the impact of your smartphone? Mapping the distribution and environmental impact of mines around the world

Image
How many different materials go into making your smartphone? According to the American Chemical Society, at least 64 different types of chemicals constitute your smartphone . Many of these materials, particularly the heavy earth metals, are extracted from the Earth through mining operations. Mining has been a part of human society for several thousand years now - from the shallow mines of earlier centuries to the more extractive ones created in the last hundred years. The challenge of mining and other extractive industries is that they are essential to many of our needs today - coal for energy, silicon for solar panels and heavy earths for electronics being among the well known ones. At the same time, they cause significant damage to local environments through the disturbance to the land caused by establishing the mine itself, to environmental impacts on water, land and ecosystems from mining operations.   So far, it’s been difficult to map the spatial extent and environmental impa...
Read more

Our online community space is now live!

 Our online community space is now open to anyone who has signed up for a free or paid course on our website! In addition to everyone who signed up for our cohort-based courses, we're now expanding it to all the members of our community. If you've already signed up for any of our courses, check your email for the invitation for the space. It's where we'll get together to talk about all things data science and clean technology related, discuss the latest research, network and make connections with other professionals in the sector. It's an invitation only , no bots and no trolls allowed space - so come on over! Here's where you can check out our courses and join our community !
Read more