The Correlation Between AI and Fresh Water Consumption
Have you ever thought about the environmental impact of using AI chatbots like OpenAI’s ChatGPT or Google’s Bard?
These large language models require massive amounts of energy to train, and the data centers that house them need to be cooled, making them incredibly thirsty. In fact, new research suggests that training GPT-3 alone consumed 185,000 gallons (700,000 liters) of water. That means that each time you have a conversation with ChatGPT, you’re essentially wasting a large bottle of fresh water. With the chatbot’s popularity on the rise, researchers are concerned about the impact this could have on our water supplies, especially in light of the historic droughts and environmental uncertainty we’re facing.
In a pre-print paper titled Making AI Less ‘Thirsty, researchers from the University of California Riverside and the University of Texas Arlington estimated the amount of fresh water required to train GPT-3. They found that it was equivalent to the amount needed to fill a nuclear reactor’s cooling tower. And while OpenAI hasn’t disclosed how long it takes to train GPT-3, Microsoft, which has partnered with the AI startup and built supercomputers for AI training, says that its latest supercomputer contains 10,000 graphics cards and over 285,000 processor cores. That gives us an idea of the scale of the operation behind artificial intelligence. Each time ChatGPT complete a basic exchange with a user consisting of rough 25–50 questions, it would need to “drink” a 500-mL water bottle.
Did you know that water consumption is a major issue for tech companies, not just for AI models like OpenAI’s ChatGPT or Google’s Bard?
In 2019, Google requested over 2.3 billion gallons of water for its data centers in just three states. And with 14 data centers across North America powering Google Search, workplace products, and large language models like LaMDA and Bard, the company’s water consumption is staggering. In fact, according to recent research, training LaMDA alone could require millions of liters of water.
But it’s not just about water. These large language models also require huge amounts of electricity to train. A recent report from Stanford AI estimated that training OpenAI’s GPT-3 released 502 metric tons of carbon. That’s enough energy to power an average American home for hundreds of years.
As Kevin Kent, CEO of Critical Facilities Efficiency Solution, said in an interview with Time, “The race for data centers to keep up with it all is pretty frantic. They can’t always make the most environmentally best choices.” So next time you use an AI model or search the web, remember that there’s a hidden environmental cost.
When it comes to AI’s water consumption, there’s a difference between water “withdrawal” and “consumption.” Withdrawal refers to physically removing water from a source, while consumption refers to the loss of water through evaporation when it’s used in data centers. The research on AI’s water usage focuses primarily on consumption, where the water can’t be recycled.
Data centers need to be kept cool to prevent equipment from malfunctioning. This is a challenge because the servers themselves generate heat as they convert electrical energy. Cooling towers are often used to counteract this heat by evaporating cold water. However, this process requires large amounts of water — around a gallon for every kilowatt-hour expended in an average data center.
Not just any type of water can be used in data centers. They need clean, fresh water to avoid corrosion or bacteria growth that can come with seawater. Freshwater is also essential for humidity control in the rooms. Data centers are also held accountable for the water needed to generate the electricity they consume, something the researchers call “off-site indirect water consumption.” So, while AI models like ChatGPT and Bard may seem like they’re just virtual entities, their operation has a very real impact on our environment.
Water scarcity is already a major issue in the US, with 2.2 million residents lacking access to water and basic indoor plumbing, and another 44 million living with inadequate water systems. Climate change and population growth are expected to exacerbate this problem, with Stanford estimating that by 2071, nearly half of the country’s 204 freshwater basins will be unable to meet monthly water demands. Many regions could see their water supplies cut by a third in the next 50 years.
In light of these challenges, AI’s water consumption is becoming a growing concern. As data requirements for large language models continue to increase, companies will need to find ways to improve their data centers’ water efficiency. Researchers suggest that there are several ways to reduce AI’s water consumption, such as training models at cooler times of day or in more efficient data centers. Chatbot users could also engage with the modules during “water-efficient hours.”
However, achieving these demand-side changes will require greater transparency from tech companies building these models. The researchers recommend that AI model developers and data center operators be more transparent about when and where their models are trained, as well as any third-party colocation data centers or public clouds used.
In conclusion, AI’s water consumption is a complex issue that requires a multi-faceted approach to address. While there are steps that can be taken to reduce water usage, greater transparency and cooperation from tech companies will be essential in ensuring a sustainable future for both AI and our planet.
