In our wonderful world of technology, artificial intelligence (AI) emerges as a transformative force, reshaping industries and our daily lives with astonishing speed. From virtual assistants to complex medical diagnostic models, AI promises a future of efficiency and innovation. However, behind the digital facade and instant answers, there is a significant and seldom-discussed environmental cost: a massive consumption of water.
The thirst of artificial intelligence is not metaphorical. It resides in the pulsating heart of the infrastructure that sustains it: data centers. These vast facilities, filled with high-performance servers, are the brains behind AI operations. The intense processing required to train and run complex language models, like those that power chatbots and image generators, creates an immense amount of heat. To prevent overheating and ensure continuous operation, these servers must be constantly cooled, and water is the most commonly used resource for this purpose.
Recent studies are beginning to quantify this water footprint. Research indicates that a simple conversation with a chatbot can consume the equivalent of a bottle of water. While it may seem small, the massive scale of daily interactions with AI systems worldwide raises this consumption to alarming levels. Tech giants like Google and Microsoft report a significant increase in the water consumption of their data centers, a trend directly correlated with the growing demand for artificial intelligence services. We are talking about billions of liters of water annually, a volume that could supply entire cities.
The environmental impact of this consumption is particularly concerning in regions already facing water stress. The strategic location of many data centers, sometimes in areas with arid climates but with good connectivity and affordable energy, creates a direct conflict with the needs of local communities and ecosystems. The “cloud,” which we perceive as something ethereal and immaterial, actually has a physical manifestation with a very real thirst, competing for a resource that is fundamental to life.
Fortunately, awareness of this issue is growing, and with it, the search for innovative solutions and more sustainable practices. The tech industry is beginning to explore more efficient cooling methods, such as closed-loop systems that recycle water, and the use of non-potable water sources, like seawater or treated wastewater. Furthermore, building data centers in colder climates can drastically reduce the need for artificial cooling.
The development of more energy-efficient hardware, which consequently generates less heat, is also a crucial front for advancement. Chips and processors designed specifically for AI tasks can decrease the demand for cooling. Optimizing the AI models themselves, making them smaller and more efficient without sacrificing performance, represents another promising path to mitigate this water footprint.
As users and technology enthusiasts, it is essential that we have a complete view of the impact of the tools we use. Artificial intelligence will undoubtedly continue to be a driving force for progress. However, its development and implementation must be guided by sustainability. The hidden thirst of AI reminds us that every innovation has an environmental cost and that true intelligence lies in recognizing and addressing these challenges responsibly, ensuring that technological advancement does not come at the expense of our most precious natural resources.
The Drastic Evolution in the Water Consumption of LLMs
In recent years, the rise of Large Language Models (LLMs) has redefined the frontiers of technology. However, behind the almost magical ability to generate text, images, and code, there is an increasingly evident environmental consequence: a water consumption that has grown exponentially. The evolution of this digital “thirst” is a story of scale, complexity, and the transition from a one-time cost to a continuous global demand.
The Pre-LLM Era: A Diluted Consumption
Before the explosion of LLMs, around 2018, data centers were already large consumers of water, using it mainly to cool the servers that supported the internet, social networks, and cloud services. However, this demand was more distributed and predictable. Intensive processing was not as concentrated in single, colossal tasks like training an artificial intelligence model with hundreds of billions of parameters.
The Turning Point: GPT-3 and the Awareness of the Water Footprint
The real game-changer, which brought the issue of AI’s water consumption into the public debate, was the training of OpenAI’s GPT-3, completed in 2020. A pioneering study by the University of California, Riverside, estimated that the training of this model alone consumed approximately 700,000 liters of fresh, clean water. This volume, used to cool Microsoft’s supercomputers, would be enough to produce more than 300 electric cars. GPT-3 not only demonstrated a new level of linguistic capability but also revealed the immense resource footprint required to achieve such a feat.
The Post-2022 Explosion: The Continuous Thirst of Inference
If the training of GPT-3 was a massive, one-time consumption event, the global popularization of tools like ChatGPT and the integration of even more complex models, like GPT-4, inaugurated a new era of water consumption: inference.
Inference is the process of using the already-trained model to respond to the requests of millions of users every day. Every question, every generated sentence, every suggested line of code requires computational power, which generates heat and, consequently, consumes water for cooling. Research indicates that a casual conversation with a chatbot, involving 20 to 50 interactions, can consume the equivalent of a 500 ml bottle of water.
This shift from a “training” consumption (one-time and massive) to an “inference” consumption (continuous and global) has caused water demand to skyrocket. GPT-4, being considerably larger and more complex than its predecessor, has an even more significant water footprint per inference, although exact numbers on its training are not public.
The Reflection in the Tech Giants
The evolution of this consumption is clearly visible in the sustainability reports of the very companies leading the AI race:
- Microsoft: The main infrastructure partner for OpenAI reported an impressive 34% increase in its water consumption between 2021 and 2022, a jump directly correlated with its massive expansion in AI to train and operate OpenAI’s models.
- Google: In the same period, while developing its own competing models like LaMDA and PaLM, Google registered a 20% increase in its water consumption.
In short, the evolution of water consumption by LLMs can be summarized in three phases: an already high consumption baseline from traditional data centers; a quantum leap with the training of massive models like GPT-3; and an exponential acceleration driven by the daily and global use of generative AIs by millions of people, transforming artificial intelligence into one of the new and thirstiest sectors of the digital economy.
But it is important to say that efficiency in water use per unit of computation is increasing.
Improvements in Efficiency (How Companies Are Using Less Water per Task)
The tech giants are aware of the financial and reputational costs associated with excessive water consumption and are investing heavily in efficiency. The improvements come from several fronts:
- Cooling Innovations: The most modern data centers no longer use the old method of simply “dumping” cold water and discarding the hot. Current techniques are much more sophisticated.
- Closed-Loop Systems: Water circulates in a closed system, being cooled and reused continuously, which drastically reduces the need to draw new water.
- Use of Alternative Sources: Companies like Google and Microsoft are building data centers that can be cooled with non-potable water sources, such as seawater, industrial water, or treated wastewater, preserving fresh water for communities.
- Smart Evaporative Cooling: Instead of relying solely on chillers (mechanical coolers), many data centers use evaporative cooling, which is more efficient. AI systems now control these processes, using the minimum amount of water necessary based on external temperature and humidity.
- Strategic Location: The choice of where to build a data center is crucial. Increasingly, companies are opting for cold-climate regions (like the Nordic countries). In these locations, they can use what is known as “free cooling,” pumping cold outside air to cool the servers for most of the year, drastically reducing the dependence on water-based cooling.
- Hardware and Software Efficiency:
- Hardware: New generations of GPUs (Graphics Processing Units) and TPUs (Tensor Processing Units) are designed to be more energy-efficient. Less energy waste means less heat generated, and less heat means less need for cooling.
- Software: Techniques like “model distillation” and “quantization” are being developed, which create smaller and “lighter” AI models capable of performing tasks with much less computational power and, consequently, a smaller water footprint.