Unveiling the Unseen Water Expense of AI: A Guide to Computing Your Own Consumption

Unveiling the Hidden Water Consumption of AI: Learn About Your Individual Usage

In the realm of technology, transparency is key to fostering trust and promoting fair comparisons among providers. This is particularly true for Artificial Intelligence (AI), as when companies release their data, the public, policymakers, and researchers can gain a better understanding of what is achievable with AI.

However, the impact of AI on our environment, specifically its water footprint, is a topic that demands attention. It's a common assumption that AI is inherently harmful due to its vast energy and water consumption. But, as we delve deeper, the story becomes more nuanced.

The type of AI model being queried plays a significant role in its resource requirements. For instance, more complex models require greater levels of complexity, hardware, and processor power, which in turn increase their energy and water usage.

Take, for example, the popular AI system, GPT-4o. A medium-length query to this system uses approximately 1.75 watt-hours of energy, which, when multiplied by the water factor of 2 milliliters per watt-hour, results in a water footprint of 3.5 milliliters per response. On the other hand, Google's AI software, Gemini, consumes roughly 0.26 milliliters of water for a text query, mainly for cooling data centers.

OpenAI reports a staggering 2.5 billion prompts per day, including queries to various AI systems like GPT-4o, GPT-4 Turbo, GPT-3.5, and GPT-5. However, there is no public breakdown of how many queries are issued to each particular model.

Multiplied by millions, AI queries' water use adds up. All GPT 4o medium prompts use about 8.8 million liters per day, while All GPT 5 medium prompts use about 97.5 million liters per day. But, if data centers are more efficient, using 1.3 milliliters per watt-hour, the numbers drop significantly, with about 25 milliliters for GPT-5 and 2.3 milliliters for GPT-4o.

Google's disclosure serves as a beacon of what is possible when systems are optimized. Through the use of specialized chips, efficient cooling, smart workload management, recycling water, and locating data centers in cooler, wetter regions, they have managed to reduce their water footprint.

However, it's important to note that AI usage involves understanding not just how to write prompts, but also the infrastructure, trade-offs, and civic choices that surround AI. Behind every AI query are two streams of water use: on-site cooling of servers and the electricity used by power plants to power the data center.

Power plants, including coal, gas, nuclear, hydropower, and concentrated solar plants, use large volumes of water for steam cycles and cooling. Meanwhile, renewable energy sources like wind turbines and solar panels use almost no water once built, aside from occasional cleaning.

Independent researchers and industry reports suggest that a reasonable range for the water per unit of electricity (combining usage for cooling and power) today is about 1.3 to 2.0 milliliters per watt-hour. Newer approaches like immersion cooling and Microsoft's zero-water cooling design offer promising alternatives to reduce water usage, but are not yet mainstream due to cost, maintenance complexity, and difficulty of conversion.

As we navigate the future of AI, it's crucial to remember that every decision we make has an impact. AI usage is not just about writing prompts; it's about understanding the infrastructure, making informed choices, and striving for sustainability.

Leo S. Lo, Dean of Libraries; Advisor to the Provost for AI Literacy; Professor of Education at the University of Virginia, echoes this sentiment, stating, "AI usage involves understanding not just how to write prompts, but also the infrastructure, trade-offs, and civic choices that surround AI."

In conclusion, while AI undeniably has a water footprint, it's not an insurmountable problem. With optimized systems, innovative approaches, and informed decisions, we can ensure a sustainable future for AI.