Exploring the hidden environmental impact of generative AI and its implications for global water resources
Every ChatGPT conversation sips water that could otherwise fill thousands of glasses
Explore the ResearchThis case study provides a clear, story-driven examination of ChatGPT's global water use and why it matters. We break down how water is used in both the operation of ChatGPT and the embodied water in its hardware.
Liters of water consumed globally each day – equivalent to a small town's usage
Milliliters of water per user query – about 30 questions uses a 500mL water bottle
ChatGPT does NOT use 39 million gallons daily. That's closer to annual usage.
Difference between water usage in efficiently cooled versus water-intensive regions
Each time you ask ChatGPT a question, it uses between 15-30 milliliters of water, depending on where and how the query is processed. This adds up to significant volumes when multiplied by millions of daily users.
Try typing a prompt below to see how much water it would use:
Our investigation into ChatGPT's water footprint combined rigorous analysis of both direct and indirect water consumption. Below we detail our approach to ensure transparency and reproducibility.
We aggregated data from multiple sources:
To estimate water usage, we applied the following formula:
Total Water = (Energy per Prompt × Prompts per Day × PUE × WUE) + (Energy per Prompt × Prompts per Day × PUE × Water-Energy Factor)
Where:
We modeled three major regions for ChatGPT's operations based on available data:
For each region, we applied the corresponding PUE and WUE values from Microsoft's published data.
Our methodology has several limitations:
Key challenges in this research include:
These challenges highlight the need for greater transparency in AI resource reporting.
ChatGPT's water footprint varies significantly by region. Data centers in different climates have vastly different cooling requirements, and the power grid mix also impacts indirect water consumption.
Select a region on the map to see detailed water usage information.
Approximately half of ChatGPT's operational water use occurs in the Americas. This region contributes the largest share because a large volume of traffic is served from U.S. data centers, with some using water-intensive cooling.
The EMEA water use is relatively low despite significant usage share, because many European facilities have ultra-efficient cooling. For instance, Microsoft's data center in Ireland has a WUE of only ~0.02 L/kWh.
The Asia-Pacific region shows mixed efficiency. Places like Singapore have extremely low water usage (WUE ~0.02), but other locations might have higher water intensity depending on climate and cooling technology.
Out of the ~300,000–500,000 liters consumed daily, roughly 20% is from on-site data center cooling, and the remaining 80% is from power generation. This ratio varies by region – if a data center is powered by renewable energy, the off-site water is minimal.
For every liter of water evaporated directly at a data center to cool ChatGPT's servers, roughly four additional liters are consumed upstream at power plants. Even if a data center doesn't use much water on-site, the electricity it buys may hide a water cost elsewhere.
Different cooling technologies have dramatically different water efficiency. Traditional evaporative cooling consumes substantial water, while newer technologies like liquid cooling in closed loops can reduce or eliminate water consumption.
Uses evaporative cooling towers that consume 1-1.5 liters of water per kWh of IT electricity.
Combines evaporative cooling with air economizers, using 0.2-0.5 liters of water per kWh.
Advanced closed-loop liquid cooling that consumes negligible water (less than 0.05 liters per kWh).
Many claims about AI's water usage circulate online - some accurate, some wildly exaggerated. Let's separate fact from fiction by exploring common myths about ChatGPT's water footprint.
Click to reveal the truth
This viral claim is incorrect. The figure of 39 million gallons is closer to ChatGPT's annual water consumption, not daily. The actual daily usage is around 80,000-130,000 gallons (~300,000-500,000 liters).
Click to reveal the truth
A single ChatGPT query uses about 15-30 milliliters of water on average - closer to a few tablespoons, not a full glass. However, in less efficient data centers, it can approach 50ml per query.
Click to reveal the truth
Only about 20% of the water is used directly for cooling. The remaining 80% is consumed indirectly at power plants generating the electricity that powers AI systems.
Click to reveal the truth
Many modern data centers, especially in cooler climates, use free air cooling or closed-loop systems that require minimal or zero water for cooling. Microsoft has developed zero-water cooling data centers.
How much water does your AI usage consume? Use this interactive calculator to estimate your personal ChatGPT water footprint based on your usage patterns.
There are several strategies that can dramatically reduce ChatGPT's water usage while maintaining performance. These span from technology improvements to operational decisions.
Not all data centers are equal in water efficiency. One straightforward approach is to shift more of ChatGPT's workload to regions or times where cooling can be water-free (or at least water-light).
Scheduling tasks in data centers with the lowest current water usage efficiency can save significant amounts of water.
At night or during cooler seasons, even hot regions might use mostly outside air cooling, saving water.
Running workloads in cool locations like Finland or Sweden when it's hot elsewhere reduces water consumption.
The data center industry is moving toward more water-efficient cooling systems. Traditional air conditioning with cooling towers is giving way to hybrid cooling and liquid cooling that recirculates coolant in closed loops.
Microsoft recently announced a next-generation data center design that uses zero water for cooling, even for AI workloads.
Immersing servers in non-conductive liquid eliminates the need for water-intensive air cooling systems.
Water or coolant circulated in sealed pipes/radiators to dissipate heat, without evaporating away.
The indirect water use from power generation can be reduced by shifting to energy sources that use less water. Solar and wind power use negligible water to generate electricity.
Solar and wind power require virtually no water to generate electricity, unlike traditional thermal power plants.
Running workloads when the grid is powered by more renewables reduces the water intensity of electricity.
Data centers with on-site solar or wind generation can bypass water-intensive grid electricity.
One of the reasons for wild claims is the lack of easily accessible data. AI providers should disclose more granular water footprint information for better accountability.
Consistent methodologies for measuring and reporting AI's water usage would enable meaningful comparisons.
Microsoft has pledged to replenish more water than it consumes by 2030, driving innovation in water conservation.
Real-time monitoring and reporting of water usage per service would drive accountability and innovation.
This research draws on multiple sources across academia, industry, and public policy to provide a comprehensive picture of ChatGPT's water footprint.
ChatGPT's global water use, while often out of sight, should not be out of mind. We found that the scale is meaningful but manageable – it's a new factor to consider in sustainable tech.
ChatGPT uses significant but not catastrophic amounts of water – roughly 500,000 liters daily worldwide.
Technology solutions exist today that could reduce this footprint by 80-90% through better cooling and energy choices.
Regional variation is enormous – the same AI task might use 5-10× more water in a hot region versus a cool climate.
By clearly understanding the numbers and mechanisms, we can debunk false claims and focus on real solutions. The story of ChatGPT's water footprint is still being written, and with the right choices, it can become a success story of innovation meeting responsibility.