AI queries consume 500ml of water each, new research shows

Water consumption hidden in every AI interaction

Every time someone asks ChatGPT to draft an email or summarize a document, the request triggers water consumption roughly equivalent to a 16-ounce bottle — about 500 milliliters per query, according to new research from the Association for Computing Machinery.

The estimate applies to what researchers define as a "medium-sized" GPT-3 interaction: approximately 800 words of input and fewer than 300 words of output. With hundreds of millions of users making multiple requests hourly, the cumulative water demand represents a substantial and largely invisible environmental cost.

For context, 500 milliliters is nearly as much water as 75 percent of Americans consume during an entire day, the research notes.

Why it matters

As companies race to build AI infrastructure, water scarcity is becoming a planning constraint, not just an environmental footnote. More than 500 new data centers are planned for regions that have already experienced severe drought. In Texas alone, data centers could account for nearly 9 percent of statewide water consumption by 2040. These facilities draw from the same municipal and regional water supplies that communities depend on, creating potential conflicts over resource allocation during drought conditions.

Cooling drives the demand

The water consumption stems primarily from cooling requirements. AI systems operate in massive data centers packed with servers that generate intense heat. Preventing equipment from overheating requires substantial water resources, either directly through onsite cooling systems or indirectly through the electricity generation that powers air conditioning.

The scale becomes clearer in specific examples documented in the research. Training AI models in Microsoft's U.S. data centers consumed 185,000 gallons of onsite water and approximately 1.4 million gallons overall. A single Google-owned data center used 6.07 billion gallons of freshwater for onsite cooling in 2023 alone.

Demand trajectory points to growing pressure

McKinsey & Company projects that companies will spend $5.2 trillion on data centers by 2030 to meet AI demand. Roughly two-thirds of new AI facilities are being constructed in some of the driest regions of the country, intensifying pressure on already-stressed water systems.

The researchers behind the study argue that efficiency improvements in future AI models may not resolve the problem if overall demand growth outpaces those gains — a pattern common in technology adoption.

Recommendations for mitigation

The research team identifies several potential solutions. They emphasize that companies should implement rigorous measurement and public reporting of water use, including both direct cooling water and the water embedded in electricity generation.

Additionally, they recommend that AI systems be built and trained in locations with better water efficiency, and that computational workloads be shifted to times of day when water resources are less constrained. The researchers stress that transparency around water consumption must become standard practice as AI infrastructure expands.

These findings were first reported by BGR, drawing on research published by the Association for Computing Machinery.