New estimate reveals surprising amount of water tied to typical AI prompt

A new estimate published by the Association for Computing Machinery suggests that a single medium-sized AI query consumes roughly 500 milliliters of water, or about a 16-ounce bottle. For AI builders shipping inference-heavy products, this figure points to a resource constraint that rarely appears on pricing pages or latency dashboards: the water footprint of every prompt.

What happened

Researchers estimated that a medium-sized GPT-3-like request, defined as roughly 800 words of input and fewer than 300 words of output, can require 500 milliliters of water. Most of that water goes to cooling the servers in massive data centers. The study highlights specific examples: training AI in Microsoft's U.S. data centers used 185,000 gallons of onsite water and about 1.4 million gallons overall, while a Google-owned data center consumed 6.07 billion gallons of freshwater for onsite cooling in 2023.

McKinsey projects that companies will spend $5.2 trillion on data centers by 2030 to keep up with AI demand. More than 500 new facilities are planned in regions that have already experienced severe drought. In Texas, data centers could account for nearly 9% of statewide water consumption by 2040.

Why AI builders should care

If you are deploying models at scale, your infrastructure choices directly affect regional water stress. Data centers draw from the same local water supplies that communities rely on, especially in drought-prone areas. The researchers behind the paper call for better transparency: companies should measure, report, and track water use, including both onsite cooling and the offsite water tied to electricity generation.

For product teams, this means water use could emerge as a compliance or reputational factor. Enterprise customers and regulators increasingly ask about environmental metrics beyond carbon. Water intensity per inference is not yet standard reporting, but the data gap is narrowing. Builders who track water efficiency early may have a competitive advantage in procurement conversations.

Practical implications

What changes operationally? The researchers recommend three actions that map directly to builder decisions:

• Site selection matters. Prioritize data centers in regions with better water efficiency or that use recirculated cooling systems.
• Shift workloads by time. Schedule training jobs and batch inference during hours when water demand on local grids is lower.
• Accept that efficiency gains may not be enough. More efficient models and hardware reduce per-query water use, but rising total demand can erase those savings.

For teams using cloud providers, ask about water usage effectiveness (WUE) in your region. Some providers publish sustainability dashboards. If your vendor cannot provide data, that is a signal worth acting on.

Caveats

The 500 milliliter estimate is for a single GPT-3-like model under specific input/output conditions. Real-world water consumption varies by model architecture, hardware generation, and cooling technology. Google, for example, has claimed that a Gemini text prompt uses roughly 5 drops of water, a dramatically lower figure. The discrepancy underscores how little standardized data exists. Treat the 500ml number as a useful order-of-magnitude indicator, not a precise bill.

The study also focuses on cooling water. Additional water is consumed indirectly through electricity generation, and that offsite use is harder to track. The researchers warn that without consistent reporting, water footprints will remain an under-reported part of AI's environmental impact.