ChatGPT's massive thirst revealed in Iowa

Every time you fire off a prompt to ChatGPT, a response pops up in seconds. It is easy to think of this as pure code running in the cloud, maybe burning a bit of extra electricity. But keeping those silicon brains from melting requires something far more physical: millions of gallons of fresh water.

An investigation published by the Associated Press on September 9 pulled back the curtain on the massive environmental toll of generative AI. While the engineers may be based in Silicon Valley, Microsoft executives let slip that OpenAI's GPT-4 was actually trained next to the cornfields of West Des Moines, Iowa, drinking heavily from local rivers to stay cool.

A supercomputer next to the cornfields

Training a cutting-edge large language model takes staggering computing power. In 2020, Microsoft built a dedicated supercomputer for OpenAI packed with more than 285,000 CPU cores and 10,000 GPUs. All that hardware generates intense heat, and data centers rely heavily on water-cooling towers to dump that thermal energy outside.

West Des Moines seemed like a smart choice on paper. For most of the year, cool Midwestern air is enough to chill the servers. Microsoft publicly noted that it only pulls water into its cooling towers when temperatures climb above 29.3 degrees Celsius.

Yet summer heatwaves tell a different story. In July 2022, just a month before OpenAI wrapped up GPT-4 training, Microsoft pumped about 11.5 million gallons of water into its local data center cluster. According to local utilities, that single month accounted for roughly 6% of all the water used in the district, which also supplies the town's drinking water.

Half a liter per conversation

It is not just training that is thirsty. Everyday inference adds up fast.

Shaolei Ren, a researcher at UC Riverside, calculated that querying ChatGPT roughly 5 to 50 times consumes about 500 milliliters of water, roughly equivalent to a standard 16-ounce water bottle. That figure accounts for both direct cooling at the servers and indirect water consumed by the power plants generating electricity for the cluster.

Corporate reports back this up. Microsoft's global water consumption spiked 34% between 2021 and 2022 to nearly 1.7 billion gallons, enough to fill over 2,500 Olympic swimming pools. Google saw a 20% jump over the same window. Researchers attribute the bulk of that surge straight to generative AI workloads.

Microsoft responded that it is researching better ways to track AI efficiency while sticking to its goal of becoming water positive by 2030, and OpenAI echoed similar sentiments. But local officials are already pushing back. The West Des Moines Water Works cautioned that future data center expansions will only be considered if Microsoft can prove it has significantly cut its peak water usage.

Why it matters

This story shatters the illusion that cloud AI is an immaterial technology. Behind every clever chatbot is a massive physical footprint competing directly with local communities for basic utilities.

For businesses and developers, this shifts the calculus around model selection. When calculating ROI, pure token costs won't be the only metric that matters. Regulatory scrutiny and public pushback over resource usage will increasingly favor compact, efficient models that achieve similar results without boiling the server rack.

For the tech giants, the AI arms race is colliding head-on with infrastructure limits. Microsoft was relatively clever to build in Iowa rather than a desert state like Arizona, using free outdoor air cooling whenever possible. But peak summer loads remain a glaring vulnerability. If data centers start competing with local taps for drinking water, local governments won't hesitate to pull the plug on new permits. The next phase of AI supremacy won't just be won by who has the most GPUs, but by who solves the thermal engineering puzzle.