It’s the kind of detail that initially seems nearly impossible to comprehend. While you wait for ChatGPT to complete a birthday message, a server rack is heating up somewhere—perhaps in Iowa, rural Georgia, or the high desert outside of Phoenix—and water is subtly evaporating into the sky to keep it cool. Give or take half a liter for each brief session. That seems insignificant. Until you realize that every day, about 2.5 billion prompts are sent in.
The math quickly becomes uncomfortable. In 2023, data centers in the United States used about 17 billion gallons of water, which is sufficient to cover the yearly requirements of a mid-sized American city. In 2022, Google’s facilities alone consumed about 5 billion gallons, a 20% increase from the previous year. According to reports, it took 5.4 million liters to train a single large model like GPT-3 in Microsoft’s American data centers. Furthermore, training is just the start. After that, each question adds a tiny sip.
This is developing in a peculiar way. The majority of the AI boom takes place behind closed doors, such as windowless buildings in industrial parks that are fenced off, unmarked, and humming nonstop. Unless your power bill suddenly increased or the local reservoir began to appear low, you wouldn’t know one was your neighbor. A Meta data center in Newton County, Georgia, is said to draw about 500,000 gallons per day. That would provide several thousand people with enough water to drink. The majority of the local population never consented to that trade.
The contradiction becomes almost theatrical in Phoenix. Data centers throughout the city pull about 385 million gallons annually just for cooling, and that’s not even accounting for the water used upstream to produce the electricity those facilities consume. The metro area is experiencing a protracted drought, with farmers chopping alfalfa fields and homeowners tearing out lawns. It seems more like a pattern than a coincidence that approximately one-fifth of all data centers worldwide are located in water-stressed areas. There, land is less expensive. There are more tax breaks. The question of water is addressed later.
The part that is rarely seen is the cooling towers. The H100s and their successors push thermal limits that older server farms never had to worry about because AI chips are hot. The majority of facilities still use evaporative cooling, which is basically the same concept as a swamp cooler: draw in fresh water, allow it to absorb heat, then release the steam. Approximately 80% of the money taken out is never returned. Dry cooling is an alternative, but it uses a lot more electricity, which means it needs more water somewhere else to produce the power. It’s an obstinate loop.
Some of the suggested solutions are creative, even bizarre. A few startups are proposing data centers that filter and reuse low-quality water, including runoff from pig farms in one instance. Immersion cooling, in which servers are submerged in non-conductive fluid, is evolving from a novel experiment to a significant infrastructure risk. Although it’s still unclear if replenishing an aquifer in Oregon truly makes up for draining one in Arizona, tech giants continue to talk about being “water positive” by 2030. At best, the accounting seems generous.
The silence surrounding all of this makes it more difficult to understand. Water use at the facility level is not required to be disclosed by data centers in the US. Without being aware of their long-term appeal, local authorities approve them. Only when their wells stop producing or when the utility raises rates to finance new pipelines and pumping stations do residents typically find out. Researchers predict that by 2027, the world’s AI water demand could reach 1.1 to 1.7 trillion gallons, which is comparable to California’s annual household use. However, public discourse hasn’t quite caught up.
The asymmetry is difficult to ignore. A query is typed by a user. A model responds in a matter of seconds. Water evaporates into an already dry sky somewhere far away. The ease of use feels unrestricted. In locations that refused to pay, the expense comes later.
