Paris Marx’s recent book, Hyperscale, highlights how the surge in artificial-intelligence training has hidden environmental costs. A 2023 university study estimated that building the large language model behind the first ChatGPT version required roughly 185,000 gallons of water. The author notes that each extended interaction with the chatbot effectively discards the volume of a typical water bottle, underscoring the hidden resource intensity of AI development.
Tech firms have responded by suggesting that a single query consumes only a few drops, but independent analysts dispute that narrow accounting. Researchers point out that a hyperscale data center can draw between one and five million gallons each day, a volume comparable to the daily water use of a small city of ten to fifty thousand residents. When multiple facilities cluster, the strain on local supplies becomes starkly apparent.
The rise of generative-AI hardware intensifies the problem. Graphics processing units, which power large language models, generate substantially more heat than traditional server chips, demanding greater cooling capacity. Moreover, data centers require clean water to avoid damaging equipment, leading many operators to tap municipal drinking supplies because treating water themselves would be costlier. This practice pits corporate cooling needs against community water security.
Chile provides a vivid illustration of the conflict. Google’s proposed $200 million facility in Cerrillos was projected to consume 169 liters of water every second, equating to the annual drinking water use of the city’s 80,000 inhabitants in a single day. The plan called for extraction from an underground aquifer already under a long-standing restriction, later upgraded to a total ban. As Sebastián Lehuedé, a researcher at King’s College London, explained, “Usually these reports use quite technical language, and this is a working-class community where not many people are lawyers, for example, or experts on ecology.”
Local activists organized under MOSACAT to scrutinize the proposal and mobilize public opinion. After months of community meetings, a December 2019 referendum showed 49 percent of voters rejecting the project, while 38 percent supported it. Subsequent legal challenges forced Google to revise its cooling strategy, offering an air-conditioning-only design in early 2022 rather than relying on water extraction, marking a rare concession after sustained grassroots pressure.
Similar resistance emerged elsewhere in Latin America. In Uruguay, researchers uncovered that Google’s planned Canelones site would require two million gallons of water per day for cooling, a figure revealed after a court request by Daniel Pena of the University of the Republic, who noted, “Only a tiny proportion of water in Uruguay is used for human consumption… Now we have Google planning to use enormous quantities of water.” In Brazil’s drought-prone Fortaleza region, multiple megaprojects,including a TikTok-linked center the size of twelve football fields,are slated for a free-trade zone despite a history of emergency water shortages.