We often talk about the cloud as if it’s abstract. It isn’t. It’s physical infrastructure, and it’s starting to hit very real limits. Water is one of them.
A recent study published in Patterns by researcher Alex de Vries-Gao estimates that by 2025, global AI workloads could consume as much water as the entire bottled water industry. Not as a comparison. As a literal match.
Prompts, Models, and Water Use
Global bottled water consumption is roughly 446 billion liters per year. According to the study, AI-related data center operations are projected to consume between 312.5 and 764.6 billion liters annually.
That puts AI in the same resource category as a global consumer industry.
What matters here isn’t just the volume, but the type of water involved. Data centers rely on potable water, not gray or recycled sources, because mineral-heavy water damages cooling systems. In other words, the water keeping AI systems running is often the same water meant for households, farms, and cities.
Energy Changes the Equation Further
Water is only one part of the footprint.
The study estimates AI systems could emit up to 79.7 million tons of CO₂ per year, roughly on par with the emissions of New York City. Power demand is projected at 23 gigawatts, now exceeding Bitcoin mining.
Even these figures are incomplete. They don’t fully account for “embodied” costs, the water and energy required to manufacture advanced chips and supporting infrastructure. Once those are included, the total impact increases again.
Why This Rarely Gets Talked About
Large tech companies regularly promote long-term sustainability goals, including promises to become “water positive” by 2030. At the same time, their absolute water usage continues to rise.
The reason is simple. Modern AI chips run hot. Extremely hot. To keep them operating reliably, data centers use evaporative cooling systems that consume water and release it into the atmosphere. That water doesn’t return to local supplies.
In regions like Arizona and Iowa, these facilities draw from the same aquifers used by agriculture and nearby communities. The trade-off exists, even if it’s rarely stated outright.
The Bottom Line
Earlier projections underestimated how resource-intensive AI would become. This is no longer just a software story. AI now behaves like heavy industry, with the same physical constraints.
Without clear, mandatory reporting on indirect water use, including the water consumed by power generation, the real cost stays hidden. And by the time shortages show up at the tap, the accounting will already be too late.
