By Taylor Mills

The Failure of Air

Older “brownfield” data centers were designed for server racks consuming between 5 and 15 kilowatts (kW) of power. Today, the solid growth in AI-centric workloads is pushing rack densities to an astonishing 40 to 140 kW. Air is a fundamentally poor thermal conductor. To prevent processors from overheating and failing under this immense load, operators are being forced to gut traditional HVAC systems and pivot to direct-to-chip liquid cooling.

By plunging advanced processors into liquid or running coolant directly across the silicon, facilities can manage the intense heat of AI training. However, this engineering marvel simply trades an electrical bottleneck for an insatiable drain on a rapidly depleting physical resource: freshwater.

The Evaporative Toll

Every query processed by a generative AI model has a physical, liquid cost. According to scientists at the University of California, Riverside, generating a single 100-word AI prompt consumes approximately 519 milliliters of water—roughly the equivalent of a standard plastic water bottle.

When hot water leaves the server racks, it is piped into massive cooling towers outside the facility, where the heat is dissipated through evaporation.

The macro-level statistics of this evaporation are staggering. In 2023 alone, data centers in the United States directly consumed approximately 17 billion gallons of water. The vast majority of this consumption—84%—was driven by hyperscale and colocation facilities. As the AI arms race continues, hyperscale data centers alone are expected to consume up to 33 billion gallons of water annually by 2028. A single large-scale facility can consume up to 5 million gallons of drinking water every single day.

The Unequal Burden on Minority Communities

This massive resource consumption is not distributed equally. Big Tech infrastructure systematically places a disproportionate environmental burden on working-class, Black, and Latine communities.

In California, an astounding 82% of data centers are situated in communities suffering from poor air quality, specifically regarding high diesel particulate matter levels. Furthermore, nearly one-third of all operational and planned California data centers are located in neighborhoods ranked in the top 10% for the worst diesel particulate pollution.

This pattern of environmental injustice is heavily pronounced in the American South. In Bessemer, Alabama—a city with a predominantly Black population—a proposed $14.5 billion hyperscale data center was recently paused due to fierce community opposition. The facility demanded 2 million gallons of water daily. This volume represents roughly the same amount of water required to sustain two-thirds of the city’s entire population. Because low-income Black Americans already face the highest national mortality rates tied to fine particulate matter air pollution, the introduction of massive, energy-and-water-hungry server farms exacerbates a preexisting public health crisis.

The Future of the Grid

Ultimately, the future of artificial intelligence is tethered irrevocably to the physical world. While neural networks operate in the abstract, their survival depends entirely on concrete resources: megawatts of electricity, billions of gallons of potable water, and miles of copper wiring.

The winners of the ongoing AI race will not simply be the organizations that write the most sophisticated algorithms. The true victors will be the thermodynamic engineers, infrastructure managers, and civic planners who can figure out how to keep the world’s most advanced digital brains from burning themselves—and the communities that host them—to the ground.