AI Data Centers May Force $58B in Water Infrastructure Upgrades

The real water cost of AI isn't measured in bottles per prompt

When communities evaluate data center proposals, developers typically share annual water consumption figures. Those numbers miss the engineering reality: water systems must be sized for the hottest day of the year, not the average Tuesday in April.

Research from UC Riverside and Caltech estimates that U.S. community water systems may need between $10 billion and $58 billion in new infrastructure by 2030 to serve data center growth, assuming current efficiency trends continue. The wide range reflects uncertainty about expansion rates and cooling technology adoption. The core finding remains consistent—peak daily demand, not annual totals, determines what cities must build.

Why it matters

Local ratepayers typically fund water infrastructure through utility bills. When a data center requires new treatment capacity, pipelines, and pumps to handle summer peak loads, those costs get distributed across the community unless developers contractually share the burden. The infrastructure sits underutilized most of the year but must be financed and maintained for decades.

Per-prompt figures vary wildly and obscure the issue

Public debate has focused on how much water a single AI query consumes. In 2024, The Washington Post and UC Riverside researcher Shaolei Ren estimated 519 milliliters for a 100-word GPT-4 response at an average U.S. data center, including cooling and electricity generation. Ren has since revised that figure to approximately 15 milliliters for current systems, with about five milliliters for direct cooling.

OpenAI CEO Sam Altman cited roughly 0.32 milliliters per average ChatGPT query, though without sufficient methodology to compare directly. A 2025 benchmarking study titled "How Hungry is AI?" found estimates ranging from under two milliliters for efficient models to more than 150 milliliters for complex reasoning tasks.

These figures shift with model architecture, output length, hardware generation, cooling design, power source, and location. They're useful for comparing models but tell city planners little about capacity requirements.

Peak demand multiplies annual averages by up to 30x

The UC Riverside and Caltech team found that facilities using evaporative cooling can see daily water demand rise to six to 10 times their annual average during hot weather. Some planned projects show peak multipliers exceeding 30. Large facilities can consume over one million gallons on a hot day, with some under construction holding allocations up to eight million gallons daily.

Without efficiency improvements, researchers estimate U.S. water systems could need an additional 697 million to 1.45 billion gallons of peak daily capacity by 2030—roughly equivalent to New York City's daily supply. That capacity remains idle most of the year while requiring continuous financing and maintenance.

A 2024 Berkeley Lab report estimated all U.S. data centers directly consumed about 17.4 billion gallons in 2023, projecting hyperscale facilities alone could reach 16 billion to 33 billion gallons annually by 2028. Annual totals demonstrate scale but don't reveal the infrastructure burden.

Cooling choices shift costs between water and power grids

Evaporative systems efficiently remove heat but consume water. Dry cooling and air-cooled chillers reduce direct water use while typically increasing electricity demand. The regional power mix matters—some electricity generation methods consume far more water than others, affecting the indirect footprint.

Closed-loop systems recirculate water internally but still require a final heat rejection method, which may use air, evaporation, or both. These decisions are locked in during site selection and design, making procurement agreements more consequential than operational adjustments.

Four numbers belong in every data center agreement

Executives and local officials should require peak daily water demand under extreme heat conditions, identification of water sources during drought restrictions, separation of direct consumption from electricity-related water use, and a contractual commitment to fund new infrastructure capacity.

The UC Riverside and Caltech researchers recommend that developers help finance verifiable water system improvements so expansion costs don't fall entirely on local ratepayers. Water utilities should consider large-user tariffs or capacity charges that assign dedicated infrastructure costs to the customers creating the demand.

These details were first reported by Robert J. Szczerba in Forbes, drawing on the UC Riverside and Caltech research.