The Physical Footprint of Centralized AI: Energy, Water, and Land

The debate about AI's resource cost is usually conducted in adjectives. This page is our attempt to conduct it in numbers, drawn from federal laboratory reports, congressional research, and peer-reviewed sources. We maintain this page as a living document.

Energy

The authoritative source is the Lawrence Berkeley National Laboratory's 2024 United States Data Center Energy Usage Report, produced for the Department of Energy at Congress's request. Its findings:

• U.S. data centers consumed roughly 176 terawatt-hours in 2023 — about 4.4% of all U.S. electricity — up from 58 TWh in 2014.
• By 2028, consumption is projected to reach 325–580 TWh, or 6.7% to 12% of total U.S. electricity. Load growth tripled over the past decade and is projected to double or triple again in four years.
• The growth rate itself is accelerating: roughly 7% annually from 2014–2018, 18% from 2018–2023, and a projected 13–27% annually through 2028 — driven overwhelmingly by accelerated AI servers.
• The concentration is extreme in places: in Virginia, data centers already consume more than one in four kilowatt-hours of the state's electricity. Over 700 new facilities are under construction across 38 states.

Notably, LBNL itself cautions that these figures are estimates constrained by a lack of transparency in the sector — its 2016 projections failed to anticipate AI server growth entirely. The uncertainty cuts both ways, and it is one reason independent measurement (our EQAT program) matters.

Water

Data centers consume water two ways: directly, through evaporative cooling (up to 85% of withdrawn water evaporates and does not return to the supply), and indirectly, through the water embedded in electricity generation and chip manufacturing.

• Direct U.S. data center water consumption was roughly 17.5 billion gallons in 2023 (LBNL), projected to reach 38–73 billion gallons by 2028.
• The indirect footprint dwarfs the direct one: an estimated 211 billion gallons in 2023 through electricity generation alone — roughly 1.2 gallons consumed per kWh nationally.
• Close to home: a Houston Advanced Research Center / University of Houston study projects Texas data centers will use 49 billion gallons in 2025, rising to as much as 399 billion gallons by 2030 — a drawdown equivalent to lowering Lake Mead by more than 16 feet in a year.
• A single large facility can consume up to 1.8 billion gallons annually — the water use of a town of 40,000–50,000 people — often drawn from the same municipal systems supplying homes.

For perspective and honesty: total data center water use remains small next to agriculture's roughly 26 trillion gallons annually. The concern is not the national total; it is the localization — gigawatt campuses siting in water-stressed regions, and growth rates that outpace municipal planning.

Land

The land story follows the energy story. Gigawatt-class AI campuses are measured in hundreds of acres, and their siting increasingly collides with residential growth, farmland, and grid capacity. With 700+ facilities under construction across 38 states, land-use disputes have become one of the primary sources of public opposition to AI infrastructure — a political constraint that arrives before the physical ones.

Why this motivates our research

Every projection above assumes AI inference must happen in a building built for it. Our research program asks the demand-side question the supply-side debate skips: how much of the interactive AI workload could run, instead, on the efficient hardware Americans already own — at 15–60 watts, with no new land, no cooling towers, and no water? Nobody currently knows, because nobody can measure it honestly. That measurement is our first flagship project.