Your AI Prompts Aren’t Drinking Bottled Water – But Data Centers Are Draining Local Basins

US data centers consumed 66 billion liters in 2023, with drought-zone siting and cooling design driving the sharpest local conflicts

Nikshep Myle Avatar
Nikshep Myle Avatar

By

Image: Wikimedia Commons

Key Takeaways

Key Takeaways

  • Viral 500ml-per-prompt claim is outdated; Google reports Gemini uses just 0.26ml.
  • U.S. data centers consumed 66 billion liters directly in 2023, with far more projected by 2030.
  • Strategic siting and closed-loop cooling designs determine local water stress, not per-prompt usage.

A Phoenix water board receives a permit application for a new data center in an arid basin. That scenario captures the infrastructure reality behind a viral claim you’ve probably encountered: “Every ChatGPT query drinks 500 milliliters of water.” Alarming. Also now considered outdated by researchers, as model efficiency has improved significantly.

Google reported in 2025 that a median Gemini text prompt uses about 0.26 milliliters — roughly five drops, covering serving infrastructure and on-site cooling. So your chatbot habit isn’t the issue. What happens when you multiply that efficiency across millions of queries, build the servers to handle them, and plant those servers in a drought zone — that’s where it gets real.

Scale, Not Sips

The aggregate numbers tell a very different story than the per-prompt panic.

U.S. data centers used 66 billion liters of water directly in 2023. Electricity generation to power them consumed roughly 800 billion liters more, according to Lawrence Berkeley National Laboratory data. Those are not rounding errors.

  • Direct U.S. data-center water use, 2023: ~66 billion liters (LBNL)
  • Indirect use via electricity generation: ~800 billion liters (LBNL)
  • Projected U.S. annual consumption by 2030: 731 billion to 1,125 billion liters (Reuters, 2025)
  • Agriculture still dwarfs these totals nationally — but like one Costco opening rewriting a suburb’s traffic patterns, a single data center can overwhelm a local watershed

Strategic placement isn’t a nice-to-have. It’s the whole argument.

The Real Fight Is Local

Zoom out to national averages and you miss where the pressure actually lands.

In Scottsdale last year, residents packed a city council meeting over a proposed data campus near an already-stressed municipal well. That kind of conflict is becoming familiar. Reuters 2025 reporting flagged the following as places where rapid data-center expansion is already straining local supply:

  • Arizona
  • New Mexico
  • Southern California
  • Sydney, Australia

The buildout is accelerating. Cooling design shapes the outcome as much as geography does. Newer liquid-cooling systems retain water in closed loops, reducing direct losses compared with older evaporative designs, though some hybrid systems still draw on supplemental cooling during heat spikes. That engineering choice, made in a boardroom, can determine whether a local reservoir feels the difference by summer.

Amazon has pledged water-positive status by 2030. Google aims to replenish 120 percent of the freshwater its data centers consume through stewardship projects by the same deadline. Pledges and projections are not pipelines.

What to Actually Watch

The real story was never your query count. Watch the permitting fights. Watch whether new AI campuses pair with renewables and closed-loop cooling. Watch whether municipal water contracts become the next flashpoint in the buildout debate. The problem is manageable — until geography, drought, and unchecked expansion decide otherwise.

Share this

At Gadget Review, our guides, reviews, and news are driven by thorough human expertise and use our Trust Rating system and the True Score. AI assists in refining our editorial process, ensuring that every article is engaging, clear and succinct. See how we write our content here →