Researchers: GPT-4 generating a 100-word email uses 1,468ml of water at its Washington data center, 925ml in Arizona, 464ml in Illinois, and 235ml in Texas
AI bots generate a lot of heat, and keeping their computer servers running exacts a toll. — Roughly a quarter of Americans … Forums: Hacker News Forums: Hacker News : A bottle of water per email: the hidden environmental costs of using AI chatbots
Context & Ripple Effects
Earlier research had already put ChatGPT's water demand at an estimated 500ml for five to 50 prompts, alongside reported year-over-year increases in water use at Microsoft and Google. This report adds a more operational point: the environmental cost of AI inference can differ substantially by where the workload runs.
Later company-level estimates for an average ChatGPT query and a Gemini text prompt are far smaller, underscoring that prompt length, model, measurement method and data-center location all matter when comparing AI resource claims.
First-order effects
- The findings make data-center location an immediate variable in GPT-4's water footprint: the same 100-word task carries markedly different reported water use across the four named sites.
- AI providers and data-center operators face greater pressure to explain the local water assumptions behind efficiency and sustainability claims, rather than presenting a single model-wide figure.
Second-order effects
- Siting and workload-routing decisions gain a water constraint alongside computing capacity: operators may favor locations or cooling arrangements with lower water intensity where feasible.
- Conflicting per-query estimates—such as OpenAI's later average-query water claim and Google's median Gemini prompt estimate—raise the value of comparable, task-specific reporting for customers and policymakers.
Third-order effects
- If AI deployment keeps moving from occasional experimentation to routine business tasks, water availability and local acceptance can become practical limits on where AI infrastructure expands.
- The broader shift is toward treating AI compute as utility-like infrastructure, with environmental impacts assessed at the facility and region level rather than only through model capability.
The trend: AI's infrastructure debate is shifting from aggregate energy narratives toward location-specific accounting for the water and other local resources consumed by inference.