How Much Water Do AI Data Centers Really Use?

A high-profile correction to Karen Hao’s book — acknowledging a 1,000x unit error in an estimate of a Google data center’s water demand near Santiago — has refocused attention on how data centers consume water and the policy, reputational and operational risks of rapid AI infrastructure expansion. Specialists stress that onsite water use (mainly for cooling) varies greatly by climate, technology and whether potable or recycled water is used, that including offsite water tied to power generation is a contested accounting choice, and that disclosure is often opaque (Google’s thirstiest U.S. facility used roughly 2.7 million gallons/day in 2024), leaving room for local opposition and regulatory action as in Chile where a project was paused. For investors, the practical implication is to underwrite AI and data-center projects with granular, location-specific water-stress analysis, clear assessments of cooling/energy trade-offs, and vigilance about disclosure gaps and community/policy risk.

Analysis

A high-profile factual correction by Karen Hao — acknowledging a 1,000x unit error in her book’s estimate of a Google data center’s water demand near Santiago — has sharpened scrutiny on data-center water use and the reputational, permitting and litigation risk that can follow. The correction and ensuing debate (including Andy Masley’s Substack, a 230-group letter to Congress, and industry pushback from the AI Infrastructure Coalition) illustrate that public sentiment and policy attention can move quickly around perceived environmental harms.

Technical experts quoted in the article emphasize that onsite water use (primarily cooling) varies materially by climate, design and source of supply, that companies like Amazon, Meta and Apple increasingly use treated municipal wastewater, and that electricity-for-water trade-offs and cooling-chemical risks (PFAS) change both operating cost and ESG profiles. Discrepancies in accounting — e.g., whether to include offsite water from power generation or how to interpret Sam Altman’s “one fifteenth of a teaspoon” figure — mean headline metrics can be misleading without standardized disclosure.

Concrete local risks are evident: Google paused a Chile project after a court order over aquifer impacts, Google’s thirstiest U.S. site used ~2.7 million gallons/day in 2024, and growth in arid regions such as Arizona raises community opposition and regulatory scrutiny. Investors therefore face exposure to site-specific water stress, opaque NDAs and potential permitting delays; these warrant project-level due diligence and incorporation of water- and policy-related contingencies into valuations.

Key Decisions for Investors

• Re-evaluate near-term exposure to companies with opaque water disclosures (notably projects tied to GOOGL/GOOG) and factor permitting, litigation and reputational risk into short-term downside scenarios
• Require project-level water-stress analysis, cooling-technology assessment and source-of-water disclosure before underwriting or expanding data-center investments
• Prefer/overweight operators that disclose water metrics and use recycled or non-potable sources (as cited for AMZN, META, AAPL), and discount or hedge positions in assets sited in high-stress basins
• Monitor local regulatory actions, court cases (e.g., Chile and Oregon precedent), major NGO mobilizations, and federal/state policy shifts as triggers for construction delays or increased operating costs