Empire of AI is wildly misleading on AI water use (andymasley.substack.com)

A reader and analyst of Karen Hao’s Empire of AI found multiple significant errors in the book’s treatment of data-center water use, most notably conflating water “withdrawal” with water “consumption” and overstating potable use by orders of magnitude. Hao cites a UC Riverside projection of 4.2–6.6 billion cubic meters of water withdrawal in 2027 (1.1–1.7 trillion gallons) but presents it as consumptive, drinkable use. The study actually estimates consumption at 0.38–0.60 bcm (100–158 billion gallons), and only about 15% of that consumption happens on-site in data centers. The study’s on-site estimate is 150–280 billion liters (≈40–74 billion gallons), and when accounting for the share that’s potable (~80% in some operators), the plausible potable figure is ~32–59 billion gallons — roughly 3% of the number Hao implies. In one stark case the book claims a Google data center would use “>1,000×” the annual water of a municipality of 88,000 people; the critic shows the real figure is ~0.22× the city’s use and only ~3% of the municipal system — an error of about 4,500×. This matters because Hao’s book has helped shape public and policy conversations about AI’s environmental impact. The critique underscores two technical takeaways for the AI/ML community and reporters: always distinguish withdrawal vs. consumptive use, and separate on-site potable use from off-site non-consumptive withdrawals (e.g., cooling water from power plants that’s largely returned). Methodological choices — how hydropower evaporation is treated, or whether returned water is counted as “use” — can dramatically change conclusions, so accurate, contextualized metrics are essential for sound debate and policy.