Behind every AI kilowatt-hour is a hidden river of evaporated water. Training a model like GPT-3 consumed 636,000 gallons—enough to fill an Olympic pool—while daily inference "drinks" a bathtub per user every week. This post reveals that data centers are massive water evaporators, with AI-driven cooling demands sometimes exceeding energy impacts. We compare water usage to office lighting, map water-rich data centers, and outline practical steps for mitigation, including liquid cooling retrofits and smarter water-rights management. By 2030, a water-light AI landscape is possible, but requires proactive action from tech leaders and policymakers today.removed