Does

Every text query sent to a modern large language model evaporates a small but measurable quantity of fresh water, directly through data-centre cooling and indirectly through power generation. At the scale of a popular consumer API, this aggregates to volumes equivalent to the household water consumption of a small town. We ask a narrow, practical question: by how much can a user reduce that footprint simply by changing how the prompt is written? We run a controlled experiment across four open-weight models served by OpenRouter, twenty standardised prompts spanning factual recall, reasoning, summarisation, and coding, and three prompting conditions, for a total of 266 controlled inferences. We separate direct (on-site cooling) from indirect (grid electricity) water, an accounting distinction the academic literature treats as essential [2, 3] but corporate sustainability disclosures routinely collapse [6]. On a fully sampled 20-billion-parameter model, prompts that ask for shorter answers reduce output tokens by 62-65% and water by 54-56% relative to an unconstrained baseline, with no measurable quality loss across all four task categories. Two cross-model findings sharpen the picture. On a 1.2-billion-parameter edge model, the same instruction reduces tokens but causes a quality cliff under one phrasing and not the other. On a 30-billion-parameter reasoning-tuned MoE model, an instruction to “answer in under 50 words” increases output tokens by 24%, the model interprets the instruction as a request for more careful reasoning rather than for shorter output. Prompt design is a real, immediately deployable user-side lever for AI sustainability; it is also an architecturally fragile one whose effect must be characterised per model class rather than assumed.