Hidden Current estimates the energy and water used by each Claude response based on the length of the text it generates.
Character count is converted to an estimated token count - the unit AI models use internally to process text. From there, published energy figures for modern GPU hardware are applied to arrive at a watt-hour estimate, then converted to water using the industry-standard Water Usage Effectiveness (WUE) figure of 1.9 litres per kWh.
Energy (Wh) = (input tokens x 0.0001) + (output tokens x 0.0003) + 0.002
Water (mL) = Energy x 1,900
These are estimates. AI companies don't publish per-query consumption data, and the true figure varies depending on which data centre handles a request, the hardware it runs on, and local conditions like climate and cooling systems. The numbers are based on the best publicly available research, but should be read as informed approximations rather than precise measurements.
The goal isn't precision - it's visibility.