Behind EcoPrompt: How I modeled the hidden physical cost of AI prompts (Zero Telemetry)

A developer built EcoPrompt, an open-source Chrome extension that estimates the water, energy, and carbon footprint of individual AI prompts in real time on interfaces like ChatGPT, Gemini, and Claude. The tool draws on peer-reviewed research including UC Riverside's "Making AI Less Thirsty" and Hugging Face's "Power Hungry Processing," and assigns per-prompt estimates by model tier, from roughly 3–5 mL of water for lightweight models to 250–400 mL for image generation. It runs entirely client-side with zero telemetry, storing history in chrome.storage.local under Manifest V3.

GenAI web interfaces ChatGPT, Gemini, Claude are pristine, minimal, and fast. But behind every single prompt lies a massive array of GPU clusters consuming real electricity, requiring evaporative cooling water, and generating carbon emissions. I created EcoPrompt , an open-source Chrome Extension, to make this invisible footprint transparent to users in real time. Before diving into the methodology, here is a quick 30-second promo teaser giving an overview of the concept: Watch on YouTube: EcoPrompt Concept Teaser https://youtu.be/yK ZUyT9lQM ❓ Why I Built This Most environmental discussions around AI fall into two extremes: complete ignorance of resource consumption or overwhelming guilt trips about using modern tools. I wanted a middle ground: unobtrusive, guilt-free awareness . 1. Physical Transparency: When you send a prompt, you should know if that specific request cost a teaspoon of water or a full glass. 2. Behavioral Nudging: Seeing real-time metrics encourages better prompting habits — batching queries, choosing lightweight models Flash/Haiku for simple tasks, and saving heavy reasoning models o1/Opus or image generation for when they are truly needed. 3. Financial Alignment: Bridging the gap between free web interfaces and underlying API token costs. 🔬 What Is EcoPrompt Based On? The metrics aren't arbitrary guesses. They rely on peer-reviewed research and industry sustainability reports: - Water Consumption Scope 1 & 2 : Based on research from UC Riverside "Making AI Less Thirsty" , Li et al. , combining direct evaporative cooling at the datacenter with indirect water used for electricity generation. - Energy & Carbon Footprint: Calibrated using research from Hugging Face "Power Hungry Processing" , Luccioni et al. alongside regional grid carbon intensity averages and provider datacenter efficiency metrics PUE . - Tiered Model Granularity: Different model architectures are categorized into distinct compute tiers: - Lightweight GPT-4o mini, Gemini Flash, Claude Haiku : ~3–5 mL water / ~0.0005 kWh - Standard GPT-4o, Gemini Pro, Claude Sonnet : ~20–30 mL water / ~0.003 kWh - Reasoning / Extended Thinking o1/o3-series, Claude Opus/Thinking : ~120–250 mL water / ~0.02 kWh - Image Generation DALL-E 3, Imagen 3 : ~250–400 mL water / ~0.035 kWh 🛡️ Privacy & Architecture Choice Building an extension that interacts with pages like chatgpt.com or claude.ai carries a heavy privacy responsibility. To ensure 100% user privacy : - Zero Telemetry: No tracking, no external analytics server, no remote APIs. - Client-Side Only: Character length and model types are parsed in temporary browser memory. - Local Persistence: All history and trends stay in your browser's chrome.storage.local . - Manifest V3 & Vanilla JS: Lightweight footprint with no external npm dependencies. 🔗 Try It & Explore EcoPrompt is completely free and open-source under CC BY-NC-SA 4.0. How do you approach tracking or optimizing your daily AI usage? Let's discuss in the comments