﻿# AI Footprint Calculator - MVP Development Guide & Specification


**Last Updated:** May 21, 2025
**Version:** 1.0 (MVP)


## 1. Project Overview


### 1.1. Purpose
The AI Footprint Calculator is a web application designed to:
* Enable users to estimate the operational environmental footprint (energy consumption, water usage, and associated carbon emissions) of their artificial intelligence (AI) activities. This covers both cloud-based services and AI models run on users' local hardware.
* Provide users with aggregated footprint totals over various practical timeframes (daily, weekly, monthly, yearly), allowing them to understand their cumulative impact.
* Empower users with actionable guidance and curated resources for understanding and purchasing credible carbon and water credits to offset their calculated footprint.
* Promote awareness and encourage more sustainable AI practices by making the environmental costs of AI usage more transparent and understandable.


### 1.2. Core Philosophy for MVP
* **Reasonable Estimates:** The calculator will provide the best possible "reasonable estimates" based on publicly available research, benchmarks, and clearly stated assumptions, acknowledging that precise footprint data for all AI systems is often scarce or proprietary.
* **User-Friendliness:** The web app will offer a simple, intuitive interface for common use cases, with optional advanced inputs for users who have more specific data or technical knowledge.
* **Transparency:** The calculator will be transparent about its methodologies, the general sources of its embedded data, key assumptions made (e.g., for PUE, WUE, hardware profiles), and the inherent uncertainties in its estimations.
* **Actionability:** The tool will go beyond mere calculation to provide users with the information and resources needed to take meaningful offsetting actions.


## 2. MVP Scope & Essential Features


### 2.1. Core Calculation Capabilities


#### 2.1.1. Cloud AI Footprint Estimation
* **Mechanism:** Based on a tiered system for different AI categories. Users can select popular AI models by name (which will map to a predefined tier) or select a generic tier directly if their specific model is not listed.
* **User Inputs:** AI task type, specific model or generic tier, volume of use (e.g., number of queries, images generated, minutes of video/audio, tokens processed), and the primary compute region (for grid carbon intensity).
* **AI Categories & Tiers for MVP:**
    * **Text Generation:**
        * Tier 1: Lightweight / Simple Q&A / Efficient Models / Analytical Tasks
        * Tier 2: General-Purpose / Moderate Complexity
        * Tier 3: Advanced / Complex Reasoning / Long-Form
    * **Image Generation:**
        * Tier 1: Standard Image Generation
        * Tier 2: High-Detail / Complex Image Generation / Upscaling
    * **Video Generation:**
        * Tier 1: Short-Form / Standard Definition / Efficient (MVP, with high uncertainty)
        * *(Tier 2: Longer-Form / HD / Complex - Fast Follow-up, not MVP)*
    * **Audio Generation:**
        * Tier 1: Basic Audio / Standard TTS (MVP, with high uncertainty)
        * *(Tier 2: Complex Music / High-Fidelity - Fast Follow-up, not MVP)*
    * **Analytical / Classification Tasks:**
        * Tier 1: Standard Analytical / Classification Tasks
* **Named Models Mapped to Tiers (Examples for MVP):**
    * OpenAI: GPT-4o, GPT-3.5 (Text Tiers); DALL-E 3 (Image Tiers); Sora (Video Tier 1, high uncertainty).
    * Anthropic: Claude 3 Sonnet, Claude 3 Haiku (Text Tiers); Claude 3 Opus (Text Tier 3).
    * Google: Gemini (general text, image, video capabilities will map to appropriate Tiers 1 using best available estimates).
    * Meta: Llama 2/3 (common sizes like 7B, 13B, 70B map to Text Tiers).
    * Mistral: Mistral 7B, Mixtral (map to Text Tiers).
    * Stable Diffusion (reference for Image Tiers).
    * Suno (Audio Tier 1, high uncertainty).
    * *Other models (DeepSeek, Qwen, Amazon Titan/Nova, Pyxa, 1minAI) will require users to select the closest generic tier for the MVP.*


#### 2.1.2. Local AI Footprint Estimation
* **Mechanism:** Based on user-selectable hardware profiles or direct user-provided power measurements.
* **MVP Estimation Levels:**
    * **Level 1 (Basic Estimation - Default):** User selects a "Hardware Category" and "Local AI Task Type." Calculator uses pre-defined average system power draw for that profile.
    * **Level 3 (Advanced Estimation - Optional):** User inputs their own measured average system power (Watts) and task duration.
* **User Inputs (Level 1):** Hardware Category, Local AI Task Type, Estimated Task Duration (or Number of Outputs which converts to duration), User's Region/Country (for local grid carbon intensity).
* **User Inputs (Level 3):** Measured Average System Power (Watts), Actual Task Duration, User's Region/Country.
* **Hardware Categories for MVP (Level 1):** High-End GPU Desktop, Mid-Range GPU Desktop, Entry GPU Desktop/CPU-only, High-Performance AI Laptop, Standard Laptop (iGPU/NPU), Apple M-Series Laptop.


#### 2.1.3. Agentic AI Platform Footprint Estimation
* **Mechanism:** User deconstructs the agent's overall task into its constituent AI calls.
    * **Orchestrator Overhead:** User estimates the number of "reasoning/planning" calls the agent makes to its orchestrating LLM (selecting a Text Generation Tier, e.g., Tier 2 or 3, and number of calls. Default: 3-5 calls to Text Gen Tier 2).
    * **Discrete Sub-Tasks:** User adds each distinct AI task performed by the agent (e.g., one image generation, two text summaries) using the standard Cloud AI or Local AI tier system.
    * The calculator sums these components. UI must clearly guide this input.


#### 2.1.4. Footprint Metrics Calculated
* Operational Energy Consumption (Wh or kWh).
* Operational Water Consumption (mL or Liters) - derived from energy and infrastructure data.
* Operational Carbon Emissions (gCO2e or kgCO2e) - derived from energy and relevant grid intensity.


### 2.2. Aggregation & Totals
* Users can define multiple "AI Activities" (recurring or project-based).
* For each activity, users specify the AI task/tier, volume per instance, and frequency (e.g., per day/week/month) or define as a "One-off Project."
* The calculator provides aggregated totals for energy, water, and carbon for user-selected timeframes: **Daily, Weekly, Monthly, and Yearly**.
* Clear visual and numerical distinction between "Recurring/Ongoing Usage" and "One-off/Project" contributions to totals.


### 2.3. Offsetting Guidance
* Direct display of equivalent carbon and water credits needed to offset calculated footprints.
* Dedicated educational section on:
    * Understanding Carbon Credits (quality criteria: additionality, permanence, leakage, verification, co-benefits).
    * Understanding Water Credits (volumetric benefit, water stress context, ecological uplift, verification, co-benefits).
    * Key Verification Standards (e.g., Gold Standard, Verra for carbon; BEF WRCs, VWBA for water).
    * Factors Influencing Credit Prices.
* Curated, neutral list of **3-5 reputable carbon credit providers** and **1-3 reputable water credit providers** for MVP, with factual info (standards, project types, links).
* A "Due Diligence Checklist" for users evaluating credit providers.
* An optional, user-selectable "Offsetting Buffer/Safety Margin" (e.g., 0-100%) to account for estimation uncertainty.


### 2.4. User Accounts & Data Persistence
* MVP will include simple user account functionality (e.g., email/password login) to:
    * Save defined recurring AI activities and user preferences (e.g., default regions).
    * Enable meaningful calculation of aggregated totals over time based on saved profiles.


### 2.5. Uncertainty Communication
* Clear text disclaimers throughout the app stating that all figures are estimates.
* Brief explanations of key sources of uncertainty (e.g., "AI model energy use varies...", "Water data is based on...").
* The user-selectable offsetting buffer directly addresses this.


### 2.6. Data Management for MVP
* All tier data, hardware profiles, PUE/WUE values, grid emission factors, and provider lists will be manually curated from prior research and embedded into the application's backend data structures (e.g., JSON files or a simple database schema). Updates will be manual for the MVP.


## 3. Core Calculation Engine & Embedded Data


### 3.1. Cloud AI Footprint Data & Tiers


**General Cloud Assumptions for Water Calculation (derived from "How Hungry is AI?" methodology):**
* Default PUE (Power Usage Effectiveness): `1.12` (representative of efficient hyperscalers like Azure/OpenAI).
* Default WUE_site (Water Usage Effectiveness, on-site cooling): `0.30 L/kWh_IT` (Azure example).
* Default WUE_source (Water Usage Effectiveness, off-site electricity generation): Varies by region (e.g., US average `3.142 L/kWh_grid`, China average `6.016 L/kWh_grid`).
* **Total Water (mL/unit) Formula:**
    `Water_direct (mL) = (Energy_IT (Wh/unit) / 1000) * WUE_site (L/kWh_IT) * 1000`
    `Water_indirect (mL) = (Energy_IT (Wh/unit) / 1000 * PUE) * WUE_source (L/kWh_grid) * 1000`
    `Total_Water = Water_direct + Water_indirect`
    *(For quick reference, an approximate factor of **0.35 - 0.40 mL total water per Wh of IT energy** is often seen for efficient US-based cloud infrastructure, but the full calculation should be used where region/provider PUE/WUE can be specified).*


---
**Category: Text Generation**


* **Tier 1: Lightweight Text / Simple Q&A / Efficient Models / Analytical Tasks**
    * Description: <500 total tokens, simple lookups, classification.
    * Examples: Llama-3.2 1B, GPT-4.1 nano (short); text classification.
    * Energy: `0.002 - 0.2 Wh` per query/task.
    * Water: `0.01 - 2.0 mL` per query/task.
    * Confidence: Energy: High (benchmarked). Water: Medium (benchmarked).


* **Tier 2: General-Purpose Text / Moderate Complexity**
    * Description: 500 - 2000 total tokens, standard chatbot use, content drafting.
    * Examples: GPT-4o (short/medium), Claude 3.7 Sonnet (short/medium). *Gemini, Qwen, Titan map here with lower confidence.*
    * Energy: `0.3 - 3.0 Wh` per query/interaction.
    * Water: `1.5 - 15 mL` per query/interaction.
    * Confidence: Energy/Water: High (benchmarked models). Low-Medium (others).


* **Tier 3: Advanced Text / Complex Reasoning / Long-Form**
    * Description: >2000 tokens (up to 100k+), in-depth analysis, code gen, reasoning models.
    * Examples: OpenAI "o3", GPT-4.5; Claude 3.7 Sonnet ET, Claude 3 Opus; DeepSeek-R1. *Gemini Advanced, large Qwen/Titan map here with lower confidence.*
    * Energy: `5 - 40+ Wh` per complex query/task.
    * Water: `20 - 300+ mL` per complex query/task.
    * Confidence: Energy/Water: Medium (benchmarked models). Low (others).


---
**Category: Image Generation**


* **Tier 1: Standard Image Generation**
    * Description: ~0.25-1MP (e.g., 512x512-1024x1024px), 20-50 diffusion steps.
    * Examples: Stable Diffusion v1.5/XL 1.0 (H100); Midjourney, DALL-E 3 (standard). *Gemini Image, Qwen Image, Titan Image map here with lower confidence.*
    * Energy: `0.1 - 1.0 Wh` per image.
    * Water: `0.05 - 0.5 mL` per image.
    * Confidence: Energy: Medium. Water: Medium.


* **Tier 2: High-Detail / Complex Image Generation / Upscaling**
    * Description: >1-4MP, >50-100 diffusion steps, complex prompts, upscaling.
    * Examples: SDXL (older hardware/high settings); professional print/art.
    * Energy: `1.0 - 12.0+ Wh` per image.
    * Water: `0.5 - 40 mL` per image.
    * Confidence: Energy: Medium-Low. Water: Medium-Low.


---
**Category: Video Generation (MVP - Tier 1 Only, High Uncertainty)**


* **Tier 1: Short-Form / Standard Definition / Efficient Video**
    * Description: <15-30s, ~480p-720p, simpler animations (e.g., Synthesia-like at ~0.0088 Wh/s, Amazon Nova Reel ~1.17 Wh/s).
    * Examples: Amazon Nova Reel, OpenAI Sora (lower-end/shorter clips), Open-Sora (derived 480p). *Pyxa, 1minAI map here.*
    * Energy: `1.0 - 5.0 Wh` per second.
    * Water: `0.35 - 2.0 mL` per second.
    * Confidence: Low-Medium.
    * *(Tier 2: Longer/HD Video - For Future Release)*


---
**Category: Audio Generation (MVP - Tier 1 Only, High Uncertainty)**


* **Tier 1: Basic Audio / Standard TTS / Simple Loops**
    * Description: Standard quality TTS, short simple instrumental loops.
    * Examples: Efficient TTS (Google Cloud TTS est. ~0.017 Wh/min), AudioLDM (~0.0012 Wh/min for simple). *Suno (simple outputs) maps here.*
    * Energy: `0.001 - 0.05 Wh` per minute.
    * Water: `0.00035 - 0.02 mL` per minute.
    * Confidence: Energy: Medium (benchmarked open models/TTS). Water: Medium.
    * *(Tier 2: Complex Music / High-Fidelity - For Future Release)*


---
**Category: Analytical / Classification Tasks**


* **Tier 1: Standard Analytical / Classification Tasks**
    * Description: Pattern identification, input classification, object detection.
    * Examples: Text classification (DistilBERT), image classification (ResNet-50).
    * Energy: `0.002 - 0.04 Wh` per query/item.
    * Water: `0.0007 - 0.016 mL` per query/item.
    * Confidence: Energy: High. Water: Medium-Low.


---
### 3.2. Local AI Footprint Data & Profiles


* **Level 1: Basic Estimation (Hardware Category & Task Type)**
    * User selects from Hardware Categories (based on "Local AI Data PDF" Table 1).
    * **Representative Average System Power Draw (Watts) during AI Task (Illustrative - to be populated from "Local AI Data PDF" Section 4):**
        * **High-End GPU Desktop (e.g., RTX 4080/4090):** LLM (13B+ Q): `350-550W`; Image (SDXL High): `300-500W`.
        * **Mid-Range GPU Desktop (e.g., RTX 4060Ti/4070):** LLM (7B-13B Q): `200-350W`; Image (SD Standard): `180-300W`.
        * **Entry GPU Desktop / CPU-only:** LLM (Small 7B Q): `100-200W` (GPU) / `80-180W` (CPU); Image (SD Light): `100-180W` (GPU).
        * **High-Performance AI Laptop:** LLM (7B-13B Q): `100-200W`; Image (SD Standard): `90-180W`.
        * **Standard Laptop (iGPU/NPU):** LLM (Small 7B Q, NPU/iGPU opt.): `20-60W`; Image (SD Light): `25-70W`.
        * **Apple M-Series Laptop:** LLM (Small-Medium Q, ANE/GPU opt.): `15-45W`; Image (SD CoreML): `20-50W`.
    * Logic: `Energy (Wh) = Representative_System_Power_Profile (W) * Task_Duration (h)`.
    * Performance rates (tok/s, img/s) from "Local AI Data PDF" (Tables 2,3,4) used to convert "Number of Outputs" to "Task_Duration".
* **Level 3: Advanced Estimation (User-Provided Measurements - Optional in MVP)**
    * User Inputs: Measured Average System Power (Watts), Actual Task Duration, Region.
    * Logic: `Energy (Wh) = User_Measured_Power (W) * Task_Duration (h)`.


### 3.3. Agentic AI Platform Footprint
* **Orchestrator LLM Calls:** User estimates N calls to a selected Text Generation Tier (Default: 3-5 calls to Text Gen Tier 2).
* **Discrete Sub-Tasks:** User adds each sub-task using standard Cloud/Local tiers.
* Total is sum of orchestrator + sub-tasks.


## 4. Data Aggregation & Presentation


### 4.1. User Input for Aggregation
* Users define "AI Activities" (e.g., "Work Emails").
* Link activity to a calculated single instance footprint (Cloud Tier or Local Profile).
* Specify frequency ("X times per [Day/Week/Month]") or mark as "One-off Project."
* User profiles save these definitions.


### 4.2. Aggregation Logic
* Calculate footprint per activity for the chosen period.
* Sum across all activities.
* Conversions: 1 Month = 4.345 Weeks (avg); 1 Year = 12 Months.
* Differentiate "Recurring" vs. "Project" totals.


### 4.3. Results Dashboard
* **Timeframe Selector:** Daily, Weekly, Monthly, Yearly.
* **Key Totals:** Energy (kWh), Water (Liters/m³), Carbon (kgCO2e/tCO2e), Equivalent Credits Needed.
* **Visualizations:** Bar/Pie charts for breakdown by activity/type.
* **Equivalencies:** Carbon (miles driven, etc. via EPA data), Energy (appliance hours), Water (showers, etc.).
* **Offsetting Buffer:** User-selectable % (0-100%) to add to credits.
* **Uncertainty Disclaimer:** Prominent text.


## 5. Offsetting Guidance (MVP Content)


### 5.1. Educational Content (Concise for MVP)
* "What are Carbon Credits?" (Additionality, permanence, verification).
* "What are Water Credits?" (Volumetric benefit, context, verification).
* Brief on key standards (Gold Standard, Verra; BEF WRCs).
* "Understanding Credit Prices" (key factors).


### 5.2. Curated Provider Lists (Neutral, Factual)
* **Carbon (3-5 examples):** CNaught, Patch, Cool Effect, Terrapass, Gold Standard Marketplace. (List name, link, standards, typical projects).
* **Water (1-3 examples):** Bonneville Environmental Foundation (BEF), Water Footprint Implementation/Act4Water. (List name, link, metrics/standards).
* Disclaimer: Informational, not endorsement for price.


### 5.3. Due Diligence
* Provide the "Due Diligence Checklist" (from "Totals & Offsetting PDF" Table 8: Third-Party Verification, Public Registry, Project Docs, Additionality, Permanence, Leakage, Co-benefits, Host Country Approval, Provider Transparency, Credit Retirement).


## 6. Web App Structure & User Flow (High-Level)


### 6.1. Key Sections/Pages
1.  **Homepage:** Intro, CTA.
2.  **Calculator Input:** Define/Manage AI Activities (Cloud/Local, Recurring/Project).
3.  **User Profile:** (Login/Register, Saved Activities).
4.  **Results Dashboard:** Aggregated totals, visualizations, credits, buffer.
5.  **Offsetting Guidance Hub:** Education, provider lists, checklist.
6.  **Methodology/About:** Assumptions, data sources (general), limitations.


### 6.2. Basic User Flow
1.  Homepage -> Start Calculation.
2.  (Optional) Login/Register or use "Typical Usage" templates.
3.  Define AI Activities (select tiers/profiles, input volume, frequency, location/hardware).
4.  View Results Dashboard (select period, see totals, apply buffer).
5.  Explore Offsetting Guidance.
6.  (If registered) Save configurations.


## 7. Style Guide


* **Overall Vision:** Minimalist, high-resolution, white-on-black, modern interpretation of vintage monochrome CRT. Professional, tech-oriented, clear, usable. Aligned with TowerIO brand.
* **Color Palette:**
    * Page Background (Absolute): `#000000` (Pure Black)
    * Main Content Area Background: `#0D0D0D` (Very Dark Grey)
    * Primary Text Color: `#FFFFFF` (White)
    * Secondary Text Color (Labels, helper text): `#B0B0B0` (Light Grey)
    * Tertiary Text Color / Subtle Borders: `#707070` (Medium Grey)
    * Focus Outline (Accessibility): `#FFFFFF` (White), `2px solid` or `2px dashed`.
* **Typography:**
    * Primary UI Font (Labels, Data, Inputs, UI Button Text): `IBM Plex Mono`, fallbacks: `Consolas, Menlo, Monaco, monospace`. Regular (400), Semibold (600) or Bold (700).
    * Informational Text Font (Longer educational reads): `IBM Plex Sans`, fallbacks: `Inter, Roboto, sans-serif`. Regular (400). Line Height: `1.7`.
    * Example Sizes: H1: `32px` (UI Font, Bold), H2: `24px` (UI Font, Bold), Body/Labels (UI): `15px` (UI Font, Regular), Body (Info): `16px` (Info Font, Regular), Button Text: `15px` (UI Font, Semibold, ALL CAPS).
* **UI Elements:**
    * **Primary Buttons:** Transparent/`#0D0D0D` bg, `#FFFFFF` text/border (`1px solid`). Padding `12px 24px`. Corner radius `3-5px`. Font: Primary UI, `15px`, Semibold, ALL CAPS.
        * **Hover/Focus:** `#FFFFFF` bg, `#0D0D0D` text (inverted). Border `1px solid #FFFFFF`.
    * **Secondary Buttons:** Transparent bg, `#B0B0B0` text, `#707070` border. Hover: `#222222` bg, `#FFFFFF` text, `#B0B0B0` border.
    * **Text Links:** `#A0A0A0` (Light Grey), Bold (700), Underlined. Hover: `#FFFFFF`.
    * **Input Fields:** `#000000` or `#0A0A0A` bg, `#FFFFFF` input text, `#707070` border/placeholder. Padding `10px 15px`. Corner radius `3-5px`. Focus: `#FFFFFF` border.
    * **Dropdown Menus:** `#0D0D0D` bg, `#FFFFFF` text. Hover/Selected item: `#222222` bg.
    * **Radio/Checkboxes:** White outlined circle/square. Selected: solid `#FFFFFF` fill/checkmark. Label: Primary UI Font, `15px`, `#B0B0B0`.
    * **Icons:** Simple, white/light grey line-art (e.g., Feather Icons).
    * **Dividers:** `1px solid #333333` (Subtle Dark Grey), use sparingly.
* **Layout & Spacing:** Minimalist, clean. Generous "blackspace." Grid-based. Max content width `~1280px-1440px`. Base spacing unit `8px`.
* **Branding:** TowerIO logo (white) in header. Charts: monochrome (white/greys on black).
* **Effects:** Minimal. No scanlines. Crisp text. Subtle CSS transitions (`0.15s ease-out`).


## 8. Post-MVP Considerations
* Adding Tier 2 for Video and Audio generation as more data becomes available.
* Implementing Local AI Level 2 estimation (specific component selection).
* Dynamic data updates for grid intensity factors (e.g., via API).
* Expanding lists of credit providers and potentially integrating with rating agencies if feasible.
* User history tracking and comparative footprint analysis over time.
* Incorporating embodied footprint of AI hardware and usage.
* More sophisticated UI for agentic task breakdown.
* "What-if" scenarios for users to see impact of changing AI habits.