Understanding AI Environmental Impact Metrics

At SustAIn, we are dedicated to quantifying the environmental footprint of AI models. Our methodology is built upon rigorous research and real-world data extrapolation, ensuring transparency and accuracy.

Why AI Has an Environmental Footprint

"Asking an AI isn't just digital — it taps into energy-hungry machines that burn power, release heat, and rely on constant cooling to keep running."

The Physical Reality of Digital AI

AI systems may seem purely digital, but they rely on massive physical infrastructure. Here's what happens when you use an AI service:

Your Query

Server Processing

Energy Consumption

Heat Generation

Cooling Needs

Key Environmental Impacts Explained

Electricity Consumption

AI models require significant computational power, drawing electricity for both processing operations and supporting infrastructure. Large language models can consume thousands of kilowatt-hours during training and inference.

Carbon Emissions

AI servers require electricity, often generated by fossil fuels. Different regions have different carbon intensities based on their energy mix (coal, natural gas, renewables).

Water Consumption

Data centers use water directly for cooling and indirectly through power generation. A single AI query can consume a small amount of water—millions of queries add up.

Factors That Influence Environmental Impact

Model Size: Larger AI models with billions of parameters require more computational resources.
Query Complexity: Complex questions or creative tasks need more processing power than simple queries.
Conversation Length: Longer conversations with many back-and-forth exchanges increase resource usage.
Data Center Location: The regional energy grid's composition significantly affects carbon emissions.
Time of Day: Energy sources often vary throughout the day (solar during daylight, etc.).

Did You Know? Training a large language model can consume as much electricity as several hundred American homes use in a year.

By measuring and optimizing these factors, we can enjoy the benefits of AI while minimizing its environmental footprint—which is exactly what our SustAIn metrics help you do.

Research & Data Sources

We base our calculations on established research papers and empirical datasets. Notably, the AI Energy Score Initiative provides invaluable insights into energy consumption trends across AI models.

Adaptive Time-Based Scaling

We now account for the duration of each prompt to more accurately reflect real-world resource consumption. Some queries take milliseconds, others minutes — this difference matters.

To smooth fluctuations, we maintain a moving baseline for duration using an Exponential Moving Average (EMA):

baseline = (1 - α) × previousBaseline + α × newDuration
α = 0.1 (adaptive but stable)

This approach provides:

One long task: Slight increase in baseline
Multiple short tasks: Gradual decrease
Burst of variation: Smooth averaging

We calculate a time-based scaling factor to adjust resource metrics:

scalingFactor = averageDuration / baseline

We bound this factor to prevent manipulation:

For shorter tasks (factor < 1): Apply 80% of base values
For typical tasks (1 ≤ factor < 10): Apply proportional scaling
For extended tasks (factor ≥ 10): Cap at 1000% of base values

Example Effect: A 6-second query with a 3-second baseline applies a 2.0 scaling factor (200% of base values), while a 120-second query might be capped at 10.0 (1000%).

Energy Consumption Scaling

To estimate real-world AI model energy consumption, we apply a parameter-based scaling approach. Since larger models incorporate optimizations, we use an efficiency-adjusted formula:

𝐸_new = 𝐸_base × (P_base / P_new)^b

Where:

E_base: Energy consumption of a reference model
P_base: Parameter count of the reference model
P_new: Parameter count of the target model
b: Efficiency scaling exponent

Carbon Emission Estimation

We determine a server's geographical location through browser APIs and geolocation services. The regional carbon intensity factor (gCO2eq/kWh) is retrieved from the ElectricityMaps API to assess emissions.

Datacenter Efficiency & PUE Factor

Power Usage Effectiveness (PUE) varies based on external conditions. Our calculations account for these fluctuations:

Peak summer demand: PUE = 1.6 (higher energy loss)
Winter off-peak hours: PUE = 1.1 (optimal efficiency)

Water Consumption Calculation

Water usage per kWh is estimated based on cooling system dependencies. Our methodology factors in regional variations in water consumption for cooling.

Query Classification

To enhance precision, queries are first categorized (e.g., text generation, image processing) using our machine learning model deployed on Google Cloud's Vertex AI.

Further Resources

For a deeper dive into our methodology, explore our repos on GitHub and consult our Documentation Repository.

Our Team

Tharun Kumarr A

GitHub LinkedIn Portfolio

Thanus Kumaar A

GitHub LinkedIn Portfolio