GreenRouter

Inspiration

Data center electricity demand is projected to more than double by 2030, to roughly 945 TWh.

Increasing AI demand causes that unprecedented growth and strain on the grid, our water resources, and clean air. Billions of LLM API calls are made every day, and the research shows many of those calls go to reasoning models for tasks a small model could handle and to datacenters in already water-scarce regions.

This is where GreenRouter intervenes.

What it does

Green AI Router is an OpenAI-compatible environmental layer that sits between an app and Azure-provided OpenAI models. it does three things:

Right-sizes the model. It classifies each prompt as simple, medium, or complex and sends it to the smallest model that can handle it.
Routes by region. For a single call, it sends the request to the region with the lowest current carbon and water impact. For planning, it returns a weighted split of $N$ calls across regions, so no single grid or watershed takes all the load.
Accounts for every call. It logs energy, CO₂, and stress-weighted water for each call and compares them against a naive baseline, ready-to-use for our users' corporate sustainability groups.

How we built it

Stack below-

Backend: FastAPI proxy with a DeBERTa-based complexity classifier, routing to Azure OpenAI deployments in West US and North Central US. Data: hourly EIA-930 grid data backfilled for a year, EIA-860/923 power plant data, and USGS National Water Availability data, all stored in TigerData (TimescaleDB) hypertables. Frontend: a React + ArcGIS dashboard with a live demo and a grid-and-impact map, plus an MCP server integration.

Challenges we ran into

AI companies and Cloud providers are obscure about their emissions, water draw, and energy data. We draw our stats from published research on them.
Dealing with gaps in grid and water data
Not making the problem worse by using a large LLM to decide which model [simple or large] to call would cancel out the savings. A small classifier is used.
Avoiding herding. Sending all traffic to the "greenest" region simply moves the strain somewhere else. We separated per-call winner-take-all routing from distribution-based planning.

Accomplishments that we're proud of

What we learned

There is a real trade-off between carbon and water, and the datacenter that is the "best" often switches by time of day. Important to know so your API call can be sent to the right place.

What's next for GreenRouter

Scale to other Azure regions in the US and internationally
Expand to AWS and GCP from just Azure
Include Anthropic models and other OpenAI models (limited by student Azure subscription)