AQUASHIFT

Logo

Inspiration

AI inference is consuming water at a scale nobody is talking about. ChatGPT alone processes 2.5 billion prompts a day — and every one of them generates heat that data centers cool with water. Evaporative cooling in Phoenix at 42°C uses 3× the water per GPU-hour as a data center in Dublin at 13°C. That gap exists right now, every hour, and nobody is acting on it — because no system exists to close it autonomously. We built AquaShift to be that system.

What it does

AquaShift is a weather-aware workload intelligence layer for data centers. You give it a list of data centers and their job queues. It runs a three-agent pipeline continuously:

Thermal Scout pulls live weather forecasts for every location and computes a Water Stress Index (WSI) — a single score combining ambient temperature, wet-bulb temperature, humidity, and drought status. Shift Planner identifies which jobs are shiftable (batch inference, model training) versus fixed (real-time API calls), and calculates exactly how many liters of water per hour a migration would save. Water Broker executes the shift — pausing the hot-region queue, spinning up capacity in the cooler region, and logging every decision with a full audit trail. The result: a live dashboard showing jobs moving in real time, a ticking water savings counter, and a traceable record of every autonomous decision made.

How we built it

Weather layer: Open-Meteo API for hyperlocal hourly forecasts (wet-bulb + ambient temp per data center lat/long), architected to swap in Jua AI for 24×/day satellite-updated forecasts Agent pipeline: Three pure TypeScript agents (Thermal Scout, Shift Planner, Water Broker) orchestrated server-side with full audit logging Action layer: Composio wired to Slack — every shift decision dispatches a real trigger action workflow run, creating an audit trail over a Slack Channel. Dashboard: Live SSE-streamed UI showing data center stress levels, job migration, and gallons saved in real time

Challenges we ran into

We had to model wet-bulb temperature from raw humidity and ambient data since most weather APIs don't expose it directly.

Accomplishments that we're proud of

The water savings math is real. We're not estimating — every gallon shown on the dashboard is calculated from actual WUE (Water Usage Effectiveness) deltas between locations, live job power consumption, and measured temperature differentials. When the demo shifts jobs from Phoenix to Dublin, the 18,000 gallons/hour figure is a derived physical quantity, and not a random number.

What we learned

Scheduling is the most underrated lever in sustainable AI infrastructure. The hardware is fixed, the cooling physics are fixed — but where and when a job runs is almost entirely discretionary for batch workloads. We learned that wet-bulb temperature, not ambient temperature, is the right signal for evaporative cooling efficiency, and that the WUE gap between a hot-desert data center and a temperate one on a summer afternoon can exceed 3× — making intelligent routing the highest-ROI intervention available today without touching a single piece of hardware.

What's next for AQUASHIFT

Jua AI integration: swap Open-Meteo for 24×/day satellite-updated hyperlocal forecasts with 6-hour wet-bulb predictions Grid carbon intensity overlay: shift jobs not just to where water is cheap, but where the grid is greenest Real queue integration: connect to actual job schedulers (SLURM, Kubernetes, Ray) instead of simulated workloads Hyperscaler API: package AquaShift as a scheduling middleware layer that cloud providers can drop in front of their existing regional routing logic Prove it at scale for various data centres.

** Demo link video in additional links below **