We will be undergoing planned maintenance on January 16th, 2026 at 1:00pm UTC. Please make sure to save your work.

K1

As data centers move to orbit, developers will soon choose where to run AI jobs: on Earth or in space. K1 helps to decide how much to run in orbital data centers vs terrestrial ones.

Comment

Inspiration

AI training evaporates 1.8 liters of water per kWh. Grid carbon varies 24x globally: Sweden's hydro at 13 gCO₂/kWh whereas Germany's coal at 320 gCO₂/kWh. Same GPU-hour costs 10x more in the wrong location. For eg, Starcloud-1 launched, on Nov 3, 2025, first NVIDIA H100 in orbit, confirming energy is 22x cheaper in space. Zero water.

K1 answers: Where and when should AI run?

What it does

K1 routes AI workloads across Earth + orbital data centers using live APIs.

Input: GPU-hours, dataset size, priority.

K1 calls UK National Grid, Electricity Maps (200+ zones), and Azure for real-time carbon and cost data.

Output: Optimal routing with 41% less carbon, 53% less water, 47% lower cost. For flexible jobs, recommends time-shifting: "Wait 4 hours for solar peak → Save 156kg more CO₂."

How we built it

Stack: Next.js 14, TypeScript, react-globe.gl (3D Earth), Tailwind

APIs: UK National Grid (free), Electricity Maps (500/day free tier), Azure Retail Prices

Algorithm: Weighted scoring (40% carbon, 30% cost, 30% latency) → 70% to best region, 30% to rest.

Data centers: Quebec (2 gCO₂/kWh), UK (LIVE), California (LIVE), Germany (LIVE), Sweden (LIVE), Orbital LEO-1 (8 gCO₂/kWh), Orbital LEO-2 (10 gCO₂/kWh).

Challenges we ran into

Making it real: Most "green AI" projects are mockups. Solution: Starcloud-1 launched Nov 2025 (operational hardware). Live data reliability: Graceful fallbacks to cached data (30 min old) with clear indicators.

Accomplishments that we're proud of

First orbital routing demo using real Starcloud-1 specs (launched 3 weeks ago, in orbit now) Live APIs proving 24x variance on screen in real-time Clean UI with 3D globe showing exactly where jobs run and why

Impact at scale: If K1 routed 1% of global AI training → 150,000 tons CO₂ saved, 6.5 billion liters water conserved, $1.8B cost reduction annually

What we learned

Location is everything: Same GPU-hour produces 24x different emissions depending on grid. Most ML frameworks don't even ask "where should this run?"

Water is the hidden crisis: AI labs obsess over carbon offsets. Meanwhile California data centers drain aquifers in drought zones. Orbital radiative cooling uses zero freshwater, that should be the headline.

Space is cheaper: Solar in orbit generates 5x more power (no atmosphere, no night). $0.002/kWh vs $0.045/kWh on Earth grid.

What's next for K1

Expand to 10+ regions (ERCOT, CAISO, NYISO), historical carbon charts, export to Terraform/Kubernetes for real deployments.

Starcloud-2 telemetry integration when it launches, batch job scheduling, ML carbon forecasting (48-hour ahead predictions).

Built With

Share this project:

Updates