AI training is wasteful
and expensive.

We fix both.

The Problem

The Status Quo is Broken

GPU training is opaque

Developers lack visibility into the underlying energy source and real-time costs of their compute clusters.

It's an always-on paradigm

Jobs run immediately regardless of grid stress, punishing cost-sensitive users who can tolerate scheduling flexibility.

Ignores the energy grid

Massive compute loads run during peak fossil-fuel hours while gigawatts of clean energy are curtailed elsewhere.

How We Built the Forecasts

Data Pipeline & Modelling

DATA SOURCES INGEST & BUILD ML MODELS OUTPUT ENERGY APIS — NATIONAL GRID OPERATORS FRANCE 🇫🇷 services-rte Generation mix Load · Demand UK 🇬🇧 bmrs.elexon Generation mix Load · Demand FINLAND 🇫🇮 Fingrid API Generation mix Load · Demand ITALY 🇮🇹 dati.terna Generation mix Load · Demand WEATHER API 🌤️ Open-Weather Wind speed · Irradiance Temperature · Cloud cover DATASET SPECS Granularity Hourly History ~10 years Features Energy + Weather Target surplus (MW) TRANSFORM Featurize lag features (1–24h) · rolling stats (mean, std) · calendar signals (hour, dow, month) · weather interaction terms · target: surplus = generation − load ENERGY GAP FORECASTING MODELS — ONE PER COUNTRY · XGBoost · 5-fold TimeSeriesSplit CV FRANCE 🇫🇷 MAE h+1 811 MW R² h+1 0.883 MAE h+12 2 410 MW R² h+12 0.131 R² decay h+1→h+12 0.883 0.131 UK 🇬🇧 MAE h+1 438 MW R² h+1 0.855 MAE h+12 1 704 MW R² h+12 0.111 R² decay h+1→h+12 0.855 0.111 FINLAND 🇫🇮 MAE h+1 202 MW R² h+1 0.934 MAE h+12 687 MW R² h+12 0.259 R² decay h+1→h+12 0.934 0.259 ITALY 🇮🇹 MAE h+1 557 MW R² h+1 0.775 MAE h+12 1 246 MW R² h+12 0.065 R² decay h+1→h+12 0.775 0.065 FORECAST OUTPUT Energy Gap Forecast → 1h · 2h · 4h · 6h · 12h horizons Router picks country with largest predicted surplus window for incoming job ROUTEABLE 95% of the week Best model: Finland 🇫🇮 MAE h+1 202 MW R² h+1 0.934

XGBoost · 5-fold TimeSeriesSplit CV · holdout week 2026-02-09→16 · Finland best model: MAE h+1 = 202 MW, R² = 0.934

The Solution

Intelligent, Carbon-Aware Routing

1

Submit

Users submit a GitHub link + scheduling preferences (max wait time).

2

Forecast

ML models forecast grid gaps across UK, France, Finland, and Italy.

3

Route & Monitor

Job routed to the cheapest green window with injected carbon monitoring.

4

Deploy

Automatically deployed and executed via AWS, Azure, or GCP infrastructure.

Under the Hood

Backend Architecture — 9-Step Flow

■ USER/INFRA ■ REDIS QUEUES ■ CLAUDE CODE ■ ML MODELS ■ CLOUD
INPUT LAYER JOB QUEUE CODE LAYER ENERGY FORECAST DEPLOY LAYER METRICS LAYER USER 👤 DEVELOPER GitHub link Compute prefs Country allow-list submit job 1 JOB QUEUE ⚡ REDIS Q-1 Job enqueue Priority ordering Deduplication trigger 2 AI CODE AGENT 🤖 CLAUDE CODE Restructure code Inject carbon hooks Estimate runtime 3 enriched job JOB QUEUE ⚡ REDIS Q-2 Enriched job payload User prefs attached Country filters 4 dispatch INFRA 🚂 RAILWAY Trigger ML models Orchestrate runs Route by country 5 invoke models ENERGY GAP FORECASTING MODELS — RAILWAY HOSTED MODEL 🇬🇧 UK MODEL 12h horizon Gap forecast (MW) MODEL 🇫🇷 FRANCE MODEL 12h horizon Gap forecast (MW) MODEL 🇫🇮 FINLAND MODEL 12h horizon Gap forecast (MW) MODEL 🇮🇹 ITALY MODEL 12h horizon Gap forecast (MW) DATABASE 🗄️ ENERGY DB ENTSO-E historical generation data DATABASE 🌤️ WEATHER DB Open-Meteo API wind, temp, cloud 6 gap + runtime est. CLAUDE CODE — ESTIMATOR 🧠 HW SELECTOR Runtime estimate GPU/CPU selection Spot vs reserved 7 package BUILD 🐳 DOCKERIZE Build container Carbon code bundled HW config embedded 8 deploy to CLOUD ☁️ AZURE AKS / Spot VMs EU regions CLOUD 🟠 AWS EKS / Spot Fleet EU regions CLOUD 🔵 GCP GKE / Preemptible EU regions — OR — — OR — 9 📊 CARBON CODE METRICS kWh consumed · carbon saved · surplus captured · training efficiency ↑ Claude Code estimates runtime from code analysis Deploy window = forecasted energy surplus hours

Market Size & Target

Capitalising on Flexible Compute

TAM
$8–12B
Flexible GPU Training
SAM
$2.5–4.5B
Non-Hyperscaler Buyers
SOM
$3–20M
Capture in 3–5 Years

Primary Users: Research Labs, Bootstrapped AI Startups, Indie Developers.

Traction & Metrics

Proven Models, Tangible Impact

95%

Of the week features actionable routing windows.

99%

Lower carbon intensity vs EU grid baseline during surplus windows.

2.2 TWh

France renewable surplus identifiable per week.

0.86 R²

Consistent one-hour-ahead accuracy across major European power markets.

Significantly lowers global footprint: Net CO₂ negative vs EU grid baseline.

Cheaper compute.
Cleaner grid.
No config.
Contact us to join the private beta / seed round.
Click or Arrow Keys