live map · auto-refresh 5m
Grid carbon-intensity where AI compute runs.
Grid regions across North America, Europe and Asia-Pacific that host the major LLM providers' inference workloads. Click any card for the 72-hour forecast and a cost-and-carbon best-window finder.
how it works
Methodology and source data
grid feeds
Multi-source, free-public-data first: UK National Grid ESO (GB), US EIA (US ISOs), ENTSO-E Transparency Platform (EU). Electricity Maps as universal fallback when a key is set; deterministic mock otherwise.
scoring
The map cards show a single number per region using a flat typical-task estimate (~0.0015 kWh, PUE 1.15). Actual ebb-ai dispatches (v0.10+) use per-model coefficients from Patterson 2021, Luccioni 2024, and the Hugging Face AI Energy Score — e.g. claude-sonnet-4 ≈ 3 Wh/call (500 in + 500 out), claude-opus ≈ 9 Wh, claude-haiku ≈ 0.9 Wh, llama-3.1-8b ≈ 0.6 Wh. The cleanest hour inside the deadline still wins; now grams scale with which model the agent actually runs. See /docs → energy.
fallback
When no API key is configured, the dashboard serves a deterministic UTC-aligned synthetic curve so demos and CI runs are reproducible. Bands are identical to the production thresholds.