The Infrastructure Zoo
Regional Grids and Sustainability Baselines
The Infrastructure Zoo provides the Environmental Context for ML deployments—the carbon intensity of regional electricity grids and datacenter efficiency profiles. Every value is sourced from published government energy data and IEA reporting.
Training the same model in Quebec (hydro-dominated, ~20 gCO₂/kWh) vs. Poland (coal-dominated, ~820 gCO₂/kWh) can differ by ~41× in carbon footprint. Use the Geography Tutorial to explore these tradeoffs interactively.
Regional Electricity Grids
| Region | Carbon Intensity | Typical PUE | Primary Source |
|---|---|---|---|
| Norway (Hydro) | 10.0 gCO2/kWh | 1.06 | Hydro |
| Poland (Coal) | 820.0 gCO2/kWh | 1.58 | Coal |
| Quebec (Hydro) | 20.0 gCO2/kWh | 1.06 | Hydro |
| US Average | 429.0 gCO2/kWh | 1.12 | Mixed |
Datacenter Rack Profiles
| Rack Class | Power Density | Cooling Type |
|---|---|---|
| AI Cluster (Standard) | 70.0 kW | Liquid |
| Traditional Enterprise | 12.0 kW | Air |
How to Read the Infrastructure Zoo
Carbon Intensity: The Biggest Lever
The single most impactful decision for ML sustainability is where you train. Carbon intensity varies by ~41x across the grids above. Quebec’s hydro-dominated grid produces ~20 gCO2/kWh, while Poland’s coal-dominated grid produces ~820 gCO2/kWh. Same model, same hardware, vastly different environmental cost.
Connecting to TCO
Infrastructure specifications feed directly into both the SustainabilityModel (carbon and water footprint) and the EconomicsModel (electricity costs in TCO). Regional electricity prices vary as much as carbon intensity — a factor the TCO model accounts for.
Textbook Connection
The Sustainable AI chapter uses these grid profiles to quantify the carbon footprint of training runs. The Compute Infrastructure chapter connects PUE to total facility cost. Try the Geography Tutorial to explore these tradeoffs interactively.
Note: For carbon and water usage formulas, see the SustainabilityModel API Reference.