> CO2 (PPAs/RECs)
- We currently use location-based grid factors (national/regional) and do not net out PPAs/RECs. That is the conservative choice for a public baseline.
- If a workload is known to be contract-matched (hourly/locational), we can apply a market-based view; I plan to expose a toggle (location- vs market-based) so both views are visible.
> Water (WUE & power-generation water)
- DC_water_per_kWh (WUE): when operators publish site/region values we use them. Otherwise we assign a cooling class (evaporative / closed-loop / seawater / air-only) and take a central value from published ranges. That gives order-of-magnitude accuracy without claiming site precision.
- PowerGen_water_intensity: technology-specific consumption factors (not withdrawals) by fuel/tech (gas, coal, nuclear, hydro, etc.), weighted by the grid mix of the region when it’s known; otherwise a conservative aggregate. Hydropower is treated as low consumption, high withdrawal.
> Electricity (IT load, utilization, PUE)
- IT_load / Training: bottom-up from reported compute for frontier runs + known fleet sizes; extrapolated to mid-scale using public training reports.
- IT_load / Inference: top-down from usage volumes (requests/tokens/images) × energy per unit by model class, calibrated from published perf/W measurements and vendor/benchmark data. We don’t simply sum GPU TDP; we use perf/W + utilization.
- utilization: ranges by workload class; we take a conservative central value (higher for sustained training, lower/peaky for inference). These are sensitivity levers and shown in the methodology.
- PUE: operator/region-specific when disclosed; otherwise we apply a conservative default for hyperscale vs. generic DCs (kept distinct). PUE is another sensitivity knob we surface.