Definition
CFAI measures the accuracy of AI-generated demand forecasts at the 12–18 month planning horizon by comparing predicted power, cooling, space, and network demand against actual consumption at the forecast maturity date.
The Measurement Gap
Traditional capacity planning uses simple linear extrapolation of historical growth trends, a method that worked adequately in stable, predictable colocation environments but fails when AI workloads introduce volatile, event-driven demand spikes. Current utilization percentage and trend lines do not capture the step-function nature of large AI deployment events or the geographic concentration risk from hyperscaler customers. CRM pipeline data sits in a separate system from power and cooling actuals, making integrated demand forecasting manual and error-prone. No standard KPI tracks how well AI-generated demand models perform at the 12–18 month horizon that capital procurement requires, leaving procurement decisions backed by unvalidated model outputs.
Decision Supported:
CFAI determines whether AI demand forecasts are accurate enough at the capital planning horizon to justify procurement commitments, or whether forecast error is large enough to require additional margin, staged procurement, or alternative contracting structures.
Formula
Plain language: For each forecast that has reached its maturity date (the point in time when the forecast was supposed to apply) compare what the AI predicted to what actually happened. Average those prediction errors across Q comparison points and subtract from 1.0 to get the accuracy index. Also compute Forecast Bias separately to detect whether the system systematically over- or under-predicts.
Notation:
CFAI = 1 − (1/Q) × Σ_q |Forecast_q − Actual_q| / Actual_q
Companion metric: FB = (1/Q) × Σ_q (Forecast_q − Actual_q) / Actual_q
Variables:
| Variable | Definition | Unit |
|---|---|---|
| Forecast_q | AI-generated demand forecast for comparison point q, recorded at forecast creation date with a specified horizon | kW, kBTU/hr, cabinet-count, or Gbps depending on resource type |
| Actual_q | Actual measured demand at comparison point q at the forecast maturity date | Same unit as Forecast_q |
| Q | Total number of forecast-to-actual comparison points with matching timestamps | Count |
| q | Index running from 1 to Q | n/a |
| FB | Forecast Bias: positive = systematic over-prediction, negative = systematic under-prediction | Decimal |
Track CFAI and FB separately for power demand, cooling demand, space demand, and network demand. Positive FB (over-prediction) wastes capital; negative FB (under-prediction) risks capacity shortfall. Both are failures, but in different directions.
Proposed formula, pending validation pilots.