Definition
Measures the degree to which AI-optimized workload placement creates power distribution imbalance across PDU circuits, UPS modules, or bus segments compared to a non-optimized baseline.
The Measurement Gap
Workload orchestration systems optimize placement for thermal and performance objectives, but their decisions have power distribution consequences that are not part of the optimization objective function. When AI places compute-intensive workloads in adjacent rack positions to capture thermal efficiency, it may simultaneously concentrate load on a subset of PDU circuits or UPS modules. A single-circuit failure that would normally affect 5% of load under uniform distribution might affect 30% under AI-optimized placement. No current operational metric measures whether the AI's placement decisions are expanding the blast radius of a power component failure. Uptime metrics capture the outcome after a failure but cannot reveal the concentration that made it worse.
Decision Supported:
Is the workload orchestrator's thermal optimization silently concentrating power load in ways that expand the blast radius of a single power component failure?
Formula
Plain language: Compute the coefficient of variation (standard deviation divided by mean) of load across all PDU circuits during AI-optimized operation. Divide by the same coefficient of variation from a baseline period. A ratio above 1.0 means AI has increased load variability, concentrating power.
Notation:
TPDDI = CV_power_AI / CV_power_baseline
Variables:
| Variable | Definition | Unit |
|---|---|---|
| CV_power_AI | Coefficient of variation of load (kW) across PDU circuits during AI-optimized workload placement | Dimensionless |
| CV_power_baseline | Coefficient of variation during uniform or non-optimized distribution from commissioning data or controlled baseline | Dimensionless |
| CV | Standard_deviation(loads) / mean(loads) | Dimensionless |
A TPDDI of 1.0 indicates no additional imbalance introduced by AI. A TPDDI of 2.0 means AI doubles the power distribution variability relative to baseline. For normalization across heterogeneous rack densities, load should be expressed as percentage of circuit-rated capacity rather than absolute kW.
Proposed formula, pending validation pilots.