How the estimate works
The capacity assessment is a deterministic scenario model. The same inputs produce the same result. Its coefficients are model assumptions; they are not presented as validated industry benchmarks or measured PhysaFlow deployment outcomes.
Capacity model
Unused capacity rate = 1 − utilization / 100. Multiply that rate by the cooling and workload factor below, add the tier and visibility adjustments, then cap the resulting stranded rate at 65%. Stranded MW = commissioned IT capacity × stranded rate.
| Cooling | AI training | Other workloads |
|---|---|---|
| Air | 0.52 | 0.45 |
| Liquid or immersion | 0.38 | 0.30 |
| Hybrid | 0.45 | 0.375 |
Tier I/II subtracts 2 percentage points; Tier III adds none; Tier IV adds 3. Unified visibility adds none; partial visibility adds 4 percentage points; no visibility adds 8. PUE does not currently affect the calculation.
The homepage scenario
At 80 MW, 60% utilization, air cooling, AI training, Tier III, and partial visibility: (1 − 0.60) × 0.52 + 0.04 = 24.8%, or 19.84 MW of modeled stranded capacity. Changing the homepage slider recomputes the same formula.
No systems are connected, no intervention is recommended or executed, and no capacity change has been observed. Utilization represents an assumed annual average; there is no measurement period or customer dataset behind this scenario. The calculation does not establish that this capacity is safe to recover.
Economic assumptions
Annual exposure multiplies stranded MW by an assumed $0.9M–$2.8M per MW per year. Potentially recoverable value applies an assumed 55% recovery fraction to the low value and 70% to the high value. These are scenario inputs embedded in the model, not verified prices, savings, or recovery rates.
GPU-hour equivalent
For AI workloads, the model multiplies stranded MW by an assumed 400 GPUs per MW and 8,760 hours per year. This is a theoretical annual equivalent, not additional GPU hours demonstrated or guaranteed. Enterprise workloads do not produce a GPU-hour estimate.
Limits and next steps
The model does not simulate thermal transients, redundancy, network dependencies, workload eligibility, or site-specific control limits. Use it to scope further measurement. Engineering review and operating data are needed to establish usable capacity and approve any changes.