Scenario: high power cost
Canonical worked example. Same box as the capacity listing: 2× A100 80GB, us-west-2, weekday 18:00–23:00 PT. Power is a CAISO-like industrial band.
Inputs
| line | value | cite |
|---|---|---|
| Power | 1.00 kW × 1.4 × $0.18 = $0.252 / h | Power cost |
| Hardware | $29,000 ÷ 26,280 = $1.103 / h | Hardware |
| Overhead | $0.226 / h | Overhead |
| Floor | $1.581 / h | sum |
| Ask (25%) | $1.581 × 1.25 = $1.976 / h | Margin |
Ask field
price = 1.976 ÷ 576 = $0.00343 per 1,000 tokens → "0.00343"
capacity = 50000
ttl_seconds = 300
model_id = "gpt-4"Rest only inside the 18:00–23:00 PT window. If three asks expire unfilled, cut 10% toward the floor ($1.581 / h → "0.00275"), not below it. See when to adjust.