Inference hosting break-even worksheet
- Task and accepted-request unit (including failed/retried work):
- Same quality, safety and latency bar for both candidates:
- Measured request-length/concurrency/peak distribution:
- Dated quotes and currency:
- API fixed/variable costs:
- Owned fixed/variable costs, including spare/idle/rollout capacity and staff:
- Benchmarked capacity at required latency and availability:
- Migration cost and payback assumptions:
- Mandatory data-route and fallback constraints:
Illustrative model: API = 300 + 0.006Q; owned = 3000 + 0.001Q. Crossing: 540,000 requests/month. At 100,000: 900 versus 3100. At 600,000: 3900 versus 3600. A 400,000-request capacity ceiling invalidates that crossing for the current configuration. A 12,000 transition cost divided by 300 monthly saving is 40 months, conditional on unchanged assumptions. These are invented inputs, not market prices.
Record volume, peak-to-average ratio, fixed capacity steps and staff-cost sensitivity. Do not count a dedicated accelerator pool twice as both capacity and token inference. Choose, defer or reject with an owner, evidence links and review trigger.