When does self-hosting pay off?

GPU vs API · self-hosting payback

The decision every FDE and startup eventually faces: run the model on your own GPUs, or pay per token to an API. Put in your real numbers — your GPU rate, your measured goodput, your monthly overhead, the API price — and we compute your self-hosted monthly cost, your all-in cost per million output tokens, the volume where self-hosting starts to beat the API, and what the API would charge for that same throughput. We deliberately don't pre-fill any price: a number you didn't measure is a number we won't guess for you.

Hours per month uses the standard 730-hour convention (the toolkit workbook exposes it as an editable input if your deployment differs). Every other number is your own. This is planning math, not a quote.

Go deeper

This is the same model that ships in The FDE Deployment Toolkit 2026 — with the latency sizing (Little's law), the per-task and per-user math this page leaves out. The full workbook recalculates every row as you change an input, so you can reconcile it against a real invoice.

$19 · open the full model ↗