Self-host vs API
Compare the per-token cost of renting a GPU against paying an API - the answer is all about utilisation.
Self-host
Self-host cost: $3.13 / 1M
API
The API is cheaper - by $2.33 per 1M tokens. You'd need higher utilisation to beat it.
Break-even at ~100% utilisation.
Ignores setup time, engineering, redundancy and cold starts. A rough model, not a quote.