Self-Host vs API Break-Even Calculator
GPU rental, ops time and real utilisation against per-token API spend, solved for the monthly request volume at which running your own box starts to win.
The workload, and what the API charges for it
The box you would run instead
You are at 2,000,000 — 49.5% of the way there. The API costs $2,700 a month at that volume; the box costs $5,450 whether you send it one request or a million.
- API cost per request
- $0.0014
- API cost per month
- $2,700
- GPU rental per month
- $3,650
- Ops time per month
- $1,800
- Self-hosted cost per month
- $5,450
- Capacity at 100% utilisation
- 4,730M output tokens / mo
- Capacity at 35.0% utilisation
- 1,656M output tokens / mo
- Output tokens you actually need
- 800M / mo
- Utilisation this workload implies
- 16.9%
- GPUs needed at your utilisation
- 2 is enough
- Self-hosted, per 1M output tokens at 35.0%
- $3.29
- Self-hosted, per 1M output tokens at 100%
- $1.15
- API, per 1M output tokens (blended)
- $3.38
- Difference per month at your volume
- $2,750
Utilisation is the term that decides this and it is the term most published comparisons omit. The arithmetic people usually run divides a GPU's hourly rate by its peak throughput, arrives at a cost per million tokens far below any API rate, and concludes that self-hosting is obviously cheaper. It would be, if the GPU generated tokens every second of every hour you paid for it. It does not. Traffic is diurnal, traffic is bursty, and you have to size for the peak, which means the trough is idle capacity you are still renting.
Run the sensitivity yourself: halve the utilisation field and watch the self-hosted cost per million double. Nothing else on this page has that leverage. A twenty per cent improvement in throughput moves the answer by twenty per cent; going from 30% utilisation to 60% halves your cost per token. That is why the serious version of this decision is a scheduling problem — filling the troughs with batch work, consolidating several workloads onto one fleet — rather than a hardware-selection problem.
The break-even is expressed as a monthly request count because that is the number you can check against your own dashboard tomorrow. Below it, an API is cheaper and also carries none of the risk. Above it, the box is cheaper on paper, and the remaining question is whether the ops burden — which is included here as a line item and is routinely underestimated — is one your team wants. There is a wide band around the crossing point where the honest answer is that the money is a wash and you should decide on other grounds: data residency, model choice, latency floor, or simply which failure you would rather be woken up for.