Three scenarios for your volume: token API, local hardware (amortised) and Ollama Cloud (flat). We show which is cheapest and where the break-even sits. Every assumption is explicit and editable.
Break-even (API vs local): 209,877 requests/month (above this volume local hardware beats the token API.)
API cost / request: 0.0072 PLN · Every assumption is explicit and editable. Indicative result.
| GPU | VRAM | BW (GB/s) | CAPEX | OPEX/m | Fit (FP8) |
|---|---|---|---|---|---|
| NVIDIA RTX 5090 32GB | 32 GB | 1792 | 24000 PLN | 450 PLN | 2x GPU |
| NVIDIA H200 NVL 141GB | 141 GB | 4800 | 165000 PLN | 2100 PLN | 1x GPU |
| AMD Instinct MI300X 192GB | 192 GB | 5300 | 145000 PLN | 2300 PLN | 1x GPU |