API billed per token, GPUs you rent, or hardware you own — which is cheapest at your volume? With real sustained utilization and itemized overhead. The open-source companion to LeCompute. Use one currency throughout.
tokens/day × 30 × price_per_Mtok / 1e6.ceil(volume / (throughput × util × 86400)) × (rental + overhead) × 24 × 30. The two terms that decide it: sustained utilization (a GPU at 10% load costs ~10× per token vs 90%) and the overhead on top of the sticker rental — network, storage, orchestration, and amortized on-call.capex / amort_months + electricity — the "buy a box" case.Why utilization, not the hourly rate, is what usually decides API vs self-host: the cost write-up. Current EU GPU rental prices (open data): lecompute-benchmarks.