Self-host vs API — at what volume does renting a GPU win?
Open-weight models can run on rented GPUs (~$1–$4/hr for an A100 / H100) or via per-token APIs. Plug in your numbers; we'll show the breakeven curve.
Monthly input tokens
30,000,000
@ 1,000,000/day
Monthly output tokens
15,000,000
@ 500,000/day
Self-host GPU
Intel Gaudi2
needs 82 GB VRAM
Self-host monthly
$751
24/7 on cheapest provider
Self-host on Intel Gaudi2
Cheapest fitting GPU for Qwen2.5 32B at FP16. Live at Cyfuture AI at $1.044/hr → $751/month at 24/7 utilisation.
Ad